Patent Application
20170004455
This application relates to the technical fields of software and/or hardware technology and, in one example embodiment, to nonlinear featurization of decision trees for linear regression modeling in the context of on-line social network data.
An on-line social network may be viewed as a platform to connect people in virtual space. An on-line social network may be a web-based platform, such as, e.g., a social networking web site, and may be accessed by a use via a web browser or via a mobile application provided on a mobile phone, a tablet, etc. An on-line social network may be a business-focused social network that is designed specifically for the business community, where registered members establish and document networks of people they know and trust professionally. Each registered member may be represented by a member profile. A member profile may be include one or more web pages, or a structured representation of the member's information in XML (Extensible Markup Language), JSON (JavaScript Object Notation), etc. A member's profile web page of a social networking web site may emphasize employment history and education of the associated member. An on-line social network may include one or more components for matching member profiles with those job postings that may be of interest to the associated member.
Embodiments of the present invention are illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like reference numbers indicate similar elements and in which:
Nonlinear featurization of decision trees for linear regression modeling in the context of an on-line social network is described. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of an embodiment of the present invention. It will be evident, however, to one skilled in the art that the present invention may be practiced without these specific details.
As used herein, the term “or” may be construed in either an inclusive or exclusive sense. Similarly, the term “exemplary” is merely to mean an example of something or an exemplar and not necessarily a preferred or ideal means of accomplishing a goal. Additionally, although various exemplary embodiments discussed below may utilize Java-based servers and related environments, the embodiments are given merely for clarity in disclosure. Thus, any type of server environment, including various system architectures, may employ various embodiments of the application-centric resources system and method described herein and is considered as being within a scope of the present invention.
For the purposes of this description the phrase “an on-line social networking application” may be referred to as and used interchangeably with the phrase “an on-line social network” or merely “a social network.” It will also be noted that an on-line social network may be any type of an on-line social network, such as, e.g., a professional network, an interest-based network, or any on-line networking system that permits users to join as registered members. For the purposes of this description, registered members of an on-line social network may be referred to as simply members.
Each member of an on-line social network is represented by a member profile (also referred to as a profile of a member or simply a profile). The profile information of a social network member may include personal information such as, e.g., the name of the member, current and previous geographic location of the member, current and previous employment information of the member, information related to education of the member, information about professional accomplishments of the member, publications, patents, etc. The profile information of a social network member may also include information about the member's professional skills, such as, e.g., “product management,” “patent prosecution,” “image processing,” etc.). The profile of a member may also include information about the member's current and past employment, such as company identifications, professional titles held by the associated member at the respective companies, as well as the member's dates of employment at those companies.
An on-line social network system also maintains information about various companies, as well as so-called job postings. A job posting, for the purposes of this description is an electronically stored entity that includes information that an employer may post with respect to a job opening. The information in a job posting may include, e.g., the industry, job position, required and/or desirable skills, geographic location of the job, the name of a company, etc. The on-line social network system includes or is in communication with a so-called recommendation system. A recommendation system is configured to match member profiles with job postings, so that those job postings that have been identified as potentially being of interest to a member represented by a particular member profile are presented to the member on a display device for viewing. In one embodiment, the job postings that are identified as of potential interest to a member are presented to the member in order of relevance with respect to the associated member profile.
Member profiles and job postings are represented in the on-line social network system by feature vectors. The features in the feature vectors may represent, e.g., a job industry, a professional field, a job title, a company name, professional seniority, geographic location, etc. A recommendation engine may include a binary classifier (e.g., in the form of a logistic regression model) that can be trained using a set of training data. The set of training data can be constructed using historical data that indicates whether a certain job posting presented to a certain member resulted in that member applying for that job. A trained binary classifier may be used to generate, for a (member profile, job posting) pair, a value indicative of the likelihood that a member represented by the member profile applies for a job represented by the job posting. A value indicative of the likelihood that a member represented by the member profile applies for a job represented by the job posting may be referred to as a relevance value or a degree of relevance. Those job postings, for which their respective relevance values for a particular member profile are equal to or greater than a predetermined threshold value, are presented to that particular member, e.g., on the news feed page of the member or on some other page provided by the on-line social networking system. Job postings presented to a member may be ordered based on their respective relevance values, such that those job postings that are determined to be more relevant (where the recommendation system determined that the member is more likely to apply for jobs represented by those listings as opposed to the jobs represented by other postings) are presented in such a manner that they would be more noticeable by the member, e.g. in a higher position in the list of relevant job postings.
A recommendation engine that is provided in the form of a binary classifier trains a binary classification model on the (member profile, job posting) pairs and their corresponding labels that indicate whether or not the member represented by the member profile has applied for the job represented by the job posting. The binary classification model would learn global weights that are optimized to fit all the (member profile, job posting) pairs in the data set. If the binary classification model treats each (member profile, job posting) pair equally, the overall optimization result may be biased towards those member profiles that have been paired with a larger number of job postings as compared to those member profiles that have been paired with a fewer number of job postings. If the binary classification model treats equally each job posting pared with a member profile, regardless, e.g., of whether the associated member viewed the job posting or not, such that the respective positions of job postings in the ranked list are invisible in the learning process, the algorithm may unduly emphasize unimportant or even irrelevant job postings (e.g., those job postings that were ignored and not viewed by a respective member). In the binary classification model, the degree of relevance may not always be well modeled. For instance, it does not take into consideration that even if a member does not apply for certain jobs, a job posting that is impressed but not clicked by the member may be inferred to be less relevant than the one that is impressed and clicked by the same member.
A learning to rank approach may be utilized beneficially to address some of these problems, as it takes into consideration multiple ordered categories of relevance labels, such as, e.g., Perfect>Excellent>Good>Fair>Bad. A learning to rank model can learn from pairwise preference (e.g., job posting A is more relevant than job posting B for a particular member profile) thus directly optimizing for the rank order of job postings for each member profile. With ranking position taken into consideration during training, top-ranked job postings may be treated by the recommendation system as being of more importance than lower-ranked job postings. In addition, a learning to rank approach may also result in an equal optimization across all member profiles and help minimize bias towards those member profiles that have been paired with a larger number of job postings. In one example embodiment, a recommendation system may be configured to produce relevance labels mentioned above automatically without human intervention.
A recommendation system may be configured to generate respective multi-point scale ranking labels for each (member profile, job posting) pairs. The labels indicating different degrees of relevance may be, e.g., in the format of Bad, Fair, Good, Excellent, and Perfect. Using such label data, a recommendation system may train a ranking model (also referred to as a learning to rank model) that may be used by a ranker module of the recommendation system to rank job postings for each member profile, directly optimizing for the order of the ranking results based on a metric such as, e.g., normalized discounted cumulative gain (NDCG).
In one example embodiment, in order to train a learning to rank model, the recommendation system constructs respective five-point labels for (member profile, job posting) pairs, utilizing feedback data collected by automatically monitoring member interactions with job postings that have been presented to them. In one embodiment, the relevance labels are defined as shown below.
Bad Random: randomly generated synthetic (member profile, job posting) pair of an active member profile with an active job posting, where the job posting has not been presented to the associated member, at all or for a certain period of time.
Fair Impressed: (member profile, job posting) pair, where the job posting has been presented to the associated member (impressed), but there has been no further interaction of the associated member with the job positing , such as a click on the job posting to view the details of the posting.
Good Clicked: (member profile, job posting) pair, where the job posting has been presented to the associated member and the recommendation system also detected a click on the job posting to view the details of the posting, but no further event indicative of applying for the associated job has been detected by the recommendation system.
Excellent Applied: (member profile, job posting) pair, where the job posting has been presented to the associated member, and the recommendation system also detected that the member clicked on the job posting to view the details and applied for the associated job but did not detect a confirmation that the member has been hired for that job.
Perfect Hired: (member profile, job posting) pair, where the recommendation system detected a confirmation that the member has been hired for that job. There are multiple ways to infer hired event within our system, e.g. a) directly through recruiter feedbacks, and b) through members' job change events, which can be further inferred from member updating certain fields of their profile, such as changing of job location, job title and job company.
It will be noted that although, in one embodiment, the recommendation system uses five degrees of relevance, a recommendation system may use a lesser or a greater number of degrees, where each degree of relevance corresponds to a respective temporal sequence of events, each one sequentially closer to the final successful action of a member represented by a member profile applying to a job represented by a job posting. A learning to rank approach described herein may be utilized beneficially in other settings, e.g., where each degree of relevance corresponds to a respective geographic proximity to a given location.
In one example embodiment, a learning to rank model utilized by a recommendation system uses boosted gradient decision trees (BGDT) as the learning to rank algorithm. Once the recommendation system generates multi-point scale relevance labels, it converts these labels into numeric gains and uses the respective Discounted Cumulative Gain (DCG) values as measurements and targets for the model training Table 1 below illustrates how different labels correspond to respective relevance values (identified as “Grade” in Table 1) and respective gains (identified as “Gain” in Table 1).
In Table 1, a Gain value is calculated as expressed in Equation (1) below.
Gain=2Grade−1 Equation (1)
The Discounted Cumulative Gain (DCG) from position 1 to position p in the list of results (e.g., in the list of references to recommended job postings) can be defined as expressed below in Equation (2).
NDCG can then be calculated as the DCG of the rank ordering, divided by the DCG of the ideal ordering (as if returned by an optimal ranker), which is expressed by Equation (3) below. NDCG is always within range [0,1].
As mentioned above, the learning to rank algorithm may be in the form of boosted gradient decision trees and can be directly optimized for NDCG (as list-wise optimization). In Equation (3) above, the DCGranker is calculated using the rank scores and DCGideal is calculated using the relevance labels. The error for an intermediate ranker produced during the training process is the difference between DCG ranker and DCG ideal, which can be used in the tree training process with gradient decent. A small number of small decision trees (e.g., decision trees with five leaves on each tree) can be trained with boosting, where a relevance score for a job posting with respect to a member profile is calculated as the sum of tree scores calculated for that job posting with respect to that member profile using respective decision trees, which is illustrated in a diagram 300 shown in
In one embodiment, the decision paths from the decision trees utilized in the ranking model may be automatically converted into a format that can be consumed by a binary classifier. This approach may be utilized beneficially to combine the two types of modeling techniques—learning to rank and binary classification—without disrupting the existing binary classification modeling architecture.
Method and system are described to transform decision tree paths into nonlinear s-expression features, which may also be referred to as nonlinear featurization of decision trees for linear regression modeling. A decision tree path can represent a business rule and can be provided in an if/then else statement format or in an s-expression format. For example, a business rule may state that, for a particular member profile the recommendation system should present a first job posting (job 1) if the location of the job is in Seattle; otherwise the recommendation system should present a different job posting (job 2). This business rule can be represented in an if/then else statement format, as shown below.
{if Location==Seattle, then return Job 1
The same business rule can be represented in an s-expression format, as shown below.
(if(=location Seattle) Job1 Job2)
Using another example, the s-expression “(if(>x 0) a b)” is equivalent to the statement “if(x>0) then a else b,” where x, a, b can be either variables or s-expressions.
In one embodiment, computer-implemented converter may be configured to evaluate one or more business rules branch operators ‘if’ provided in the s-expression format and transform the s-expressions into a decision tree that could be then included in the learning to rank algorithm used by a recommendation system. The converter may also be configured to read the decision tree structure of a tree that may be included in the learning to rank algorithm, and convert each path from root to a leaf into an s-expression that can be used as an s-expression feature. Thus the decision trees can be converted to features and used in existing logistic regression models. The example below is provided with reference to a decision tree 400 shown in
A converter configured to transform s-expressions into a decision tree and also to convert a decision tree structure into s-expressions may contribute to greater flexibility to use a trained tree-based ranker in different settings. A trained tree-based ranker can be used to replace existing logistic regression models as a pure ranking solution. In some embodiments, a trained tree-based ranker, together with the converter, can be combined with existing logistic regression models. For example, the s-expressions generated based on a tree structure from a ranking model or from one or more business rules may be utilized as additional nonlinear features for training the logistic regression model. This approach provides optimized non-linear features to a linear model such that it may result in increasing of the linear model's capability of predicting nonlinear patterns. In a further embodiment, the s-expressions generated based on a tree structure from a ranking model or from one or more business rules may be utilized as weighted components in the final score produced bet the logistic regression model, without retraining the existing logistic regression model. This approach may be useful in keeping some or most of the benefits of training a ranking model, while leveraging the benefits of the existing logistic regression model.
Logistic regression model use sigmoid function to calculate the final score (score), which is expressed below using Equation (4).
The function φ(x) is modified to incorporate decision tree features in s-expression format as shown below using Equation (5).
Example method and system to utilize nonlinear featurization of decision trees for linear regression modeling in the context of an on-line social network data may be implemented in the context of a network environment 100 illustrated in
As shown in
The client systems 110 and 120 may be capable of accessing the server system 140 via a communications network 130, utilizing, e.g., a browser application 112 executing on the client system 110, or a mobile application executing on the client system 120. The communications network 130 may be a public network (e.g., the Internet, a mobile communication network, or any other network capable of communicating digital data). As shown in
As shown in
Returning to
As mentioned above, the retrieval engine 210, in one example embodiment, uses a binary classifier in the form of a logistic regression model. The logistic regression model may be retrained using additional features obtained from one or more decision trees used by the ranker 220. Architecture of the recommendation system may utilize a converter configured to transform s-expressions into a decision tree that may that maybe used by the ranker 220 and also to convert a decision tree structure into s-expressions that can be used as additional features to train logistic regression models. An example architecture 500 that includes such converter is shown in
As shown in
An example recommendation system 144 of
The converter 620 may be further configured to access one or more business rules stored in the form of s-expressions, construct a decision tree based on these business rules stored in the form of s-expressions, and include the decision tree into the ranking model used by the learning to rank module 610. A business may be, e.g., related to a job title represented by a feature from a member profile maintained in the on-line social network system 142. Some operations performed by the system 600 may be described with reference to
As shown in
The various operations of example methods described herein may be performed, at least partially, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors may constitute processor-implemented modules that operate to perform one or more operations or functions. The modules referred to herein may, in some example embodiments, comprise processor-implemented modules.
Similarly, the methods described herein may be at least partially processor-implemented. For example, at least some of the operations of a method may be performed by one or more processors or processor-implemented modules. The performance of certain of the operations may be distributed among the one or more processors, not only residing within a single machine, but deployed across a number of machines. In some example embodiments, the processor or processors may be located in a single location (e.g., within a home environment, an office environment or as a server farm), while in other embodiments the processors may be distributed across a number of locations.
The example computer system 800 includes a processor 802 (e.g., a central processing unit (CPU), a graphics processing unit (GPU) or both), a main memory 804 and a static memory 806, which communicate with each other via a bus 808. The computer system 800 may further include a video display unit 810 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)). The computer system 800 also includes an alpha-numeric input device 812 (e.g., a keyboard), a user interface (UI) navigation device 814 (e.g., a cursor control device), a disk drive unit 816, a signal generation device 818 (e.g., a speaker) and a network interface device 820.
The disk drive unit 816 includes a machine-readable medium 822 on which is stored one or more sets of instructions and data structures (e.g., software 824) embodying or utilized by any one or more of the methodologies or functions described herein. The software 824 may also reside, completely or at least partially, within the main memory 804 and/or within the processor 802 during execution thereof by the computer system 800, with the main memory 804 and the processor 802 also constituting machine-readable media.
The software 824 may further be transmitted or received over a network 826 via the network interface device 820 utilizing any one of a number of well-known transfer protocols (e.g., Hyper Text Transfer Protocol (HTTP)).
While the machine-readable medium 822 is shown in an example embodiment to be a single medium, the term “machine-readable medium” should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, and/or associated caches and servers) that store the one or more sets of instructions. The term “machine-readable medium” shall also be taken to include any medium that is capable of storing and encoding a set of instructions for execution by the machine and that cause the machine to perform any one or more of the methodologies of embodiments of the present invention, or that is capable of storing and encoding data structures utilized by or associated with such a set of instructions. The term “machine-readable medium” shall accordingly be taken to include, but not be limited to, solid-state memories, optical and magnetic media. Such media may also include, without limitation, hard disks, floppy disks, flash memory cards, digital video disks, random access memory (RAMs), read only memory (ROMs), and the like.
The embodiments described herein may be implemented in an operating environment comprising software installed on a computer, in hardware, or in a combination of software and hardware. Such embodiments of the inventive subject matter may be referred to herein, individually or collectively, by the term “invention” merely for convenience and without intending to voluntarily limit the scope of this application to any single invention or inventive concept if more than one is, in fact, disclosed.
Certain embodiments are described herein as including logic or a number of components, modules, or mechanisms. Modules may constitute either software modules (e.g., code embodied (1) on a non-transitory machine-readable medium or (2) in a transmission signal) or hardware-implemented modules. A hardware-implemented module is tangible unit capable of performing certain operations and may be configured or arranged in a certain manner. In example embodiments, one or more computer systems (e.g., a standalone, client or server computer system) or one or more processors may be configured by software (e.g., an application or application portion) as a hardware-implemented module that operates to perform certain operations as described herein.
In various embodiments, a hardware-implemented module may be implemented mechanically or electronically. For example, a hardware-implemented module may comprise dedicated circuitry or logic that is permanently configured (e.g., as a special-purpose processor, such as a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC)) to perform certain operations. A hardware-implemented module may also comprise programmable logic or circuitry (e.g., as encompassed within a general-purpose processor or other programmable processor) that is temporarily configured by software to perform certain operations. It will be appreciated that the decision to implement a hardware-implemented module mechanically, in dedicated and permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software) may be driven by cost and time considerations.
Accordingly, the term “hardware-implemented module” should be understood to encompass a tangible entity, be that an entity that is physically constructed, permanently configured (e.g., hardwired) or temporarily or transitorily configured (e.g., programmed) to operate in a certain manner and/or to perform certain operations described herein. Considering embodiments in which hardware-implemented modules are temporarily configured (e.g., programmed), each of the hardware-implemented modules need not be configured or instantiated at any one instance in time. For example, where the hardware-implemented modules comprise a general-purpose processor configured using software, the general-purpose processor may be configured as respective different hardware-implemented modules at different times. Software may accordingly configure a processor, for example, to constitute a particular hardware-implemented module at one instance of time and to constitute a different hardware-implemented module at a different instance of time.
Hardware-implemented modules can provide information to, and receive information from, other hardware-implemented modules. Accordingly, the described hardware-implemented modules may be regarded as being communicatively coupled. Where multiple of such hardware-implemented modules exist contemporaneously, communications may be achieved through signal transmission (e.g., over appropriate circuits and buses) that connect the hardware-implemented modules. In embodiments in which multiple hardware-implemented modules are configured or instantiated at different times, communications between such hardware-implemented modules may be achieved, for example, through the storage and retrieval of information in memory structures to which the multiple hardware-implemented modules have access. For example, one hardware-implemented module may perform an operation, and store the output of that operation in a memory device to which it is communicatively coupled. A further hardware-implemented module may then, at a later time, access the memory device to retrieve and process the stored output. Hardware-implemented modules may also initiate communications with input or output devices, and can operate on a resource (e.g., a collection of information).
The various operations of example methods described herein may be performed, at least partially, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors may constitute processor-implemented modules that operate to perform one or more operations or functions. The modules referred to herein may, in some example embodiments, comprise processor-implemented modules.
Similarly, the methods described herein may be at least partially processor-implemented. For example, at least some of the operations of a method may be performed by one or processors or processor-implemented modules. The performance of certain of the operations may be distributed among the one or more processors, not only residing within a single machine, but deployed across a number of machines. In some example embodiments, the processor or processors may be located in a single location (e.g., within a home environment, an office environment or as a server farm), while in other embodiments the processors may be distributed across a number of locations.
The one or more processors may also operate to support performance of the relevant operations in a “cloud computing” environment or as a “software as a service” (SaaS). For example, at least some of the operations may be performed by a group of computers (as examples of machines including processors), these operations being accessible via a network (e.g., the Internet) and via one or more appropriate interfaces (e.g., Application Program Interfaces (APIs).)
Thus, nonlinear featurization of decision trees for linear regression modeling in the context of an on-line social network has been described. Although embodiments have been described with reference to specific example embodiments, it will be evident that various modifications and changes may be made to these embodiments without departing from the broader scope of the inventive subject matter. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense.
