CONGENIALITY-PRESERVING GENERATIVE ADVERSARIAL NETWORKS FOR IMPUTING LOW-DIMENSIONAL MULTIVARIATE TIME-SERIES DATA

Abstract

Recent advances and techniques in missing data imputation suffer from inherent limitations of preserving the relationship among the input feature attributes and the target variable and temporal relations between observations spanning across timeframes because of which it is also challenging to reconcile missing data for any downstream tasks. Present disclosure provides system and method that implement for congeniality-preserving Generative Adversarial Networks (cpGAN) for imputing low-dimensional incomplete multivariate industrial time-series data. The method minimizes the rubric based on the information theory for Machine Learning (ML) between the empirical probability distributions of the reconcile data and the non-linear original data to preserve the temporal dependencies and retain the input feature-attributes and target-variable relationship and probability distributions of the original data.

Description

PRIORITY CLAIM

This U.S. patent application claims priority under 35 U.S.C. § 119 to: Indian Patent Application No. 202221011587, filed on Mar. 3, 2022. The entire contents of the aforementioned application are incorporated herein by reference.

TECHNICAL FIELD

The disclosure herein generally relates to congeniality-preserving Generative Adversarial Networks (cpGAN), and, more particularly, to congeniality-preserving Generative Adversarial Networks (cpGAN) for imputing low-dimensional multivariate industrial time-series data.

BACKGROUND

Complex industrial units are big data digital behemoths, typically the sensors record in the scale of several Gigabytes of the plant operation data. These sensory observations are incomplete due to sensor failure including other various reasons and are of less utility. The quality of the data is of high priority for deploying data-driven artificial intelligence (AI) algorithms for process control, optimization, etc., as inaccurate decision-making impacts product quality, and plant/industrial units' safety. The imputation methods are in general classified as either discriminative or generative. Recent advances in missing data imputation through generative adversarial network (GAN) architectures suffer from inherent limitations of preserving the relationship among the input feature variables and the target variable and temporal relations between observations spanning across timeframes because of which it is also challenging to reconcile missing data for any downstream tasks.

SUMMARY

Embodiments of the present disclosure present technological improvements as solutions to one or more of the above-mentioned technical problems recognized by the inventors in conventional systems. For example, in one aspect, there is provided a processor implemented method for imputing low-dimensional incomplete multivariate industrial time-series data using a congeniality-preserving Generative Adversarial Networks (cpGAN). The method comprises obtaining, via one or more hardware processors, an input training dataset, a cluster independent random noise and one or more associated cluster labels corresponding to the input training dataset; transforming, via the one or more hardware processors, the cluster independent random noise by using the one or more associated cluster labels corresponding to the input training dataset to obtain a cluster dependent random noise; generating, via the one or more hardware processors, an imputed synthetic noise based on (i) a mask variable, (ii) one or more feature embeddings of the training dataset, (iii) a flipped mask variable, and (iv) the obtained cluster dependent random noise; generating, via the one or more hardware processors, one or more imputed high-dimensional feature embeddings using the generated imputed synthetic noise; predicting, via the one or more hardware processors, one or more imputed high-dimensional target feature embeddings of the training dataset using the one or more imputed high-dimensional feature embeddings; generating, via the one or more hardware processors, one or more single-step ahead imputed feature embeddings using the one or more imputed high-dimensional feature embeddings; and generating, via the one or more hardware processors, an imputed training data using the one or more single-step ahead imputed high-dimensional feature embeddings.

In an embodiment, the cluster dependent random noise is obtained from the cluster independent random noise that is sampled from a Gaussian distribution and the one or more associated cluster labels corresponding to the input training dataset.

In an embodiment, the flipped mask variable is obtained based on a difference between a pre-defined value and the mask variable.

In an embodiment, the method further comprises minimizing a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings.

In an embodiment, the method further comprises validating the imputed training data based on a comparison of the imputed training data and the input training dataset.

In an embodiment, the method further comprises classifying the one or more imputed high-dimensional feature embeddings into at least one class type.

In another aspect, there is provided a processor implemented system for imputing low-dimensional incomplete multivariate industrial time-series data using a congeniality-preserving Generative Adversarial Networks (cpGAN). The system comprises a memory storing instructions; one or more communication interfaces; and one or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to: obtain an input training dataset, a cluster independent random noise and one or more associated cluster labels corresponding to the input training dataset; transform the cluster independent random noise by using the one or more associated cluster labels corresponding to the input training dataset to obtain a cluster dependent random noise; generate an imputed synthetic noise based on (i) a mask variable, (ii) one or more feature embeddings of the training dataset, (iii) a flipped mask variable, and (iv) the obtained cluster dependent random noise; generate, by using a generator module comprised in the cpGAN, one or more imputed high-dimensional feature embeddings using the generated imputed synthetic noise; predict, by using a critic module comprised in the cpGAN, one or more imputed high-dimensional target feature embeddings of the training dataset using the one or more imputed high-dimensional feature embeddings; generate, by using a supervisor module comprised in the cpGAN, one or more single-step ahead imputed high-dimensional feature embeddings using the one or more imputed high-dimensional feature embeddings; and generate, by using a recovery module comprised in the cpGAN, an imputed training data using the one or more single-step ahead imputed high-dimensional feature embeddings.

In an embodiment, the cluster dependent random noise is obtained from the cluster independent random noise that is sampled from a Gaussian distribution and the one or more associated cluster labels corresponding to the input training dataset.

In an embodiment, the flipped mask variable is obtained based on a difference between a pre-defined value and the mask variable.

In an embodiment, the one or more hardware processors are further configured by the instructions to minimize a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings.

In an embodiment, the one or more hardware processors are further configured by the instructions to validate, by using a discriminator comprised in the cpGAN, the imputed training data based on a comparison of the imputed training data and the input training dataset.

In an embodiment, the one or more hardware processors are further configured by the instructions to classify, by using a discriminator comprised in the cpGAN, the one or more imputed high-dimensional feature embeddings into at least one class type.

In yet another aspect, there are provided one or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause imputing low-dimensional incomplete multivariate industrial time-series data using a congeniality-preserving Generative Adversarial Networks (cpGAN) by: obtaining an input training dataset, a cluster independent random noise and one or more associated cluster labels corresponding to the input training dataset; transforming the cluster independent random noise by using the one or more associated cluster labels corresponding to the input training dataset to obtain a cluster dependent random noise; generating an imputed synthetic noise based on (i) a mask variable, (ii) one or more feature embeddings of the training dataset, (iii) a flipped mask variable, and (iv) the obtained cluster dependent random noise; generating one or more imputed high-dimensional feature embeddings using the generated imputed synthetic noise; predicting one or more imputed high-dimensional target feature embeddings of the training dataset using the one or more imputed high-dimensional feature embeddings; generating one or more single-step ahead imputed high-dimensional feature embeddings using the one or more imputed high-dimensional feature embeddings; and generating an imputed training data using the one or more single-step ahead imputed high-dimensional feature embeddings.

In an embodiment, the cluster dependent random noise is obtained from the cluster independent random noise that is sampled from a Gaussian distribution and the one or more associated cluster labels corresponding to the input training dataset.

In an embodiment, the flipped mask variable is obtained based on a difference between a pre-defined value and the mask variable.

In an embodiment, the one or more instructions which when executed by the one or more hardware processors further cause minimizing a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings.

In an embodiment, the one or more instructions which when executed by the one or more hardware processors further cause validating the imputed training data based on a comparison of the imputed training data and the input training dataset.

In an embodiment, the one or more instructions which when executed by the one or more hardware processors further cause classifying the one or more imputed high-dimensional feature embeddings into at least one class type.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the disclosed principles:

FIG. 1 depicts an exemplary system for imputing low-dimensional incomplete multivariate industrial time-series data using a congeniality-preserving Generative Adversarial Networks (cpGAN), in accordance with an embodiment of the present disclosure.

FIG. 2 depicts an exemplary high level block diagram of the system 100 illustrating the congeniality-preserving Generative Adversarial Networks (cpGAN) for imputing low-dimensional incomplete multivariate industrial time-series data, in accordance with an embodiment of the present disclosure, in accordance with an embodiment of the present disclosure.

FIG. 3 depicts an exemplary flow chart illustrating a method for imputing low-dimensional incomplete multivariate industrial time-series data, using the systems of FIGS. 1 and 2, in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION

Exemplary embodiments are described with reference to the accompanying drawings. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. Wherever convenient, the same reference numbers are used throughout the drawings to refer to the same or like parts. While examples and features of disclosed principles are described herein, modifications, adaptations, and other implementations are possible without departing from the scope of the disclosed embodiments.

As mentioned earlier, conventional imputation methods are in general classified as either discriminative or generative. Recent advances in missing data imputation through generative adversarial network (GAN) architectures suffer from inherent limitations of preserving the relationship among the input feature variables and the target variable and temporal relations between observations spanning across timeframes because of which it is also challenging to reconcile missing data for any downstream tasks. To overcome these drawbacks, embodiments of the present disclosure provide system and method that implement a Congeniality-Preserving Generative Adversarial Networks (cpGAN) that enables reconcile missing data by preserving the temporal dependencies, probability distributions of the original data and retain its utility for any downstream tasks.

More specifically, system and method of the present disclosure implement an implicit probabilistic model-based congeniality-preserving Generative Adversarial Networks (cpGAN) for imputing low-dimensional incomplete multivariate industrial time-series data with an adversarial trained generator neural network. The cpGAN architecture is presented as an alternative paradigm of research for numerical modeling of continuum mechanics and transport phenomena based on imputation techniques. The cpGAN architecture as implemented by the system and method of the present disclosure is established on two-player non-cooperative zero-sum adversarial game, based on game theory, and minimax optimization-approach. The system and method of the present disclosure also leverage artificial intelligence systems to demonstrate the random missing data imputation-utility efficacy tradeoff for downstream tasks on the open-source industrial benchmark datasets.

In other words, the congeniality-preserving Generative Adversarial Networks (cpGAN) is an architecture comprising of embedding, recovery, critic, supervisor, generator and discriminator and is implemented by the present application for imputing low-dimensional incomplete multivariate industrial time-series data. The method described herein minimizes the rubric based on the information theory for Machine Learning (ML) between the empirical probability distributions of the reconcile data and the non-linear original data to preserve the temporal dependencies and retain the input feature-attributes and target-variable relationship and probability distributions of the original data.

Referring now to the drawings, and more particularly to FIGS. 1 through 3, where similar reference characters denote corresponding features consistently throughout the figures, there are shown preferred embodiments and these embodiments are described in the context of the following exemplary system and/or method.

FIG. 1 depicts an exemplary system for imputing low-dimensional incomplete multivariate industrial time-series data using a congeniality-preserving Generative Adversarial Networks (cpGAN), in accordance with an embodiment of the present disclosure. In an embodiment, the system 100 is also referred as “congeniality-preserving Generative Adversarial Networks (cpGAN) system”, or cpGAN and may be interchangeably used herein. In an embodiment, the system 100 includes one or more hardware processors 104, communication interface device(s) or input/output (I/O) interface(s) 106 (also referred as interface(s)), and one or more data storage devices or memory 102 operatively coupled to the one or more hardware processors 104. The one or more processors 104 may be one or more software processing components and/or hardware processors. In an embodiment, the hardware processors can be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuitries, and/or any devices that manipulate signals based on operational instructions. Among other capabilities, the processor(s) is/are configured to fetch and execute computer-readable instructions stored in the memory. In an embodiment, the system 100 can be implemented in a variety of computing systems, such as laptop computers, notebooks, hand-held devices (e.g., smartphones, tablet phones, mobile communication devices, and the like), workstations, mainframe computers, servers, a network cloud, and the like.

The I/O interface device(s) 106 can include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like and can facilitate multiple communications within a wide variety of networks N/W and protocol types, including wired networks, for example, LAN, cable, etc., and wireless networks, such as WLAN, cellular, or satellite. In an embodiment, the I/O interface device(s) can include one or more ports for connecting a number of devices to one another or to another server.

The memory 102 may include any computer-readable medium known in the art including, for example, volatile memory, such as static random-access memory (SRAM) and dynamic-random access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable ROM, flash memories, hard disks, optical disks, and magnetic tapes. In an embodiment, a database 108 is comprised in the memory 102, wherein the database 108 comprises information training dataset, associated cluster independent random noise and cluster labels, associated cluster dependent random noise, mask variable, associated flipped mask variable, imputed synthetic noise, one or more imputed high-dimensional feature embeddings, one or more imputed target feature embeddings, one or more single-step ahead imputed feature embeddings, imputed training data, Gaussian distribution of the cluster independent random noise corresponding to the input training dataset, validated imputed training data, and the like. The database 108 further comprises information related to error minimization between imputed training data and the input training data, class type associated with the one or more imputed high-dimensional feature embeddings, and the like. The memory 102 further comprises (or may further comprise) information pertaining to input(s)/output(s) of each step performed by the systems and methods of the present disclosure. In other words, input(s) fed at each step and output(s) generated at each step are comprised in the memory 102 and can be utilized in further processing and analysis.

FIG. 2 depicts an exemplary high level block diagram of the system 100 illustrating the congeniality-preserving Generative Adversarial Networks (cpGAN) for imputing low-dimensional incomplete multivariate industrial time-series data, in accordance with an embodiment of the present disclosure. The congeniality-preserving Generative Adversarial Networks (cpGAN) system 100 includes a generator module 202, a critic module 204, a supervisor module 206, a recovery module 208, an embedding module 210, and a discriminator module 212. It is to be understood by a person having ordinary skill in the art or person skilled in the art that the communication between various modules as depicted in FIG. 2 shall not be construed as limiting the scope of the present disclosure. In other words, at any given point of time, any module depicted in FIG. 2 can either directly or via another module may communicate with other modules to generate specific outputs as described herein. For instance, the generator module 202 may communicate with recovery module 208 and the embedding module 210 directly, in an embodiment of the present disclosure. It is to be further understood by a person having ordinary skill in the art or person skilled in the art the above modules (modules 202 till 212) as depicted in FIG. 2 are implemented as at least one of a logically self-contained part of a software program, a self-contained hardware component, and/or, a self-contained hardware component with a logically self-contained part of a software program embedded into each of the hardware component that when executed perform the method described herein.

FIG. 3 depicts an exemplary flow chart illustrating a method for imputing low-dimensional incomplete multivariate industrial time-series data, using the systems of FIGS. 1 and 2, in accordance with an embodiment of the present disclosure. In an embodiment, the system(s) 100 comprises one or more data storage devices or the memory 102 operatively coupled to the one or more hardware processors 104 and is configured to store instructions for execution of steps of the method by the one or more processors 104. The steps of the method of the present disclosure will now be explained with reference to components of the system 100 of FIG. 1, the block diagram of the system 100 depicted in FIG. 2, and the flow diagram as depicted in FIG. 3.

The complex industrial units are big data digital behemoths. The sensory observations are incomplete due to sensor failure and thus are of less utility. The quality of the data is of high priority for deploying data-driven Artificial Intelligence (AI) algorithms for process control, optimization, etc. as inaccurate decision-making impacts product quality, and plant safety. The imputation methods are in general classified as either discriminative or generative. Recent advances in missing data imputation through GAN architectures have suffered from inherent limitations of preserving the relationship among the input feature variables and the target variable and temporal relations between observations spanning across timeframes. To overcome these drawbacks, embodiments of the present disclosure implement a congeniality-preserving Generative Adversarial Networks (cpGAN) framework to reconcile missing data by preserving the temporal dependencies, probability distributions of the original data and retaining its utility for downstream tasks. To put it briefly, congeniality-preserving Generative Adversarial Networks (cpGAN) is implemented for imputing low-dimensional incomplete multivariate industrial time-series data. The algorithmic approach minimizes the rubric based on the Information theory for Machine Learning between the empirical probability distributions of the reconciled data and the non-linear original data to preserve the temporal dependencies, retain the input feature variables and target-variable relationship, and probability distributions of the original data. The architecture as depicted in FIG. 2 is presented as an alternative paradigm of approach in comparison to the conventional first principles-based transport phenomena macroscopic mathematical modeling and numerical simulations for missing data imputation. In addition, the cpGAN generative imputation technique which fuses imperceptible noise, which is customized to the input process database can also be leveraged as a data anonymization technique, a privacy-preserving mechanism to prevent de-identification of the process-plants databases by the third-party adversary, and as defenses to adversarial attacks.

Problem Formulation

Consider representing a f(∈N)-dimensional euclidean real space (the totality of f-space) as =Π_i=11^f _i× . . . ×_f. Assume that I is a continuous stochastic real-valued variable modeled by a finite-dimensional space, . D is termed as the realizations of I and D is referred to as the set of the f-tuples with the alphabet, T_n×2N sampled from the domain, ^(f). The euclidean probability distribution of D is described by (D). The dataset, D∈^(Tn×2N,f) observed at timestamps, t={1, 2, . . . , (T_n×2N)} is denoted as a sequence of T_n×2N observations.

D={(D₁^(t), . . . ,D_f^(t))},∀t E{1,2, . . . ,T_n×2N}  (1)

D^(t)∈I^(f) consists of f-continuous feature variables observations at t-th time point. D^(t)={(D₁^(t), . . . , D_f^(t))} represents the data vector. D_j^(t) denotes the observed value of the j-th (∈f) feature variable at t-th time point. M∈^(Tn×2N,f) be a binary stochastic mask variable and it is described by,

M={(M₁^(t), . . . ,M_f^(t)},∀t∈{1,2, . . . ,T_n×2N}  (2)

M∈{0,1}^(Tn×2N,f) and it describes the missingness of the data. M_j^(t) denotes the mask value (also referred as mask variable) for the j-th (∈f) feature variable at t-th time point. In the present disclosure, the system and method described herein address the missing data imputation for the multidimensional time series dataset under the setting if the missingness of the feature variables happens altogether at random. It is mathematically described by,

(M^(t))˜(M^(t)|D^(t))  (3)

D_j^(t)|M_j^(t)=1 describes observed values of D_j for a jth-feature variable at t-th time point. D_j^(t) is observed if M_j^(t)=1 otherwise D_j^(t)|M_j^(t)=0 s absent in the recorded data. The observed data of D is described by,

D _obs={(D₁^(t)⊙M₁^(t), . . . ,D_f^(t)⊙M_f^(t)},∀t∈{1,2, . . . ,T_n×2N}  (4)

The unobserved data of D is described by,

D _mis={(D₁^(t)⊙(1−M₁^(t)), . . . ,D_f^(t)⊙(1−M_f^(t)))},∀t∈{1,2, . . . ,T_n×2N}  (5)

The incomplete dataset, D is rearranged as,

{tilde over (D)} _{n,1:T n} ∈I ^{T n ×f} ,∀n∈{1,2, . . . ,2N}  (6)

|2N| denotes the dataset cardinality. It is to be considered for n=1, {tilde over (D)}_1,1:T1∈^T1×f. T₁ denotes the finite length of the sequence, n=1. {tilde over (D)}_1,1:T1 consists of observation values of D_t at time points, t∈{1, 2, . . . , T₁}. In the same way for the sequence, n=2, {tilde over (D)}_2,1:T2∈^T2×f is an array of arrays, which consists of observation values of D_t at time points, t∈{(T₁+1), . . . , (T₁+T₂)}. The finite-length, T_n of n-th sequence is a hyper-parameter of the algorithm and is a constant for all the sequences, ∀n∈{1, . . . , 2N}. {tilde over (D)}_n,1:Tn be the observed multivariate time series of length, T_n for a particular sequence, n∈{1, . . . , 2N} corresponding to all the feature variables, f. {tilde over (D)}_n,t is the n-th sequence observation values of the feature variables, f at the t-th time point. {({tilde over (D)}_n,1:Tn)}_n=1^2N is split into two disjoint finite sets with N sequences.

$\begin{array}{cc}{\left\{{\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right\}}_{n=1}^N\bigcap {\left\{{\tilde{D}}_{{\mathrm{test}}_{n,1:T_n}}\right\}}_{n=1}^N=\varnothing & \left(7\right)\end{array}$

The missingness mask, M is also rearranged as,

M _{n,1:T n} ∈^Tn×f,∀n∈{1,2, . . . ,2N}  (8)

{(M_n,1:Tn)}_n=1^2N is also split into two disjoint finite sets with N sequences.

$\begin{array}{cc}\{{M_{{\mathrm{train}}_{n,1:T_n}}\}}_{n=1}^N\bigcap {\left\{M_{{\mathrm{test}}_{n,1:T_n}}\right\}}_{n=1}^N=\varnothing & \left(9\right)\end{array}$

The traditional Generative Adversarial Missing Data Imputation Networks consisted of generator and discriminator modules which are trained simultaneously in competing minimax game to generate imputed samples of having the same distribution as that of the fully observed data, D. Embodiments of the present disclosure implement a cpGAN algorithmic architecture to overcome the inherent limitations of the generative imputation networks to preserve the characteristics of a multidimensional fully observed time-series data such as joint distributions, temporal dynamics, the relationship between independent variables and the dependent target variable in the imputed data by operating on rearranged data, {tilde over (D)}_n,1:Tn The cpGAN architecture is trained to generate and impute the unobserved data of

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}},$

∀n∈{1, 2, . . . , N} and the imputed data is denoted by, {circumflex over (D)}_n,1:Tn, ∀n∈{1, 2, . . . , N}. The cpGAN architecture performance is evaluated and reported on

${\tilde{D}}_{{\mathrm{test}}_{n,1:T_n}},$

∀n∈{1, 2, . . . , N}. Given,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

the imputation network of the cpGAN framework learns a density {circumflex over (P)}({circumflex over (D)}_n,1:Tn) that best approximates the unknown probability density function,

$P\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right)$

such that it minimizes the weighted sum of the Kullback-Leibler (KL) divergence and the Wasserstein distance (W) of order-1 between the original,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

and imputed data, {circumflex over (D)}_n,1:Tn continuous probability distributions,

$\begin{array}{cc}\underset{\hat{P}}{\min }\left[\text{⁠}\mathrm{KL}\left(P\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right)❘❘\hat{P}\left({\tilde{D}}_{n,1:T_n}\right)\right)+\gamma W\left(P\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right),\hat{P}\left({\tilde{D}}_{n,1:T_n}\right)\right)\right]& \left(10\right)\end{array}$

The imputed data, {circumflex over (D)}_n,1:Tn should also preserve the temporal dynamics of the observed data,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

for the imputed data to be of substantial utility in downstream forecasting tasks.

$\begin{array}{cc}\underset{\hat{P}}{\min }\left[\mathrm{KL}\left(P\left({\tilde{D}}_{{\mathrm{train}}_{n,t}}|{\tilde{D}}_{{\mathrm{train}}_{n,1:t-1}}\right)❘❘1\hat{P}\left({\tilde{D}}_{n,t}|{\tilde{D}}_{n,1:t-1}\right)\right)+\mathrm{\gamma W}\left(P\left({\tilde{D}}_{{\mathrm{train}}_{n,t}}|{\tilde{D}}_{{\mathrm{train}}_{n,1:t-1}}\right),\hat{P}\left({\tilde{D}}_{n,t}|{\tilde{D}}_{n,1:t-1}\right)\right)\right],t\in 1:T_n& \left(11\right)\end{array}$

The objective of missing data imputation generative neural network is also to preserve the relationship between the independent feature variables, f_c⊂f, and target variable, f_T∈f of the observed data,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}.$

The unbiased imputed data, {circumflex over (D)}_n,1:Tn is then beneficial for utilization in the downstream predictive analytics task. It is described by,

$\begin{array}{cc}\underset{\hat{P}}{\min }\mathrm{KL}\left[\left(P\left({\tilde{D}}_{{\mathrm{train}}_{n,t}}^{f_T}|{\tilde{D}}_{{\mathrm{train}}_{nt}}^{f_c}\right)❘❘\hat{P}\left({\tilde{D}}_{n,t}^{f_T}|{\tilde{D}}_{n,t}^{f_c}\right)\right)+\mathrm{\gamma W}\left(P\left({\tilde{D}}_{{\mathrm{train}}_{n,t}}^{f_T}|{\tilde{D}}_{{\mathrm{train}}_{n,t}}^{f_c}\right),\hat{P}\left({\tilde{D}}_{n,t}^{f_T}|{\tilde{D}}_{n,t}^{f_c}\right)\right)\right],t\in 1:T_n& \left(12\right)\end{array}$

Referring to steps of FIG. 3, in an embodiment, at step 202 of the present disclosure, the one or more hardware processors 104 obtain an input training dataset

$\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right),$

a cluster independent random noise (Z_n,1:Tn) and one or more associated cluster labels

$\left({\tilde{C}}_{{\mathrm{train}}_{n,1:T_n}}\right)$

corresponding to the input training dataset (e.g., low-dimensional multivariate time-series data as known in the art).

In an embodiment, at step 304 of the present disclosure, the one or more hardware processors 104 transform the cluster independent random noise (Z_n,1:Tn) using the one or more associated cluster labels

$\left({\tilde{C}}_{{\mathrm{train}}_{n,1:T_n}}\right)$

corresponding to the input training dataset

$\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right)$

to obtain a cluster dependent random noise (Z_n,1:Tn). The cluster dependent random noise is obtained from a Gaussian distribution of the cluster independent random noise corresponding to the input training dataset, in an example embodiment of the present disclosure. The steps 302 and 304 are better understood by way of following illustrative description.

It is assumed by the present disclosure and its system and method that, Z_n,1:Tn∈R^Tn×f, ∀n∈{1,2, . . . , N} be an f-dimensional uniformly distributed random variable of length, T_n for a sequence, n with values in the half-open interval [0,1) sampled from a uniform distribution, Z. The cluster independent random noise (Z_n,1:Tn) is transformed based on the cluster labels,

${\tilde{C}}_{{\mathrm{train}}_{n,1:T_n}}\in R^{T_n},\forall n\in \left\{1,2,\dots ,N\right\}.$

The cluster labels are determined by an iterative distance-based algorithm to partition the unlabeled dataset,

$D_{{\mathrm{train}}_{n,1:T_n}}$

into k-predetermined distinct non-overlapping, non-homogeneous clusters. Label embedding vectors, e^c∈R^f′, ∀c∈{1, . . . , k} are obtained from the learnable label embedding matrix, W∈R^k×f′ based on the labels,

${\tilde{C}}_{{\mathrm{train}}_{n,1:T_n}}.$

f′ is the characteristic dimension (a hyperparameter) of the embedding matrix, W. The label embedding vectors, e^c corresponding to the labels,

$C_{{\mathrm{train}}_{n,1:T_n}}$

are concatenated to obtain the label matrix, L_n,1:Tn^c. Matrix-matrix product of Z n and L_n,1:Tn^c is performed to obtain the cluster dependent random noise (also referred as cluster membership aware noise and interchangeably used herein), Z_n,1:Tn^C, in an example embodiment of the present disclosure. The labels,

${\tilde{C}}_{{\mathrm{train}}_{n,1:T_n}}$

corresponding to the fully observed dataset,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

are obtained by performing vector quantization through the unsupervised learning technique. It is to be understood by a person having ordinary skill in the art that matrix-matrix product of Z_n,1:Tn and L_n,1:Tn^c operation shall not be construed as limiting the scope of the present disclosure. The following is determined by the k-means clustering algorithm:

• • 1. Initialize cluster centroids randomly, μ₁, μ₂, . . . , μ_k∈I^(f).
  • 2. Repeat until convergence to minimize the within-cluster sum of pairwise squared deviations:

{
for every n, while n ≤ N, ∀n ∈ {1, 2, . . . , N},
for each n, set
C~trainn,1:Tn:=argminmD~trainn,1:Tn-μm2
for each m ∈ k, set
μm:=∑t=1Tn1⁢{C~trainn,1:Tn=m}⁢D~trainn,t∑t=1Tn1⁢{C~trainn,1:Tn=m}
}

Referring to steps of FIG. 3, in an embodiment, at step 306 of the present disclosure, the one or more hardware processors 104 generate an imputed synthetic noise

$(W_{{\theta }^{\prime }}\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\odot M_{n,1:T_n}+\left(1-M_{n,1:T_n}\right)\odot Z_{n,1:T_n}^C\right)$

based on (i) a mask variable (M_n,1:Tn), (ii) one or more feature embeddings of the training dataset

$\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\right),$

(iii) a nipped mask variable (1−M_n,1:Tn), and (iv) the obtained cluster dependent random noise (Z_n,1:Tn^C). The system 100 invokes the generator module 202 of the cpGAN for generating the imputed synthetic noise. In the present disclosure, the one or more hardware processors 104 via the generator module 202 (also referred as a generator and interchangeably used herein) perform a linear transformation on the summation of the product of the mask variable with the feature embeddings of the training dataset and the flipped mask variable with the cluster-dependent random noise to obtain the imputed synthetic noise. In an embodiment, the flipped mask variable is obtained based on a difference between a pre-defined value (e.g., in this case value of ‘1’) and the mask variable (M_n,1:Tn). It is to be understood by a person having ordinary skill in the art or person skilled in the art that the pre-defined value=1 shall not be construed as limiting the scope of the present disclosure. In an embodiment, at step 308 of the present disclosure, the one or more hardware processors 104 generate one or more imputed high-dimensional feature embeddings (Ĥ_n,1:Tn) using the generated imputed synthetic noise. The generator module 202 generates one or more imputed high-dimensional feature embeddings (Ĥ_n,1:Tn) using the generated imputed synthetic noise, in an embodiment of the present disclosure. The one or more hardware processors 104 further classify the one or more imputed high-dimensional feature embeddings into at least one class type (e.g., real, or fake).

The above steps 306 and 308 that describe generation of the imputed synthetic noise

$(W_{{\theta }^{\prime }}\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\odot M_{n,1:T_n}+\left(1-M_{n,1:T_n}\right)\odot Z_{n,1:T_n}^C\right)$

and the one or more imputed high-dimensional feature embeddings (Ĥ_n,1:Tn) may be better understood by way of following illustrative description:

The generator module 202 comprised in the cpGAN takes as input the realizations of

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}},M_{n,1:T_n},Z_{n,1:T_n}^C$

and outputs a high-dimensional latent variable, H and the same is expressed way of equation below:

$\begin{array}{cc}{G}_{\mathrm{cpGAN}}:{\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\times M_{n,1:T_n}\times Z_{n,1:T_n}^C\to {\underset{\_}{H}}_{n,1:T_n}& \left(13\right)\end{array}$

The generative imputation neural network function can also be viewed as, G_cpGAN:I^(Tn,f)×{0,1}^(Tn,f)×[0,1]^(Tn,f)→H^(Tn,f), ∀n∈{1, 2, . . . , N}. The temporal latent variables H_n,1:Tn, Ĥ_n,1:Tn∈H and are computed as,

$\begin{array}{cc}{\underset{\_}{H}}_{n,1:T_n}=G_{\mathrm{cpGAN}}\left(W_{{\theta }^{\prime }}\left({\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\odot M_{n,1:T_n}+\left(1-M_{n,1:T_n}\right)\odot Z_{n,1:T_n}^C\right)\right)& \left(14\right)\end{array}$ $\begin{array}{cc}{\hat{H}}_{n,1:T_n}=M_{n,1:T_n}\odot {\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}+\left(1-M_{n,1:T_n}\right)\odot {\underset{\_}{H}}_{n,1:T_n}& \left(15\right)\end{array}$

W_θ, denotes the trainable parameter and it is shared across the sequences, n, ∀n∈{1, 2, . . . , N}. ⊙ denotes Hadamard product and 1∈R^Tn×f. The generator (also referred as the generator module) is realized by leveraging a sequential operation on a 3-layer stack of neural-network architectures comprising of unidirectional LSTM and a feed-forward neural network layer, in an example embodiment of the present disclosure. The loss function for the incomplete observed data is described below,

$\begin{array}{cc}{ℒ}_G\left({\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}},{\hat{H}}_{n,1:T_n}\right)={\sum }_{n=1}^N\left(1-M_{n,1:T_n}\right){\left({\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}-{\hat{H}}_{n,1:T_n}\right)}^2& \left(16\right)\end{array}$

It is to be understood by a person having ordinary skill in the art or person skilled in the art that the generator architecture as described above shall not be construed as limiting the scope of the present disclosure.

In an embodiment, at step 310 of the present disclosure, the one or more hardware processors 104 predict one or more imputed high-dimensional target feature embeddings (Ĥ_n,1:Tn^(T) also referred as

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}^{\left(T\right)})$

of the training dataset using the one or more imputed high-dimensional feature embeddings (Ĥ_n,1:Tn). The system 100 of the present disclosure invokes a critic module 204 from the cpGAN 100 for prediction. More specifically, in the critic module (or also referred as a critic neural network or a critic and interchangeably used herein), there comprises a F_cpGAN:H*_n,1:Tn→R, which is a mathematical function used for determining the target variable. Here, * refers to {tilde over (H)}_train^(1:f−1) or Ĥ^(1:f−1). The critic neural network function takes as input the realizations of

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}^{\left(1:f-1\right)}\mathrm{or}{\widehat{H}}_{n,1:T_n}^{\left(1:f-1\right)}$

and outputs,

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}^{\left(T\right)}\mathrm{or}{\widehat{H}}_{n,1:T_n}^{\left(T\right)}.$

The variable subset selection includes the features attributes from the set, {1, . . . , f−1} ⊂f in H_n,1:Tn^{{(1, . . . ,f−1)}*} as input variables to the model. The last feature variable in H_n,1:Tn^{(T)}* denoted by the superscript, T E f denotes the target variable to predict. The loss function for the target variable prediction is described below:

$\begin{array}{cc}{ℒ}_F\left({\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}},{\widehat{H}}_{n,1:T_n}\right)={\Sigma }_{n=1}^N{\left(F_{\mathrm{cpGAN}}\left({\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}^{\left(1:f-1\right)}\right)-F_{cpGAN}\left({\widehat{H}}_{n,1:T_n}^{\left(1:f-1\right)}\right)\right)}^2& \left(17\right)\end{array}$

The critic module preserves the relationship between independent feature columns and the target variable in the real dataset during the adversarial training of G_cpGAN to generate the relationship preserving synthetic data, {circumflex over (D)}_n,1:Tn by minimizing the _F.

In an embodiment, at step 312 of the present disclosure, the one or more hardware processors 104 generate one or more single-step ahead imputed high-dimensional feature embeddings (H*_n,1:t) using the one or more predicted imputed high-dimensional target feature embeddings. In an embodiment, at step 314 of the present disclosure, the one or more hardware processors 104 generate via a recovery module 208 an imputed training data (D*_n,1:Tn) using the one or more single-step ahead imputed high-dimensional feature embeddings (H*_n,1:Tn/H′_n,1:Tn^f*). The steps 312 and 314 may be better understood by way of following illustrative description. The system 100 of the present disclosure invokes a supervisor module 206 from the cpGAN for generation of the one or more single-step ahead imputed high-dimensional feature embeddings (H*_n,1:t). In an embodiment, the supervisor module from the cpGAN leverages a supervisor neural network function, S_cpGAN to retain conditional temporal dynamics of the original data in the imputed temporal data, {circumflex over (D)}_n,1:Tn The G_cpGAN neural network of the cpGAN framework generates the synthetic latent variables, Ĥ_n,1:Tn The auto-regressive, S_cpGAN:H*_n,1:t−1→H*_n,t, ∀n∈{1, 2, . . . , N}, t∈1:T_n takes as input H*_n,1:t−1 and predicts the one-step-ahead temporal latent variable, H*_n,t conditioned on the past latent sequences. It can also be expressed as, S_cpGAN:H_n,1:Tn*∈Π_t Π_j=i^f_j→H_n,1:Tn^f*∈Π_tΠ_j=i^f_j. The cpGAN framework effectively captures the temporal dynamics of the true data by minimizing the supervised loss,

_S=[Σ_n=1^NΣ_t∥H_n,1:T*−S_cpGAN(H_n,1:T−1*)∥₂]  (18)

The G_cpGAN by operating in the closed-loop receives the ground-truth,

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}$

from the embedding module comprising in the cpGAN. It minimizes L_S by forcing the Ĥ_n,1:Tn appraised by the improper adversary (D_cpGAN) to capture the single-step transitions of the

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}.$

The embedding module 210 comprised in the cpGAN, takes as input the realizations of

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}},$

and outputs feature embeddings,

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}.$

$\begin{array}{cc}{E}_{cpGAN}:{\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}\in {\prod }_t{\prod }_{j-i}^f𝕀\to {\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}\in {\prod }_t{\prod }_{j=i}^f{ℍ}_j,& \left(19\right)\end{array}$ $\forall n\in \left\{1,2,\dots ,N\right\}$

denotes the latent embedding space, Π_j=i^f_j. The supervisor module 206 takes as input the temporal latent feature embeddings, H_n,1:Tn* and predicts one-step ahead temporal embeddings, H_n,1:Tn^f*.

S _cpGAN :H _{n,1:T n} *∈Π_tΠ_j−i^f_j→H_n,1:Tn^f*∈Π_tΠ_j=i^f_j,∀n∈{1,2, . . . ,N}  (20)

The recovery module 208 (also referred as recovery function and interchangeably used herein) takes as input the high-dimensional latent embeddings, H_n,1:Tn* in the supervised setting or H_n,1:Tn^f* in the unsupervised setting and transforms to their corresponding low-dimensional representations, D_n,1:Tn* The superscript, * denotes for real variables,

${\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}},{\tilde{H}}_{{\mathrm{train}}_{n,1:T_n}}^{\prime },{\widehat{D}}_{{\mathrm{train}}_{n,1:T_n}}$

or for synthetic imputed variables, Ĥ_n,1:Tn, Ĥ_n,1:Tn^f and {circumflex over (D)}_n,1:Tn respectively.

R _cpGAN :H _{n,1:T n} * of H_n,1:Tn^f*∈∈Π_tΠ_j−i^f_j→D_n,1:Tn*∈Π_tΠ_j=i^f_j,∀n∈{1,2, . . . ,N}  (21)

The learnable parameters of the embedding and recovery modules are transformed by the joint training of the modules in the supervised-learning approach of reconstructing the input fully observed temporal data,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

through by minimizing a supervised loss as described below,

$\begin{array}{cc}{ℒ}_R={\Sigma }_{n=1}^N{{\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}-{\widehat{D}}_{{\mathrm{train}}_{n,1:T_n}}}_2& \left(22\right)\end{array}$

In joint training of the generator module 202, the supervisor module 206 and the recovery module 208 in unsupervised learning approach, the first moment, |D₁−D₂| and second-order moment, |√{square root over ({circumflex over (σ)}₁²)}−{circumflex over (σ)}₂²|differences, defined between the original data,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

and the imputed data, {circumflex over (D)}_n,1:Tn are minimized respectively. In other words, a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings is minimized using the critic module 204.

The sample means for fully-observed data,

${\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}$

and imputed data, {circumflex over (D)}_n,1:Tn is computed by,

${\underset{\_}{D}}_1=\frac{1}{N}{\Sigma }_{j=1}^f{\Sigma }_{n=1}^N{\tilde{D}}_{{\mathrm{train}}_{n,1:T_n}}^{\left(j\right)}\in I^{\left(f\right)}\mathrm{and}$ ${\underset{\_}{D}}_2=\frac{1}{N}{\sum }_{j=1}^f{\sum }_{n=1}^N{\widehat{D}}_{{\mathrm{train}}_{n,1:T_n}}^{\left(j\right)}\in I^{\left(f\right)}.$

The sample variances, {circumflex over (σ)}₁², {circumflex over (σ)}₂²∈I^(f) are evaluated by,

$$\begin{gathered} \begin{array}{cc}{\hat{\sigma }}_1^2=\frac{1}{N}{\sum }_{j=1}^f{\sum }_{n=1}^N{\left({\tilde{D}}_{\mathrm{trai}{n}_{n,1:T_n}}^{\left(j\right)}-{\underset{\_}{D}}_1^{\left(j\right)}\right)}^2\mathrm{and}{\hat{\sigma }}_2^2=\frac{1}{N}{\sum }_{j=1}^f{\sum }_{n=1}^N{\left({\hat{D}}_{n,1:T_n}^{\left(j\right)}-{\underset{\_}{D}}_2^{\left(j\right)}\right)}^2.\text{ \\ ℒU⁢S}=❘{\underset{\_}{D}}_1-{\underset{\_}{D}}_2❘+❘\sqrt{{\hat{\sigma }}_1^2}-\sqrt{{\hat{\sigma }}_2^2}❘& \left(23\right)\end{array} \end{gathered}$$

The goal of the generator module 202, the supervisor module 206, and recovery module 208 is to minimize the first and the second-order moment's differences between the fully-observed input date,

${\tilde{D}}_{\mathrm{trai}{n}_{n,1:T_n}},$

and the imputed temporal data, {circumflex over (D)}_n,1:Tn obtained from its corresponding high-dimensional latent representations, Ĥ_n,1:Tn^f in the unsupervised learning approach. By minimizing _US, P({circumflex over (D)}_n,1:Tn) learns the underlying probability distributions of the input temporal data,

$P\left({\tilde{D}}_{\mathrm{trai}{n}_{n,1:T_n}}\right).$

Each or the embedding and recovery modules is realized by leveraging a sequential operation on a 3-layer stack of neural-network architectures comprising of uni-directional Long-Short-Term-Memory (LSTM) and a feed-forward neural network layer.

The one or more hardware processors 104 further validate the imputed training data based on a comparison of the imputed training data and the input training dataset. The validation is described above as performed by the recovery module 208. The validation outcome which is referred as a validation dataset is utilized for the hyper-parameter tuning of the cpGAN/system 100.

The one or more hardware processors 104 further classify the one or more imputed high-dimensional feature embeddings into at least one class type. In the present disclosure, the system and method invoke a discriminator (also referred as a discriminator module 212 or a discriminator network and interchangeably used herein) comprised in the cpGAN for performing the classification. The above step of classification may be better understood by way of following description. The objective of the discriminator network, D_cpGAN in cpGAN architecture is to distinguish the observed and imputed values in Ĥ_n,1:Tn The discriminator neural network 212 performs classification of the one or more imputed high-dimensional feature embeddings into at least one class type, and is described by way of expression below,

$\begin{array}{cc}{D}_{\mathrm{cpGAN}}:{\hat{H}}_{n,1:T_n}\times {\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\times H_{n,1:T_n}\to {\hat{M}}_{n,1:T_n},p_{n,1:T_n,m}^{*},p_{n,1:T_n}^{*},P\left(H_{n,1:T_n}^{*}\right)& \left(24\right)\end{array}$

D_cpGAN takes as input the realizations of Ĥ_n,1:Tn,

${\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}$

and H_n,1:Tn (hint matrix, a random binary-valued mask matrix (also referred as mask variable)) and outputs an estimated mask matrix, {circumflex over (M)}_n,1:Tn, the predicted probability of cluster labels, p_n,1:Tn,m*, the predicted probability of adversarial ground-truth, i.e real/fake, p_n,1:Tn*, the estimated probability distributions, P(H_n,1:Tn*), ∀n∈{1, 2, . . . , N} as described by,

$\begin{array}{cc}{\hat{M}}_{n,1:T_n},p_{n,1:T_n,m}^{*},p_{n,1:T_n}^{*},& \left(25\right)\end{array}$ $P\left(H_{n,1:T_n}^{*}\right)=D_{\mathrm{cpGAN}}\left({\hat{H}}_{n,1:T_n}\odot \left(1-H_{n,1:T_n}\right)+{\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\odot H_{n,1:T_n}\right)$

The superscript, corresponds to real,

$p_{\mathrm{trai}{n}_{n,1:T_n,m}},p_{\mathrm{trai}{n}_{n,1:T_n}},P\left({\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\right)$

or synthetic, {circumflex over (p)}_n,1:Tn,m, {circumflex over (p)}_n,1:Tn, P(Ĥ_n,1:Tn). {circumflex over (M)}_n,1:Tn can also be viewed as, {0,1}^(Tn,f), ∀n∈{1, 2, . . . , N}. It gives the probability say, {circumflex over (M)}_n,t^(j) of Ĥ_n,t^(j) for j-th variable corresponding to a sequence, n at t-th timepoint being observed. In the context of the present disclosure, H_n,1:Tn∈ {0,1}^(Tn,f), ∀n∈{1, 2, . . . , N} denotes a Hint matrix. The G_cpGAN of the cpGAN framework produces synthetic outputs, Ĥ_n,1:Tn by operating on the random noise space, Z, and the discriminator neural network, D_cpGAN by operating on the adversarial learning latent space, H, tries to distinguish latent temporal embeddings,

${\hat{H}}_{n,1:T_n}\mathrm{and}{\tilde{H}}_{\mathrm{trai}{n}_{n1:T_n}}.$

The binary adversarial cross-entropy loss for classification of the sequence observation as real or fake is described by,

$\begin{array}{cc}{ℒ}_u=\frac{1}{N}{\sum }_{n=1}^N\left[-\left(y_{n,1:T_n}\log \left(p_{\mathrm{trai}{n}_{n,1:T_n}}\right)+\left(1-y_{n,1:T_n}\right)\log \left(1-p_{\mathrm{trai}{n}_{n,1:T_n}}\right)\right)+\left(y_{n,1:T_n}\log \left({\hat{p}}_{n,1:T_n}\right)+\left(1-y_{n,1:T_n}\right)\log \left(1-{\hat{p}}_{n,1:T_n}\right)\right)\right]& \left(26\right)\end{array}$

γ_n,1:Tn∈{0,1}^Tn, ∀n∈{1, 2, . . . , N} is the adversarial ground-truth, real or fake data.

$p_{\mathrm{trai}{n}_{n,1:T_n}},{\hat{p}}_{n,1:T_n}\in {\left[0,1\right]}^{T_n},\forall n\in \left\{1,2,\dots ,N\right\}$

is the predicted probability of the sequence is real and

$1-p_{\mathrm{trai}{n}_{n,1:T_n}}\&1-{\hat{p}}_{n,1:T_n}$

is the predicted probability of the sequence being fake. D_cpGAN tries to minimize, _U. The G_cpGAN tries to minimize, −_US which helps to learn {circumflex over (P)}({circumflex over (D)}_n,1:Tn) that best approximates

$P\left({\tilde{D}}_{\mathrm{trai}{n}_{n,1:T_n}}\right).$

the cross-entropy loss in binary classification for predicting the input random mask matrix is described by,

$\begin{array}{cc}{ℒ}_M=-\frac{1}{N}{\sum }_{n=1}^N\left(M_{n,1:T_n}\log {\hat{M}}_{n,1:T_n}+\left(1-M_{n,1:T_n}\right)\log \left(1-{\hat{M}}_{n,1:T_n}\right)\right)|M_{n,t}^{\left(j\right)}=0& \left(27\right)\end{array}$

The D_cpGAN attempts to maximize the probability of accurately predicting, M_n,1:Tn. D_cpGAN tries to minimize, _M and G_cpGAN tries to minimize, −_M. The loss, _M is evaluated for mask values, M_n,t^(j)=0. This is due to the sole reason that the G_cpGAN network ideally should impute the missing data in

${\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\odot M_{n,1:T_n}$

with unbiased estimates as given by, (1−M_n,1:Tn)⊙Ĥ_n,1:Tn. The prediction of cluster-label, C_n,1:Tn*(∈{0,1, . . . , k}) is a multi-class classification task. For the multiclass classification task of cluster label prediction, a separate loss for each cluster label is evaluated per observation. Summation operation of the output is then performed.

$\begin{array}{cc}{ℒ}_{LP}=\frac{1}{N}\left[{\sum }_{m=1}^k{\sum }_{n=1}^N-y_{n,1:T_n}^c\log \left(p_{{\mathrm{train}}_{n,1:T_n,m}}\right)+{\sum }_{m=1}^k{\sum }_{n=1}^N{y}_{n,1:T_n}^c\log \left({\hat{p}}_{n,1:T_n,m}\right)\right]& \left(28\right)\end{array}$

k denotes the number of predetermined cluster labels. γ_n,1:Tn^c (ground-truth) denotes the binary value (0 or 1) if clustering label m is the correct classification for observation at time point t(t∈1:T_n) corresponding to data sequence, n.

$p_{{\mathrm{train}}_{n,1:T_n,m}}$

is the predicted probability for real observation at time point t of a data sequence, n belongs to cluster label, m. {circumflex over (p)}_n,1:Tn,m is the predicted probability for imputed synthetic observation at time point t of a data sequence, n belongs to cluster label, m. The cluster labels are determined by,

$C_{\mathrm{trai}{n}_{n,1:T_n}}^p:=\arg \underset{m}{\max }\left[\mathrm{Softmax}\left(p_{\mathrm{trai}{n}_{n,1:T_n,m}}\right)\right],m\in \left\{0,1,\dots ,k\right\}$ ${\hat{C}}_{n,1:T_n}:=\arg \underset{m}{\max }\left[\mathrm{Softmax}\left({\hat{p}}_{n,1:T_n,m}\right)\right],m\in \left\{0,1,\dots ,k\right\}$

The D_cpGAN tries to minimize, £_P whereas G_cpGAN tries to minimize,

$\left(-{ℒ}_{LP}\right).C_{\mathrm{trai}{n}_{n,1:T_n}}^p$

denote the predicted cluster labels by the neural-network architecture in comparison with the ground-truth,

$C_{\mathrm{trai}{n}_{n,1:T_n}}$

corresponding to real data,

${\tilde{D}}_{\mathrm{trai}{n}_{n,1:T_n}}.{\hat{C}}_{n,1:T_n}$

denote the predicted cluster labels for the imputed temporal data, {circumflex over (D)}_n,1:Tn Wasserstein distance between the estimates of two probability distributions

$\mathbb{P}\left({\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\right)\mathrm{and}\hat{\mathbb{P}}\left({\hat{H}}_{n,1:T_n}\right)$

is also minimized. The Wasserstein loss, _W is described by,

$\begin{array}{cc}{ℒ}_W=W\left(\mathbb{P}\left({\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\right),\hat{\mathbb{P}}\left({\hat{H}}_{n,1:T_n}\right)\right)={\inf }_{\gamma ~\prod \left(P,\hat{P}\right)}{𝔼}_{\left(H_{{\mathrm{train}}_{n,1:T_n}},{\hat{H}}_{n,1:T_n}\right)~\gamma }\left[\mathbb{P}\left({\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\right)-\hat{\mathbb{P}}\left({\hat{H}}_{n,1:T_n}\right)\right]& \left(29\right)\end{array}$

$\gamma \sim \prod \left(\mathbb{P}\left({\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\right),\hat{\mathbb{P}}\left({\hat{H}}_{n,1:T_n}\right)\right)$

is the set of all possible joint probability distributions between

$\mathbb{P}\left({\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}}\right)\mathrm{and}\hat{\mathbb{P}}\left({\hat{H}}_{n,1:T_n}\right).$

The D_cpGAN tries to maximize, _W whereas the G_cpGAN tries to minimize, _W. The discriminator module is realized by leveraging a sequential operation on a stack of neural-network architectures comprising of unidirectional LSTM and a feed-forward neural network layer.Training of cpGAN/System of FIGS. 1-2:

The embedding module (E_cpGAN) and the recovery module (R_cpGAN) were jointly trained on the task of reconstructing the fully observed temporal data,

${\tilde{D}}_{\mathrm{trai}{n}_{n,1:T_n}}.$

Initially, the supervisor module (S_cpGAN) was trained in the supervised-learning approach on the single-step ahead prediction task of the fully observed temporal latent variable,

${\tilde{H}}_{\mathrm{trai}{n}_{n,1:T_n}},$

∀n∈{1,2, . . . N} by operating in the latent space, In the beginning, the critic module (F_cpGAN) was trained on the original fully observed data to minimize the target variable prediction loss, _F. Here, the objective is to minimize, min_{Φe,Φr,Φs,Φc}(_R+_S+_F). Φ_e, Φ_r, Φ_s, Φ_c denote the trainable parameters of the embedding, the recovery, the supervisor, and the critic neural network functions respectively. Let Θ_g, Θ_d describe the learnable parameters of the G_cpGAN, D_cpGAN neural network functions. The G_cpGAN outputs {circumflex over (D)}_n,1:Tn, ∀n∈{1, 2, . . . , N}, which contains both the unimputed data and imputed data as given by M_n,1:Tn, and (1−M_n,1:Tn) respectively. G was trained by eight distinct loss functions classified as unsupervised & supervised losses, _US, _G, _U, _W, _M, _S, _F and _LP. G_cpGAN is trained adversarially to minimize the weighted sum of the above loss functions,

$\begin{array}{cc}{ℒ}_G=\underset{{\theta }_g}{\min }\left[\alpha \left(\left(-{ℒ}_U\right)+\gamma \left({ℒ}_W\right)+\left(-{ℒ}_M\right)+\left(-{ℒ}_{\mathrm{LP}}\right)\right)+{ℒ}_{US}+{ℒ}_G+{ℒ}_S+{ℒ}_F\right]& \left(30\right)\end{array}$

Here, α∈R⁺ is a hyper-parameter. In the experiments conducted by the present disclosure, α=100 and γ=10. D_cpGAN was trained by four distinct loss functions _U, _W, _M, and _LP. D was trained to minimize the weighted sum of the loss functions,

$\begin{array}{cc}{ℒ}_D=\underset{{\theta }_d}{\min }\left(\alpha \left(\left({ℒ}_U\right)+\gamma \left(-{ℒ}_W\right)+\left({ℒ}_M\right)+\left({ℒ}_{LP}\right)\right)\right)& \left(31\right)\end{array}$

G_cpGAN, D_cpGAN were trained adversarially by deceptive input as follows, min_GcpGANmax_DcpGAN(G_cpGAN, D_cpGAN) It can be expressed as,

$\begin{array}{cc}\underset{{\theta }_g}{\min }\left[\alpha \left(\left(-{ℒ}_U\right)+\gamma \left({ℒ}_W\right)+\left(-{ℒ}_M\right)+\left(-{ℒ}_{\mathrm{LP}}\right)\right)+{ℒ}_{US}+{ℒ}_G+{ℒ}_S+{ℒ}_F+\underset{{\theta }_d}{\max }\left(\alpha \left(\left(-{ℒ}_U\right)+\gamma \left({ℒ}_W\right)+\left(-{ℒ}_M\right)+\left(-{ℒ}_{\mathrm{LP}}\right)\right)\right)\right]& \left(32\right)\end{array}$

In conclusion, the cpGAN architecture was trained with both the supervised and unsupervised losses. Performance of the cpGAN was evaluated on

${\tilde{D}}_{tes{t}_{n,1:T_n}}$

and is reported herein accordingly. The unbiased imputed data, {circumflex over (D)}_n,1:Tn is then beneficial for utilization in the downstream predictive analytics task and forecasting tasks. The validation dataset is utilized for the hyper-parameter tuning of the imputation network. Performance of the cpGAN is evaluated on the test dataset.

Results:

Electricity Transformer Temperature (ETT) datasets used during experiments contained 2-year data of electricity transformer usage and reported in, Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. Each dataset (ETTm1, and ETTm2) contained 70,080 (2 year*365 days*24 hours*4 times (observation at 15 min interval)) temporal observations. In addition, hour-wise recorded dataset (ETTh1, ETTh2) has been leveraged to evaluate performance of the method of the present disclosure. The train/validation/test splits are 60/20/20%. Table 1 reports the Root Mean Square Error (RMSE) in imputing the ETT test dataset by the cpGAN architecture. The cumulative imputation error increases with an accretion of the random missing percentage for each column feature attribute on all the datasets. As reported in Table 2, LSTM neural network is utilized as a standard benchmark model trained on original train data on the basis of supervised-learning driven downstream tasks of the target variable (“oil temperature” (° C.)) and one-step-ahead prediction with distinct learnable parameters respectively. The evaluation of the prediction models is performed on the original test dataset and the imputed test datasets. The performance is reported in terms of the RMSE metric. As observed in Table 2, the first column reports the prediction error on the original test dataset. The subsequent columns report prediction error on the imputed test datasets. The adversarial training of the generator imputation network (or the generator) for imputing missing data with the random missing (%) of each column attribute in the range, [2.5%, 20%] resulted in an on-par performance of the prediction model on the imputed test dataset in comparison with the original test dataset. For the single-step ahead prediction, as shown in Table 3, the error in forecasting rises as the missing (%) of feature attributes increases. In Table 4, the experimental results demonstrate the cpGAN algorithmic framework (or cpGAN as implemented by the present disclosure) efficacy in the downstream application task of the target variable prognostics and it outperforms several strong baselines in the literature, as reflected in the lower prediction error. The results reported in Tables 1, 2, 3, and 4 on the respective tasks are obtained from the arithmetic mean of five computational experimental runs. The error of deviation is at most 5% from the statistical mean value reported in Table 2, Table 3, and Table 4. The system and method of the present disclosure leveraged NVIDIA® T4 GPU for the training of deep learning models built upon the PyTorch framework.

TABLE 1
Performance of cpGAN architecture on
ETT test dataset in terms of RMSE metric
	ETTh	ETTh	ETTh	ETTh	ETTh	ETTh	ETTh	ETTh
Dataset	1	1	1	1	2	2	2	2
Missing (%)	2.5	(%)	5	(%)	10	(%)	20	(%)	2.5	(%)	5	(%)	10	(%)	20	(%)
RMSE	0.837	0.852	2.224	4.685	1.067	1.771	2.914	5.267
RMSE	1.045	1.050	1.471	3.053	1.385	1.598	3.359	7.836

TABLE 2
Results of prediction model on the target
variable prediction on ETT test dataset
Missing (%)	0	(%)	2.5	(%)	5	(%)	10	(%)	20	(%)	60	(%)	90	(%)
Dataset	LSTM	LSTM	LSTM	LSTM	LSTM	LSTM	LSTM
ETTh1	11.764	9.737	9.852	10.157	9.370	16.327	26.701
ETTh2	10.786	10.121	10.228	10.885	11.671	28.164	24.760
ETTm1	11.886	8.756	9.231	9.104	9.964	9.124	31.958
EETm2	11.196	11.196	10.498	10.935	12.400	18.240	26.588

TABLE 3
Results of forecasting model on the
single-step ahead prediction on ETT test dataset
Dataset	Missing	(%)	HUFL	HULL	MUFL	MULL	LUFL	LULL	OT
ETTm1	0	(%)	7.78	4.18	7.79	4.01	2.96	2.13	3.69
ETTm1	2.5	(%)	17.78	8.69	18.62	9.29	5.40	5.45	15.18
ETTm1	5	(%)	17.63	8.58	18.65	9.26	5.40	5.40	15.08
ETTm1	10	(%)	17.49	8.63	18.39	9.23	5.32	5.33	14.98
ETTm1	20	(%)	17.54	8.57	18.66	10.58	6.00	5.06	15.29
ETTm1	60	(%)	19.40	8.33	24.03	8.57	5.71	5.29	13.72
ETTm1	90	(%)	28.74	12.99	31.41	13.11	8.43	8.74	25.80
ETTm2	0	(%)	8.25	8.70	7.71	6.91	6.57	5.51	8.09
ETTm2	2.5	(%)	18.28	15.01	16.11	15.20	18.97	27.84	18.20
ETTm2	5	(%)	18.24	15.21	16.02	15.03	19.15	27.68	18.26
ETTm2	10	(%)	18.36	15.45	16.10	15.09	19.14	27.27	18.22
ETTm2	20	(%)	21.59	15.24	17.34	14.81	18.92	28.26	17.30
ETTm2	60	(%)	20.93	21.91	17.35	14.50	18.01	28.37	16.93
ETTm2	90	(%)	22.82	25.76	23.54	21.18	22.80	33.81	20.84

TABLE 4
Results of the target variable prediction model on the
imputed ETT test dataset obtained from various imputation
techniques in the literature. The error metric is RMSE.
	ETTh	ETTh	ETTh	ETTh	ETTh	ETTh	ETTh	ETTh
Dataset	1	1	1	1	2	2	2	2
Missing (%)	10	(%)	20	(%)	60	(%)	90	(%)	10	(%)	20	(%)	60	(%)	90	(%)
MICE	12.18	13.04	21.15	35.51	13.06	14.08	36.16	34.71
(Multivariate
imputation by
Chained
Equations-as
known in the art)
GAN	10.41	10.60	20.81	32.04	11.15	11.96	33.65	34.57
(generative
adversarial
network-as known
in the art)
BRITS	10.96	10.119	19.60	30.70	11.755	12.607	32.670	33.776
(Bidirectional
recurrent
imputation
for time series-as
known in the art)
cpGAN	10.157	9.370	16.327	26.701	10.885	11.671	28.164	24.760
(system of the
present disclosure)
MICE	10.92	11.956	11.742	39.947	12.73	13.895	23.02	34.56
(Multivariate
imputation by
Chained Equations-
as known in
the art)
GAN	9.301	10.113	11.086	36.751	11.198	12.733	21.446	32.836
(generative
adversarial
network-as
known in the art)
BRITS	9.383	10.612	10.864	36.078	11.979	13.064	21.324	31.578
(Bidirectional
recurrent
imputation for
time series-as
known in the art)
cpGAN	9.104	9.964	9.124	31.958	10.935	12.401	18.24	26.588
(system of the
present disclosure)

As described above, the cpGAN comprises of embedding, critic, supervisor, generator, discriminator, and recovery neural network functions (or also referred as modules and interchangeably used herein) to tackle the temporal non-uniformity of the incomplete observed multidimensional continuous-variable time series data. The input to the cpGAN is the multidimensional continuous-variable time series data, random mask variable, and cluster-independent random noise. The input data preprocessing involves feature scaling by transforming the scale of continuous feature variables by utilizing the min-max normalization technique(s) as known in the art to obtain the preprocessed data. The preprocessed data is split into training, validation, and test dataset respectively. Training of the cpGAN consisted of two phases. In the first phase, the following modules were trained namely, the embedding, recovery, critic, and supervisor modules of the imputation network by utilizing the training dataset. More specifically, the training dataset is fed as an input to the embedding module to obtain high-dimensional feature embeddings. The high-dimensional feature embeddings are fed as input to the recovery module to reconstruct the training dataset. The embedding and recovery modules are trained jointly in a supervised learning approach to reconstruct the input training dataset. The high-dimensional feature embeddings are fed as input to the supervisor module to obtain single-step ahead predictions of the feature embeddings. The supervisor module is trained in a supervised learning approach as a forecasting model to minimize the forecasting error predictions on the training dataset. The high-dimensional feature embeddings are fed as input to the critic module. The high-dimensional feature embeddings consist of independent feature embeddings and dependent target feature embedding. The critic module operates on the independent feature embeddings to predict the target feature embedding. The critic module is trained in a supervised learning approach as a prediction model to minimize the target variable prediction error on the training dataset.

In the second phase, the following modules namely, the generator, discriminator, recovery, critic, and supervisor modules of the imputation network/cpGAN are jointly trained by utilizing the training dataset and cluster-independent random noise and the mask variable. For instance, cluster-dependent random noise is obtained by transforming the cluster-independent random noise by using the cluster-labels corresponding to the training dataset. A linear transformation is performed on the summation of the product of the mask variable with the feature embeddings of the training dataset and the flipped mask variable with the cluster-dependent random noise to obtain the imputed synthetic noise. The imputed synthetic noise is fed as input to the generator module to output the imputed high-dimensional feature embeddings. The generator is trained to fool the discriminator to classify the imputed high-dimensional feature embeddings as real. The imputed high-dimensional feature embeddings are fed as input to the critic module to predict the imputed target feature embeddings. The critic module is trained to minimize the difference between the target feature embedding of the training dataset and the predicted imputed target feature embeddings in a supervised learning approach. The imputed feature embeddings are fed to the discriminator to assign a label as real or fake, wherein the discriminator tries to classify the imputed high-dimensional feature embeddings as fake. The imputed high-dimensional feature embeddings are fed as input to the supervisor module to generate the single-step ahead predictions of the imputed feature embeddings. The supervisor module is trained to minimize the difference between the single-step ahead predictions of the imputed feature embeddings and the feature embeddings of the training dataset in a supervised learning approach. The single-step ahead imputed feature embeddings are fed to the recovery module to obtain the imputed training data. The recovery module is trained in an unsupervised learning approach to minimize the difference between the imputed training data and the input training data. Though the experimental results depict for a specific use scenario (e.g., Electricity Transformer Temperature (ETT)) or application, such use scenario or application shall not be construed as limiting the scope of the present disclosure. For instance, the cpGAN system 100 used for imputing low-dimensional multivariate industrial time-series data may be used in Digital Twin, simulation of Industry Plants/machines/sensors (or sensor data), production units and/or manufacturing units.

The written description describes the subject matter herein to enable any person skilled in the art to make and use the embodiments. The scope of the subject matter embodiments is defined by the claims and may include other modifications that occur to those skilled in the art. Such other modifications are intended to be within the scope of the claims if they have similar elements that do not differ from the literal language of the claims or if they include equivalent elements with insubstantial differences from the literal language of the claims.

It is to be understood that the scope of the protection is extended to such a program and in addition to a computer-readable means having a message therein; such computer-readable storage means contain program-code means for implementation of one or more steps of the method, when the program runs on a server or mobile device or any suitable programmable device. The hardware device can be any kind of device which can be programmed including e.g., any kind of computer like a server or a personal computer, or the like, or any combination thereof. The device may also include means which could be e.g., hardware means like e.g., an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a combination of hardware and software means, e.g., an ASIC and an FPGA, or at least one microprocessor and at least one memory with software processing components located therein. Thus, the means can include both hardware means and software means. The method embodiments described herein could be implemented in hardware and software. The device may also include software means. Alternatively, the embodiments may be implemented on different hardware devices, e.g., using a plurality of CPUs.

The embodiments herein can comprise hardware and software elements. The embodiments that are implemented in software include but are not limited to, firmware, resident software, microcode, etc. The functions performed by various components described herein may be implemented in other components or combinations of other components. For the purposes of this description, a computer-usable or computer readable medium can be any apparatus that can comprise, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

The illustrated steps are set out to explain the exemplary embodiments shown, and it should be anticipated that ongoing technological development will change the manner in which particular functions are performed. These examples are presented herein for purposes of illustration, and not limitation. Further, the boundaries of the functional building blocks have been arbitrarily defined herein for the convenience of the description. Alternative boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. Alternatives (including equivalents, extensions, variations, deviations, etc., of those described herein) will be apparent to persons skilled in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope of the disclosed embodiments. Also, the words “comprising,” “having,” “containing,” and “including,” and other similar forms are intended to be equivalent in meaning and be open ended in that an item or items following any one of these words is not meant to be an exhaustive listing of such item or items, or meant to be limited to only the listed item or items. It must also be noted that as used herein and in the appended claims, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.

Furthermore, one or more computer-readable storage media may be utilized in implementing embodiments consistent with the present disclosure. A computer-readable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processor(s) to perform steps or stages consistent with the embodiments described herein. The term “computer-readable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., be non-transitory. Examples include random access memory (RAM), read-only memory (ROM), volatile memory, nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and any other known physical storage media.

It is intended that the disclosure and examples be considered as exemplary only, with a true scope of disclosed embodiments being indicated by the following claims.

Claims

1. A processor implemented method, comprising: obtaining, via one or more hardware processors, an input training dataset, a cluster independent random noise and one or more associated cluster labels corresponding to the input training dataset;transforming, via the one or more hardware processors, the cluster independent random noise by using the one or more associated cluster labels corresponding to the input training dataset to obtain a cluster dependent random noise;generating, via the one or more hardware processors, an imputed synthetic noise based on (i) a mask variable, (ii) one or more feature embeddings of the training dataset, (iii) a flipped mask variable, and (iv) the obtained cluster dependent random noise;generating, via the one or more hardware processors, one or more imputed high-dimensional feature embeddings using the generated imputed synthetic noise;predicting, via the one or more hardware processors, one or more imputed high-dimensional target feature embeddings of the training dataset using the one or more imputed high-dimensional feature embeddings;generating, via the one or more hardware processors, one or more single-step ahead imputed high-dimensional feature embeddings using the one or more imputed high-dimensional feature embeddings; andgenerating, via the one or more hardware processors, an imputed training data using the one or more single-step ahead imputed high-dimensional feature embeddings.

2. The processor implemented method of claim 1, wherein the cluster dependent random noise is obtained from the cluster independent random noise that is sampled from a Gaussian distribution and the one or more associated cluster labels corresponding to the input training dataset.

3. The processor implemented method of claim 1, wherein the flipped mask variable is obtained based on a difference between a pre-defined value and the mask variable.

4. The processor implemented method of claim 1, further comprising minimizing a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings.

5. The processor implemented method of claim 1, further comprising validating the imputed training data based on a comparison of the imputed training data and the input training dataset.

6. The processor implemented method of claim 1, further comprising classifying the one or more imputed high-dimensional feature embeddings into at least one class type.

7. A system, comprising: a memory storing instructions;one or more communication interfaces; andone or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to:obtain an input training dataset, a cluster independent random noise and one or more associated cluster labels corresponding to the input training dataset;transform the cluster independent random noise by using the one or more associated cluster labels corresponding to the input training dataset to obtain a cluster dependent random noise;generate an imputed synthetic noise based on (i) a mask variable, (ii) one or more feature embeddings of the training dataset, (iii) a flipped mask variable, and (iv) the obtained cluster dependent random noise;generate, by using a generator module comprised in the cpGAN, one or more imputed high-dimensional feature embeddings using the generated imputed synthetic noise;predict, by using a critic module comprised in the cpGAN, one or more imputed high-dimensional target feature embeddings of the training dataset using the one or more imputed high-dimensional feature embeddings;generate, by using a supervisor module comprised in the cpGAN, one or more single-step ahead imputed feature embeddings using the one or more imputed high-dimensional feature embeddings; andgenerate, by using a recovery module comprised in the cpGAN, an imputed training data using the one or more single-step ahead imputed high-dimensional feature embeddings.

8. The system of claim 7, wherein the cluster dependent random noise is obtained from the cluster independent random noise that is sampled from a Gaussian distribution and the one or more associated cluster labels corresponding to the input training dataset.

9. The system of claim 7, wherein the flipped mask variable is obtained based on a difference between a pre-defined value and the mask variable.

10. The system of claim 7, wherein the one or more hardware processors are further configured by the instructions to minimize a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings.

11. The system of claim 7, wherein the one or more hardware processors are further configured by the instructions to validate, by using a discriminator comprised in the cpGAN, the imputed training data based on a comparison of the imputed training data and the input training dataset.

12. The system of claim 7, wherein the one or more hardware processors are further configured by the instructions to classify, by using a discriminator comprised in the cpGAN, the one or more imputed high-dimensional feature embeddings into at least one class type.

13. One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause: obtaining, an input training dataset, a cluster independent random noise and one or more associated cluster labels corresponding to the input training dataset;transforming, via the one or more hardware processors, the cluster independent random noise by using the one or more associated cluster labels corresponding to the input training dataset to obtain a cluster dependent random noise;generating, via the one or more hardware processors, an imputed synthetic noise based on (i) a mask variable, (ii) one or more feature embeddings of the training dataset, (iii) a flipped mask variable, and (iv) the obtained cluster dependent random noise;generating, via the one or more hardware processors, one or more imputed high-dimensional feature embeddings using the generated imputed synthetic noise;predicting, via the one or more hardware processors, one or more imputed high-dimensional target feature embeddings of the training dataset using the one or more imputed high-dimensional feature embeddings;generating, via the one or more hardware processors, one or more single-step ahead imputed high-dimensional feature embeddings using the one or more imputed high-dimensional feature embeddings; andgenerating, via the one or more hardware processors, an imputed training data using the one or more single-step ahead imputed high-dimensional feature embeddings.

14. The one or more non-transitory machine-readable information storage mediums of claim 13, wherein the cluster dependent random noise is obtained from the cluster independent random noise that is sampled from a Gaussian distribution and the one or more associated cluster labels corresponding to the input training dataset.

15. The one or more non-transitory machine-readable information storage mediums of claim 13, wherein the flipped mask variable is obtained based on a difference between a pre-defined value and the mask variable.

16. The one or more non-transitory machine-readable information storage mediums of claim 13, wherein the one or more instructions which when executed by the one or more hardware processors further cause minimizing a difference between one or more target feature embeddings of the input training dataset and the one or more predicted imputed high-dimensional target feature embeddings.

17. The one or more non-transitory machine-readable information storage mediums of claim 13, wherein the one or more instructions which when executed by the one or more hardware processors further cause validating the imputed training data based on a comparison of the imputed training data and the input training dataset.

18. The one or more non-transitory machine-readable information storage mediums of claim 13, wherein the one or more instructions which when executed by the one or more hardware processors further cause classifying the one or more imputed high-dimensional feature embeddings into at least one class type.