Patent Application
20220405575
The present invention relates to a system for predicting a user drop-out rate based on artificial intelligence learning and an operation method thereof. Specifically, the present invention relates to a user drop-out rate prediction system for predicting a user drop-out rate by inputting question information to an encoder neural network and inputting response information to a decoder neural network in a transformer structure.
Recently, in line with popular utilization of the Internet and electronic devices in every field, the educational environment is also rapidly changing. In particular, the development of various educational media enables learners to choose and use a wider range of learning methods. Among them, educational services through the Internet have played the role of a major teaching and learning method thanks to the advantages of overcoming time and space constraints and attaining low-cost education.
This online educational service is combined with various models of artificial intelligence to predict the probability of a user's correct answer to a random question, which is impossible in the existing offline education environment, thereby providing more efficient learning content.
However, the users in the mobile environment are more vulnerable to various factors that cause the user to drop out of the online learning environment during learning, so dropping-out may occur more frequently in the online environment than in the offline environment. For example, whenever a call is received, a text message is received, a notification of SNS such as Facebook or Instagram is received, or various other notifications are received, the user may drop out of the online learning content in order to check information thereof, thereby lowering the learning efficiency. Although consistent research about methods of preventing the user's dropping-out in the offline environment has been made, research on dropping-out during learning in the online environment has not yet been conducted.
A transformer model among a variety of recent artificial intelligence models uses only attention, instead of RNN, to produce an encoder-decoder structure, thereby providing a very fast learning speed and superior performance to RNN, so it began to be applied to the online education field.
It is possible to predict the probability that a user will drop out of the online learning environment using the transformer model, and a method for predicting a drop-out rate more effectively using the transformer model, such as a method of obtaining inference results optimized for predicting a user drop-out rate in the online learning environment depending on configuration of data input to an encoder and a decoder of the transformer model, a method of preventing the prediction based on the questions that have not yet solved by the user due to the characteristics of the learning content, or the like, is required.
In order to solve the above problem, the present invention proposes a transformer structure-based artificial intelligence model optimized for predicting a drop-out rate and an input data format thereof, thereby providing the effect of predicting a drop-out rate of the user who learns in the mobile environment with higher accuracy.
The present invention relates to a user knowledge tracking method with more improved accuracy. An operation method of a user drop-out rate prediction system including a plurality of encoder neural networks and a plurality of decoder neural networks may include the steps of: inputting question information to a kth encoder neural network and inputting response information to a kth decoder neural network; generating query data, which is information about a question of which the correct answer probability is desired to be identified by the user, by reflecting a weight to the response information, and generating attention information to be used as a weight for the query data by reflecting a weight to the question information; and learning the user drop-out rate prediction system using the attention information as a weight for the query data.
A system for predicting a user drop-out rate based on artificial intelligence learning and an operation method thereof according to an embodiment of the present invention propose a transformer structure-based artificial intelligence model optimized for predicting a drop-out rate and an input data format thereof, thereby providing the effect of predicting a drop-out rate of the user who learns in the mobile environment with higher accuracy.
The above and other aspects of the present invention will be more clearly understood from the following detailed description exemplary embodiments taken in conjunction with the accompanying drawings, in which:
Specific structural or sequential descriptions of embodiments according to the concept of the present invention disclosed in this specification or application are only provided by way of example in order to explain the embodiments according to the concept of the present invention, and should not be construed as being limited to the embodiments described in the present specification or application so that the embodiments according to the concept of the present invention may be realized in various forms.
Although the embodiment according to the concept of the present invention may be modified in various ways and may have various forms, specific embodiments will be illustrated in the drawings, and will be described in detail in the present specification or application. However, this is not intended to limit the embodiment according to the concept of the present invention to a specific disclosed form, and should be understood to encompass all changes, equivalents, or substitutes included in the spirit and scope of the present invention.
Terms such as “first”, “second”, and/or the like may be used to describe various elements, but the elements are not be limited to the terms. The above terms are used only for the purpose of distinguishing one element from another element, for example, without departing from the scope of the present invention, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element
In the case where one element is described as being “coupled” or “connected” to another element, it should be understood that a third element may be interposed therebetween, as well as that the two elements are directly coupled or connected to each other. On the other hand, the expression “one element is ‘directly coupled’ or ‘directly connected’ to another element” should be understood that no other element is present therebetween. Other expressions describing the relationship between elements, such as “between” and “just between”, “adjacent to” and “directly adjacent to”, or the like, should be interpreted in the same manner.
The terms used herein are intended to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless clearly dictated otherwise. As used herein, terms such as “include”, “have”, or the like are intended to specify presence of the stated features, numbers, steps, operations, elements, parts, or combinations thereof, and it should be understood that the terms do not exclude the possibility of presence or addition of one or more other features, numbers, steps, operations, elements, parts, or combinations thereof.
Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by those skilled in the art to which the present invention pertains. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in the present specification.
In describing the embodiment, a description of a technique that is well known in the technical field to which the present invention pertains and is not directly related to the present invention will be omitted. This is intended to more clearly convey the subject matter of the present invention without obscuring the same by omitting unnecessary description.
Hereinafter, the present invention will be described in detail by describing a preferred embodiment of the present invention with reference to the accompanying drawings. Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.
Referring to
The transformer model is implemented only with attention while following the encoder-decoder, which is the conventional structure of seq2seq. The transformer model does not use RNN, but maintains the encoder-decoder structure in which the encoder receives an input sequence and the decoder outputs an output sequence, like the existing seq2seq, and has characteristics in which there are N encoder and decoder units.
An attention mechanism was proposed to solve the problems of information loss and vanishing gradient that arise from compressing all information into one vector of a fixed size, which is the problem of seq2seq based on RNN.
According to the attention mechanism, all input data to the encoder is referred to once again at every time step the decoder predicts an output word. However, more attention is paid to the input data part that is related to the data to be predicted, instead of referring to all input data at the same proportion.
Referring back to
The drop-out rate information may be generated on the basis of time. The drop-out rate information may be updated to have different values for every hour during learning.
In addition, the drop-out rate information may be generated on the basis of the learning characteristics of the user, such as a correct answer rate to the previously solved questions, the average time for solving questions, a usual learning time, and the like, as well as the characteristics of the question itself, such as the difficulty level of the question.
Since the users in the mobile environment are more vulnerable to various factors that cause dropping-out, the dropping-out may occur more frequently in the online environment than in the offline environment.
For example, whenever a call is received, a text message is received, an SNS notification such as Facebook or Instagram is received, or various other notifications are received, the user may drop out of the online learning content in order to check information thereof, thereby lowering the learning efficiency.
Although various offline methods have been provided to prevent the dropping-out and increase the concentration of the students, these methods are limited to the offline environment and are difficult to be applied directly to the online environment, so a method for predicting the drop-out rate and increasing the concentration by reflecting the characteristics of each user in the online environment is required.
The user drop-out rate prediction system 5 according to an embodiment of the present invention may predict the drop-out rate on the basis of the transformer structure optimized for the online learning environment, and may use input data in a format that is optimized for prediction of the drop-out rate, thereby predicting the drop-out rate with improved accuracy.
Referring back to
The question information E may be information on various types of questions having various difficulty levels, which are provided to measure the knowledge level of the user. The question information E may include metadata on a question and/or the relative position of a corresponding the question in each session. The session may be a learning unit distinguished according to time based on the time step at which the user drops out
The response information R may be information on the answer selected by the user in response to the question information E. The response information R may include information on the time taken to solve the questions given to the user, as well as the answer of the user to the question.
According to an embodiment, the question information E and the response information R may constitute a set of question-response log data. The question information E and the response information R corresponding thereto may be matched as question-response log data, and may be input into the user drop-out rate prediction system 5 as one data unit
In an embodiment, the question information E may be expressed as “E” that is the abbreviation for “Example”, the response information R may be expressed as “R” that is the abbreviation for “Response”, and the drop-out rate information d may be expressed as “d” that is the abbreviation for “Dropping-out”. The question-response log data may be expressed as “I” that is the abbreviation for “Interaction”.
The embedding units 10 and 30 may express input data, such as questions, responses, or a set of questions and responses, as the vector to perform a function of embedding the same in the user drop-out rate prediction system 5. The input data may be expressed as the vector for a latent space in various ways. For example, one of them may be a method in which words are quantified and used by artificial intelligence. Even if the user inputs different expressions or forms, the meanings of words, sentences, and text may be expressed by calculating correlation thereof and representing the same using numerical values.
As will be described later with reference to
The question information E expressed as the vector may be embedded by the embedding unit 10, and may be input into the encoder neural network 20. The response information R expressed as the vector may be embedded by the embedding unit 30, and may be input into the decoder neural network 40.
The encoder neural network 20 may generate attention information on the basis of the embedded question information E. The attention information may be question information weighted while passing through a plurality of layers of the encoder neural network 20. In particular, the attention information may be information generated through self-attention in the encoder neural network 20. The attention information may be mathematically expressed as a probability, and the sum of all attention information is 1.
The attention information may be input into the decoder neural network 40, may be used as a weight for query data of the decoder neural network 40, and may be used to train the user drop-out rate prediction system 5.
An artificial neural network may utilize the attention information in order to learn whether or not which portion is important in conformance to a target function. In particular, self-attention may indicate that attention is paid to oneself, and may be an operation of applying a weight to the portion to be considered important in specific data and reflecting the same back to the data. In the attention of the existing seq2seq, since the correlation is obtained using information on different data, i.e., data of the encoder and data of the decoder, it is impossible to obtain the information according to self-attention in the attention structure of the conventional seq2seq.
The user drop-out rate prediction system 5 may refer to again all input data E1, E2, . . . , Ek, R1, R2, . . . , and Rk−1 with respect to the encoder neural network 20 at every time step of predicting an output result dk* of the decoder neural network 40, and may pay attention to the data related to the corresponding output result depending on the attention information.
The decoder neural network 40 may generate drop-out rate information on the basis of the embedded response information R and the attention information. The decoder neural network 40 may perform multi-head attention in which aforementioned self-attention is performed one time or more, on the response information R.
As described above, the decoder neural network 40 may perform multi-head attention on query data generated from response information R on the basis of the attention information to which a weight is applied depending on the importance in the question information by the encoder neural network 20, thereby generating drop-out rate information.
According to an embodiment of the present invention, by using question information E in the encoder and response information R in the decoder, as input data optimized for prediction of the user drop-out rate, it is possible to implement a user drop-out rate prediction system 5 with more improved performance. The user drop-out rate prediction system 5 may train the user drop-out rate prediction system 5 capable of predicting the drop-out rate using the question information E and the response information R as input data Thereafter, the user drop-out rate prediction system 5 may output drop-out rate information from the learned user drop-out rate prediction system 5. This is an inference process for outputting drop-out rate information indicating the probability in which the user drops out when solving a corresponding question, listening to or watching an explanation or lecture, or the like by processing input data depending on the weight determined in the learning process.
In addition, the present invention may implement a user drop-out rate prediction system 5 capable of preventing prediction of the drop-out rate on the basis of the question that has not yet been solved by the user using upper triangular masking in the encoder neural network 20 and the decoder neural network 40 of the transformer structure optimized for the online learning environment.
Referring to
Although the embedding units 10 and 30 in
The question information E may include a plurality of questions E1, E2, . . . , and Ek expressed as vectors. The response information R may include responses R1, R1, . . . , and Rk−1 of the user to the respective questions E1, E2, . . . , and Ek expressed as vectors. The drop-out rate information d may include respective drop-out rate information d1, d2, . . . , and dk indicating the drop-out probability of the user for the respective questions expressed as vectors.
In an embodiment, the drop-out rate information dk may be information on the probability of dropping out of the learning program while the user is solving the question E1 in the case where a user response to the question E1 is R1, a user response to the question E2 is R2, . . . , and a user response to the question Ek-1 is Rk−1.
Although the drop-out rate information has been described as the information on the probability of the user terminating the learning program while solving the questions, according to an embodiment, the drop-out rate information may include the probability of terminating the learning program while the user is reading the explanation on the solved question or watching the lecture therefor.
In this case, drop-out rate information may be generated, which indicates the probability of the user dropping out of the learning program at the time exceeding the time of completion of question solving from the time at which the question is provided to the use. This will be described in detail later with reference to
The question information processor 21 may receive question information E, and may perform a series of operations related to self-attention. The operations may include dividing the question information E into queries, keys, and values, generating a plurality of head values for the respective values, generating a weight from a plurality of query head values and a plurality of key head values, performing masking on the generated weight, and applying the masked weight to the plurality of value head values, thereby generating prediction data.
The prediction data generated in the question information processor 21 may be attention information.
In particular, the question information processor 21 may perform upper triangular masking, as well as the key-query masking, during the masking operation. The key-query masking and the upper triangular masking will be described in detail later with reference to
The non-linearization unit 22 may perform non-linearization of the prediction data output from the question information processor 21. ReLU function may be used in the non-linearization.
Although not shown in the drawing, one or more encoder neural networks 20 may be provided. Attention information generated in the encoder neural network 20 may be input back into the encoder neural network 20 so that a series of operations related to the self-attention and the non-linearization may be repeated several times.
Thereafter, the attention information may be divided into a key and a value to then be input into the second response information processor 42. The attention information may be used as the weight for query data, which is input into the second response information processor 42, and may be used to train the user knowledge drop-out rate prediction system 5.
The first response information processor 41 may receive response information R, and may perform a series of operations related to the self-attention, which is similar to the question information processor 21. The operations may include dividing the question information E into queries, keys, and values, generating a plurality of head values for the respective values, generating a weight from a plurality of query head values and a plurality of key head values, performing masking on the generated weight, and applying the masked weight to the plurality of value head values, thereby generating prediction data.
The prediction data generated in the first response information processor 41 may be query data
The second response information processor 42 may receive query data from the first response information processor 41, may receive attention information from the encoder neural network 20, and may output drop-out rate information.
Attention information may be input into the decoder neural network 40 to then be used as a weight for query data of the decoder, and may be used to train the user drop-out rate prediction system 5.
The attention information may be information on a weight provided in order to intensively consider a specific area of query data. Specifically, the user drop-out rate prediction system 5 may refer to again all input data E1, E2, Ek, R1, R2, . . . , and Rk−1 with respect to the encoder neural network 20 at every time step of predicting an output result (d) of the decoder neural network 40, and may pay attention to the data related to the corresponding output result.
The second response information processor 42 may generate dk, which is user drop-out rate information for question information Ek according to the above operation.
Although not shown in the drawing, one or more decoder neural networks 40 may be provided. The drop-out rate information generated in the decoder neural networks 40 may be input back into the decoder neural networks 40 so that a series of operations related to the self-attention, the multi-head attention, and the non-linearization may be repeated several times.
Like the question information processor 21, the first response information processor 41 and the second response information processor 42 may perform upper triangular masking, as well as key-query masking, during the masking operation.
Referring to
In
In addition, although the explanation is described as being read and the lecture is described as being watched, five human senses may be utilized variously in consideration of the learning efficiency according to an embodiment.
In
In an embodiment, the user may begin to solve question 1 at t1. After the solving of question 1 is finished at t2, the learning program may provide explanation 1 and lecture 1. Explanation 1 may be provided until t3, and then lecture 1 may be provided. The user may drop out of the learning program at t4. The dropping-out may include a variety of situations in which the user no longer uses the learning program, such as the case where the screen of a mobile device is turned off, the case where the learning program switches to the background state as another application is executed, the case where the mobile device is turned off, or the like, as well as the case where the user directly terminates the learning program. In addition, the dropping-out may be determined even in the case where the user does not perform any activity (e.g., the state in which no input received from the user while the learning program is running) for a predetermined time or more. If the dropping-out of the user is defined on the basis of the time during which the user does not perform any activity, the session may be defined as a learning unit divided according to the time on the basis of the time at which the dropping-out of the user occurs.
For example, the embodiment in
The predetermined time may be variously configured in addition to one hour, and may be configured differently for each question depending on the question type (or category). For example, the listening part may be configured to have a shorter time than the reading part in the TOEIC test, the longer the passage, the longer the time in the same listening part, and the shorter the passage, the shorter the time in the same reading part.
If the user drops out of the learning program at t4, the learning program may save the state of lecture 1 that is being played at the drop-out time t4. Thereafter, if the user executes the learning program again, the automatically saved lecture may be played from the drop-out time t4. The replayed lecture may be lecture 2.
According to the process described above, lecture 2, question 2, explanation 2, and question 3 may be provided to the user. Since more than one hour has elapsed after t7 at which question 3 is provided until t8 at which explanation 3 is provided as a subsequent activity, the user drop-out rate prediction system 5 may determine that the user has dropped out.
The user drop-out rate prediction system 5 may predict the dropping-out of the user in advance using the artificial intelligence model, and may provide user-customized learning data. As described above, since the dropping-out of the user may occur even while the explanations or the lectures are provided after the question is solved, the user drop-out rate prediction system 5 may predict the drop-out time irrespective of the time for solving the question by the user.
Although
The key-query masking may be an operation of preventing execution of the attention by imposing the penalty to the null value (zero padding) as an optional choice. The value of prediction data in which key-query masking was performed may be expressed as 0, and the remaining portions may be expressed as 1.
Although the last values of the queries and keys are masked for the convenience of description in the key-query masking in
The upper triangular masking may be an operation of preventing execution of the attention by the information corresponding to the future position for the prediction of a subsequent question. For example, this may be a masking operation in order to prevent a prediction value from being calculated on the basis of the question that the user has not yet solved. Like the key-query masking, the value of prediction data in which the upper triangular masking was performed may be expressed as 0, and the remaining portions may be expressed as 1.
Then, the values of the masked prediction data may be controlled to have the probability close to zero when they are probabilistically expressed through the Softmax function by reflecting an arbitrarily large negative number.
In the conventional transformer structure, the key-query masking is performed in the encoder neural network, and the upper triangular masking is performed together with the key-query masking in the decoder neural network. In the user drop-out rate prediction system 5 of the present invention, the upper triangular masking may be performed both in the encoder neural network and in the decoder neural network, so that the drop-out rate information is able to be controlled to rely only on the question information E1, E2, . . . , and Ek previously provided to the user and the response information R1, R2, . . . , and Rk−1 that the user has already submitted.
Referring to
Question identification information may be a unique value assigned to each question. The user or a computer may identify a corresponding question through the question identification information.
Question category information may be information indicating the type of question. For example, in the TOEIC test question, the question category may be information indicating whether the question relates to a listening part or a reading part.
Start time information may be the time at which the user first encounters a given question. That is, this may be the time at which the question is provided to the user.
Input sequence position information may be information indicating the position of a corresponding question or response of the question in the overall question sequence or overall response sequence. Unlike the RNN structure, since the sequence of input data is not indicated in the transformer structure, it is necessary to separately indicate the position of each piece of data within the overall data sequence in order to distinguish the sequence thereof. The input sequence position information may be embedded together with input data, and may be added to the embedded input data, thereby being input into the encoder neural network 20 and the decoder neural network 40.
Session position information may be information indicating the position of the question or response in the session. The session position information may be counted by 1 for every question, and in the case of moving on to the next session, the session position information may be initialized to 1, and may be counted again from 1.
Response accuracy information may be information indicating whether the response of the user is a correct answer or a wrong answer. For example, if the response of the user is a correct answer, index may be “1”, whereas if the response of the user is a wrong answer, index may be “0”. And the index may be mapped to a vector.
Elapsed time information may be information obtained by converting the time taken in order for the user to solve the question into a vector. The elapsed time information may be expressed in seconds, minutes, hours, or the like, and if a predetermined time (e.g., 300 seconds) elapses, it may be determined that a corresponding time (300 seconds) has elapsed.
Input-on-time information may be information on whether or not the user responds within the time limit suggested by a domain expert. The input-on-time information may have index “1” if it is true. Otherwise, the input-on-time information may have index “0”. And the index may be mapped to a vector.
Drop-out rate information may be information indicating whether or not the user drops out while solving a corresponding question, reading an explanation, or watching a lecture. The drop-out rate information may have index “1” if it is true. Otherwise, the drop-out rate information may have index “0”. And the index may be mapped to a vector.
Referring to
Response information R may include start time information, input sequence position information, session position information, response accuracy information, elapsed time information, input-on-time information, and drop-out rate information. However, the response information is not limited thereto, and may exclude any one of them, or may further include other information according to an embodiment.
Although not shown in the table, the question-response log data I may include question identification information, question category information, start time information, input sequence position information, session position information, response accuracy information, elapsed time information, input-on-time information, and drop-out rate information. However, the question-response log data is not limited thereto, and may exclude any one of them, or may further include other information according to an embodiment.
The user drop-out rate prediction system 5 according to an embodiment of the present invention may input the question information E configured as the above-described data format into the encoder neural network 20 of the user drop-out rate prediction system 5, and may input the response information R configured as the above-described data format into the decoder neural network 40 thereof to generate drop-out rate information, thereby providing the effect of predicting a drop-out rate with improved accuracy.
Deep attentive study session dropout prediction in mobile learning environment (DAS) may be a term indicating the user drop-out rate prediction system 5 of the present invention. An area under the receiver operating characteristic curve (AUC) is a value indicating the ratio of correct prediction to overall prediction. The AUC is the value indicating the sensitivity depending on the specificity, and the higher the AUC, the higher the prediction accuracy.
Referring to
In particular, when the sequence size is 5 (DAS-5), the AUC is 0.7379, and when the sequence size is 25 (DAS-25), the AUC is 0.6856, so it is confirmed that accuracy of DAS-5 is higher than that of DAS-25.
The question information E and the response information R, as input data, may be configured as various sequence sizes.
Referring to
The abbreviation in
Referring to
In this case, if start time information st is further added to the question information E, and if start time information st, input-on-time information iot, and elapsed time information et are sequentially added to the response information R, the AUCs may be 0.6653, 0.6819, and 0.7017, respectively. Here, when session position information sp is added to the question information E, and if session position information sp and drop-out rate information d are added to the response information R, the AUC may be 0.7379.
Summarizing the above description, it is possible to implement the user drop-out rate prediction system 5 having the highest accuracy when the question information E includes question identification information id, question category information c, start time information st, input sequence position information p, and session position information sp, and when the response information R includes start time information st, input sequence position information p, session position information sp, response accuracy information r, elapsed time information et, input-on-time information iot, and drop-out rate information d.
Referring to
The question information E may include question identification information id, question category information c, start time information st, input sequence position information p, and session position information sp. (See
In step S903, the user drop-out rate prediction system 5 may generate attention information by reflecting a weight to the question information E, and may generate query data by reflecting a weight to the response information R.
Specifically, the question information E may reflect a weight to the question information E through self-attention. The self-attention may be an operation of applying a weight to the portion to be considered important in specific data and reflecting the same back to the data.
The weight may be reflected to the response information R by performing multi-head attention based on attention information, as well as the self-attention.
In step S905, query data output from the first response information processor 41 of the kth decoder neural network 40 may be input into the second response information processor 42. The query data may be prediction data output from the first response information processor 41.
In step S907, the user drop-out rate prediction system 5 may use attention information as a weight for the query data of the second response information processor 42, thereby training the user drop-out rate prediction system 5.
In step S909, the user drop-out rate prediction system 5 may compare k with N, and if k is equal to N, may perform step S911, and if k is smaller than N, may return to step S903 to repeat step S903 to step S907.
Since the encoder neural network 20 and the decoder neural network 40 may be stacked as many as N, the above process may be repeated until a series of operations including the self-attention and the multi-head attention for all the stacked encoder neural networks 20 and decoder neural networks 40 are completed.
In step S911, the user drop-out rate prediction system 5 may output drop-out rate information from the learned user drop-out rate prediction system 5.
This is an inference process for outputting drop-out rate information indicating the probability in which the user drops out when solving a corresponding question, listening to or watching an explanation or lecture, or the like by processing input data depending on the weight determined in the learning process.
Referring to
In step S1003, each element may generate a plurality of head values for the respective queries, keys, and values.
Each element may generate a weight from a plurality of query head values and a plurality of key head values in step S1005, and may perform a masking operation including key-query masking and upper triangular masking in step S1007.
The key-query masking may be an operation of preventing execution of the attention by imposing the penalty to the null value (zero padding) as an optional choice. The value of prediction data in which key-query masking was performed may be expressed as 0, and the remaining portions may be expressed as 1.
The upper triangular masking may be an operation of preventing execution of the attention by the information corresponding to the future position for the prediction of a drop-out rate for a subsequent question. For example, this may be a masking operation of preventing execution of attention for the question that the user has not yet solved in order to prevent prediction of drop-out rate information d from the question that has not yet been solved. Like the key-query masking, the value of prediction data in which the upper triangular masking was performed may be expressed as 0, and the remaining portions may be expressed as 1.
Thereafter, in step S1009, prediction data may be generated by applying the masked weight to the plurality of value head values.
The prediction data generated in the question information processor 21 may be attention information, the prediction data generated in the first response information processor 41 may be query data, and the prediction data generated in the second response information processor 42 may be drop-out rate information d.
The user drop-out rate prediction system 5 according to the present invention may use an optimized input data format, and may perform upper triangular masking both on the encoder neural network 20 and on the decoder neural network 40 in the transformer structure, thereby obtaining improved performance.
The embodiments of the present invention shown in the present specification and drawings are merely provided for specific examples to easily explain the technical concept of the present invention and help understanding of the present invention, and are not intended to limit the scope of the present invention. It will be obvious to those of ordinary skill in the art to which the present invention pertains that other modifications based on the technical spirit of the present invention are able to be implemented in addition to the embodiments disclosed herein.
