Global vector recommendation based on implicit interaction and profile data

Generating feature vector representations and feature similarity matrix through feature word embedding models solves the problem that traditional recommendation systems cannot handle cold users and implicit interactive data, and achieves more efficient and relevant recommendation results.

CN110020188BActive Publication Date: 2024-12-03ADOBE INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201810621817.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2017-10-17
Filing Date
2018-06-15
Publication Date
2024-12-03
Estimated Expiration
2038-06-15

AI Technical Summary

Technical Problem

Traditional recommendation systems cannot generate recommendations for users who do not have historical explicit interaction data, and cannot consider implicit interaction data between user profiles and projects, resulting in unrelated recommendation results.

Method used

Using a vector model based on feature word embedding, convert user profiles, projects and interaction data into words through natural language processing, generate explicit and implicit interactive sentences, and use the word embedding model to generate feature vector representations to construct feature similarity Matrix to provide customized recommendations.

Benefits of technology

It can provide customized recommendations for cold user profiles and improve the relevance and accuracy of recommendations, taking into account the implicit interactive data between user profiles and projects, providing higher recommendation accuracy than traditional technologies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN110020188B_ABST
    Figure CN110020188B_ABST
Patent Text Reader

Abstract

A digital media environment for promoting recommendations based on vectors generated using feature word embeddings is described. A recommendation system receives data describing at least one attribute of a user profile, at least one item, and an interaction between the user profile and the at least one item. The recommendation system uses natural language processing to associate each user profile attribute, each item, and each interaction between the user profile and the item as words, and combines the words into sentences. These sentences are input into a word embedding model to determine a feature vector representation, which describes the relationship between the user profile attributes, the items, and the explicit and implicit interactions. From the feature vector representation, the recommendation system determines the similarity between different features. Therefore, even for user profiles that are not associated with any historical interaction data, the recommendation system can provide customized recommendations based on implicit interactions.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0001] The computing device outputs recommendations that expose items that may be of interest to the user, even if the user is unaware of the items' existence. For example, a video streaming service may output movie or television show recommendations, an online store may output product recommendations, and so on. Many services have replaced traditional search engines with automated recommendations as the primary means for discovering content items. Thus, the system strives to tailor recommendations to individual end users so that the recommended items are actually of interest.

[0002] Conventional recommender systems utilize historical data describing explicit interactions between user profiles and items. Explicit interactions include item view-view relationships, which describe items that were viewed together in a single browsing session. Other explicit interactions include view-purchase relationships, which describe at least one item that was purchased after being viewed. Alternatively or additionally, explicit interactions include item purchase-purchase relationships, which describe items that were purchased together in a single browsing session. Conventional solutions build models from these explicit interaction data and use these models to determine recommendations for user profiles.

[0003] However, traditional recommendation systems cannot generate recommendations for user profiles that do not have historical explicit interaction data. Similarly, traditional recommendation systems cannot take into account implicit interactions between user profiles and items. Implicit interactions refer to interactions that do not explicitly indicate the relevance of certain items. For example, implicit interactions can include the number of times an item has been viewed, the amount of time spent reading a news article, the percentage of videos watched, and so on. Therefore, traditional schemes for digital recommendation cannot take into account important interaction data that would otherwise affect the recommendation decision. Therefore, traditional schemes for automatic item recommendation are limited to considering only certain types of interaction data, which often results in outputting irrelevant recommendations. Summary of the invention

[0004] A digital media environment is described to facilitate recommendations based on vectors generated using feature word embeddings. In one example, a computing device implementing a recommender system receives data describing at least one attribute of a user profile, at least one item, and an interaction between the user profile and the at least one item. Continuing with the example, the recommender system receives different sets of session data describing the user profile attributes, items, and interactions between the user profile and the items over a period of time (e.g., over a single browsing session). After receiving the data, the recommender system associates each user profile attribute, each item, and each interaction between the user profile and the item into terms using natural language processing.

[0005] From these words, the recommendation system compiles the words associated with the conversation into sentences so that the sentences include information describing the user profile, the items, and the interactions between the user profile and the items during the course of the conversation. Thus, the recommendation system generates sentences that represent individual conversations. In addition, the recommendation system distinguishes between explicit and implicit interactions between user profiles and items by generating separate sentences that describe explicit and implicit interactions for a single conversation. For example, explicit interactions refer to explicit correlations between user profiles and items, such as ratings assigned to items by user profiles, indications that user profiles purchased items, and the like. In contrast, implicit interactions refer to interactions that are not directly related to explicit correlations between user profiles and items, such as the number of times an item is viewed, the amount of time spent interacting with an item, the percentage of items consumed by a user profile, and the like.

[0006] Each sentence generated by the recommendation system is then input into a word embedding model to determine a feature vector representation that describes the similarity between user profile attributes, items, explicit interactions, and implicit interactions described in the received data. In some implementations, the word embedding model utilizes an objective function that includes data fields designed to consider implicit interactions between items and user profiles. Therefore, the recommendation system generates a feature vector representation that includes information describing user profiles, items, and interactions in a common framework. From the feature vector representation, the recommendation system can determine the similarity between different features. For example, the recommendation system can determine the similarity between user profile attributes and items based on the feature vector representations corresponding to the user profile attributes and the items. In some implementations, feature similarity is determined based on the dot product of the corresponding feature vectors.

[0007] The recommender system uses the feature vector representation to generate a feature similarity matrix, and stores both the feature similarity matrix and the feature vector representation for generating recommendations. In this way, when the recommender system receives a request for accessing an item from a user profile, information describing the user profile can be used together with the feature vector representation and the feature similarity matrix to recommend items that have relevance to the user profile. Therefore, the recommender system can provide customized recommendations even for cold user profiles that are not associated with any historical interaction data. In addition, due to the inclusion of implicit interaction data, the recommender system can identify the relevance between various features described in the feature similarity matrix with higher accuracy than traditional techniques.

[0008] This Summary introduces a selection of concepts in a simplified form that are further described in the Detailed Description below. As such, this Summary is not intended to identify essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The detailed description is described with reference to the accompanying drawings. Entities shown in the drawings may refer to one or more entities, and thus reference to the entities in singular or plural forms may be made interchangeably in the discussion.

[0010] Figure 1 is an illustration of an environment in an example implementation that is operable to employ global vector recommendations using the techniques described herein.

[0011] Figure 2 A more detailed description is given showing Figure 1 Operation of a recommender system is described in an example implementation of the system.

[0012] Figure 3 Depicted is an example implementation showing feature vector generation using word embeddings.

[0013] Figure 4 Depicted is an example implementation showing feature vector generation using word embeddings.

[0014] Figure 5 is a flowchart depicting an example process in accordance with one or more embodiments.

[0015] Figure 6 is a flowchart depicting an example process in accordance with one or more embodiments.

[0016] Figure 7 An example system is shown that includes various components of an example device that can be implemented as described in reference Figure 1-6 Any type of computing device described and / or utilized to implement embodiments of the techniques described herein. DETAILED DESCRIPTION

[0017] Overview

[0018] This article describes techniques and systems for implementing a word embedding model through at least one computing device to determine vector representations of various features (such as user profile identifiers, user profile attributes, item identifiers, and interactions between user profiles and items). By generating vector representations that include information describing these different features in a common framework, the techniques described herein accurately identify similarities between different features as a function of the corresponding vector representations.

[0019] The techniques described herein are advantageous over conventional approaches to generating recommendations because they consider information describing a user profile and implicit and explicit interactions between the user profile and items. In contrast, conventional techniques generate recommendations based solely on item co-occurrence models and are unable to consider various user profile attributes or interaction data to customize recommendations. Thus, the techniques described herein can provide customized recommendations even for cold user profiles that are not associated with any interaction data.

[0020] To ensure that recommendations are appropriately tailored to specific features, the recommendation system uses natural language processing techniques to model user profile behavior across websites as an ongoing conversation. Therefore, the recommendation system classifies data describing user profiles, items, and interactions between user profiles and items into words. These words are then combined into sentences, each sentence representing an individual session, such as a browsing session. Multiple sentences are then combined into a unified behavioral corpus for user profile interactions, which is used to generate vector representations for various features including user profiles, items, user profile attributes, and interaction data.

[0021] The recommendation system then uses the feature vector representations to generate a feature similarity matrix that can be used to define relationships between features. For example, the recommendation system can generate a feature similarity matrix based on the dot product of the corresponding feature vector representations. In this example, a high result value from the dot product corresponds to a high correlation between the features, and vice versa. Therefore, the technology described herein is configured to utilize the feature similarity matrix to generate feature recommendations, such as recommendations for items that have a high correlation with user profile attributes.

[0022] In the following discussion, an example environment is described in which the techniques described herein may be employed. Example procedures are then described that may be performed in the example environment as well as other environments. Thus, the execution of the example procedures is not limited to the example environment, and the example environment is not limited to the execution of the example procedures.

[0023] the term

[0024] The term "session" refers to a discrete period of time during which a user profile interacts with one or more items. For example, a session may refer to a browsing session of a website that begins when a user profile visits the website and ends when the user profile navigates away from the website. Alternatively or additionally, a session may be a period of no inactivity for 30 minutes in a shopping environment, no inactivity for 12 hours for a streaming environment, and so forth.

[0025] The term "explicit interaction" refers to an explicit correlation between a user profile and an item. For example, an explicit interaction may describe a rating assigned to an item by a user profile, an indication that the user profile purchased the item, and the like.

[0026] The term "implicit interactions" refers to interactions that are not directly related to the displayed relevance between a user profile and an item. For example, implicit interactions may describe the number of times an item has been viewed, the amount of time spent interacting with an item, the percentage of an item consumed by a user profile, and so on.

[0027] The term "word embedding model" refers to a mathematical model that can be used by a computing device to determine a vector representation of features based on an input sentence that includes information describing activities that occurred during a session. As described herein, the word embedding model includes an objective function that includes data fields for considering implicit interactions between user profiles and items.

[0028] The term "feature vector representation" or "feature vector" refers to a representation in a vector space that includes information describing one or more of a user profile identifier, a user profile attribute, an item identifier, data describing explicit interactions between a user profile and an item, data describing implicit interactions between a user profile and an item, or a combination thereof. The feature vector representation as described herein describes these various features in a general framework, thereby enabling comparison of feature relationships based on functions of corresponding feature vectors.

[0029] The term "feature similarity matrix" refers to a database structure that mathematically describes the relationship between different features. The feature similarity matrix advantageously describes different features in a general framework to capture the similarities between different features.

[0030] The term "cold user profile" refers to a user profile for which the recommender system is unaware of any data describing the user profile's previous interactions. For example, a cold user profile is contrasted with a warm user profile, for which the recommender system is aware of data describing historical interactions between the warm user profile and items.

[0031] Example Environment

[0032] Figure 1 1 is a diagram of an environment 100 in an example implementation that is operable to employ feature word embeddings for global vector recommendations using the techniques described herein. The illustrated environment 100 includes a service manager system 102, a recommendation system 104, and user interactions 106 (also referred to herein as users) with corresponding computing devices as a source of session data 108 transmitted via a communication system 110. Each individual computing device 106 as described herein can be associated with multiple user profiles. For example, a computing device can be shared by a family, with each family member having a corresponding user profile. In addition, individuals can use different user profiles for work-related and personal computing device interactions. Therefore, a user as described herein refers to an individual user or a profile of an individual user.

[0033] The service manager system 102, the recommendation system 104, and the communication system 110 are implemented at least in part in the hardware of one or more computing devices. For example, each system 102, 104, and 110 may be implemented in a computer-readable storage medium to be executed by a processing system of a computing device. Examples of processing systems and computer-readable storage media are described below. Figure 7 Further detailed description. In some implementations, the service manager system 102, the recommendation system 104, and the communication system 110 are implemented at a single computing device. Alternatively or additionally, the service manager system 102, the recommendation system 104, and the communication system 110 are implemented between multiple computing devices that are communicatively connected to each other via a network 112.

[0034] For example, a computing device implementing one or more of the systems described herein may be configured as a desktop computer, a laptop computer, a mobile device (e.g., assuming a handheld configuration such as a tablet or mobile phone as shown), etc. Thus, the computing devices referred to herein may range from full-resource devices with large amounts of memory and processor resources (e.g., personal computers, game consoles) to low-resource devices with limited memory and / or processing resources (e.g., mobile devices). Alternatively or additionally, a computing device may refer to a plurality of different devices, such as a plurality of servers utilized by an enterprise to perform operations "in the cloud," as described with respect to Figure 7 Further described.

[0035] The service manager system 102 represents functionality implemented at least in part in the hardware of a computing device for providing services accessible via a network 112 by users 106 via a corresponding computing device. For example, the service manager system 102 may expose a website 114 or other functionality accessible via the network 112 to expose items 116. Thus, the service manager system 102 is configured to expose items 116 for user interaction.

[0036] The communication system 110 represents functionality implemented at least in part in the hardware of a computing device for enabling access to items 116 disclosed by the service manager system 102. For example, the communication system 110 enables access to the service manager system 102 through a web browser, an application, etc. In this manner, the communication system 110 enables a computing device to obtain data from the service manager system 102 via the network 112. In some implementations, the communication system 110 employs the recommendation system 104 to provide data obtained from the service manager system 102 to the user 106. Alternatively or additionally, the service manager system 102 employs the recommendation system 104 to provide data to the user 106 via the communication system 110.

[0037] Recommendation system 104 represents functionality implemented at least in part in the hardware of a computing device for generating recommendations for user 106 via communication system 110. To generate recommendations, recommendation system 104 uses feature similarity matrix 118, which may be stored at storage device 120 of the computing device, for example, as described with respect to Figure 7 The feature similarity matrix 118 is a data structure that can be used by the recommender system 104 to generate recommendations, which is described in further detail below.

[0038] Thus, the service manager system 102, the recommendation system 104, and the communication system 110, the user profile interaction 106 with the corresponding computing device, and the storage device 120 are implemented at least in part in the hardware of one or more computing devices and may reside on a computer-readable storage medium for execution by a corresponding processing system. Examples of processing systems and computer-readable media may be found at Figure 7 Examples of computing systems can be found in the discussion of FIG.

[0039] Figure 2 A more detailed description is given showing Figure 1 200 in an example implementation of the operation of the recommender system 104 of the present invention. As shown, the recommender system 104 generates a feature similarity matrix 118 based on the session data 108. Using the feature similarity matrix 118, the recommender system 104 then provides feature recommendations 232 to the user 106. The recommender system 104 receives the session data 108 from a respective computing device implementing a user profile of the user 106. As described herein, the session data 108 represents data describing at least one of a user profile identifier, a user profile attribute, an item identifier, or an interaction between a user profile and an item. In some implementations, using the techniques described herein, the interactions described by the session data 108 can be explicit interactions, implicit interactions, or a combination thereof.

[0040] In the illustrated example, session data 108 is received from session 204 for user profile 206 and from session 216 for user profile 218. In some implementations, user profile 206 identifies an individual user that is different from user profile 218. Alternatively, user profiles 206 and 218 may identify a common individual user. For example, user profile 206 may correspond to a profile for use on an individual mobile device, while user profile 218 may correspond to a profile for use on an individual laptop computer. Thus, the recommender system 104 is configured to receive session data 108 describing multiple sessions for one or more user profiles. The received session data 108 may be stored by the recommender system 104 in a storage medium such as a computer. Figure 1 The data is stored in the storage device 120 for further use.

[0041] For each session, session data 108 includes information describing at least one of a user profile identifier, profile attributes and corresponding attribute values, explicit interaction data for an item, implicit interaction data for an item, or a combination thereof. For example, session 204 includes information describing a user identifier 208 for user profile 206. User identifier 208 may be any suitable form of information for uniquely identifying a user profile, such as a user name, an employee identification number, and the like.

[0042] The session 204 additionally includes information describing at least one profile attribute 210 and a corresponding attribute value 212 for the user profile 206. For example, an example profile attribute 210 is "city," where the corresponding attribute value 212 may specify "New York." Alternatively or additionally, the profile attribute 210 may be "gender," where the corresponding attribute value 212 specifies "female." Thus, the profile attributes 210 and attribute values ​​212 included in the session data 108 include information suitable for identifying various descriptive characteristics of the user profile 206. Using the techniques described herein, the profile attributes 210 may include any information suitable for describing the user profile 206, such as occupation, discipline, industry, country, region, experience, education, and the like.

[0043] The session 204 also includes at least one item identifier 214 that describes the item with which the user profile 206 has interacted. Figure 3 and 4 As further described, the project identifier 214 additionally includes information describing explicit interaction data, implicit interaction data, or a combination thereof for the identified project. For example, the project identifier 214 may include a project description, a project title, project metadata tags, search terms associated with the project, and the like.

[0044] Additionally, the item identifier 214 may include information describing explicit and implicit interactions between the user profile 206 and the corresponding item. For example, the item identifier 214 may include information describing the type of interaction, such as viewed, purchased, a timestamp for the interaction, a percentage of the item visited, an amount of time spent interacting with the item, etc. Thus, the item identifier 214 not only identifies a specific item with which the user profile 206 interacted, but also includes information specifically describing how the user profile 206 interacted with the specific item.

[0045] Similarly, session 216 includes information describing a user identifier 220 of user profile 218. User identifier 220 may be any suitable form of information for uniquely identifying a user profile, such as a username, employee identification number, etc. Session 216 additionally includes information describing at least one profile attribute 222 and a corresponding attribute value 224 for user profile 218. For example, example profile attribute 222 is "country," where corresponding attribute value 224 specifies "India."

[0046] Alternatively or additionally, the profile attribute 222 may specify "gender," with the corresponding attribute value 224 specifying "male." Thus, the profile attributes 222 and attribute values ​​224 included in the session data 108 describe information that may be used to identify various characteristics of the user profile 218. Using the techniques described herein, the profile attributes 222 may include any information suitable for describing the user profile 224, such as occupation, discipline, industry, country, region, experience, education, and the like.

[0047] The session 216 additionally includes at least one item identifier 226 that describes an item with which the user profile 218 has interacted. Figure 3 and Figure 4 As further described, the project identifier 226 additionally includes information describing explicit interaction data, implicit interaction data, or a combination thereof for the identified project. For example, the project identifier 226 may include a project description, a project title, project metadata tags, search terms associated with the project, and the like.

[0048] Additionally, item identifiers 226 may include information describing explicit and implicit interactions between user profile 218 and the corresponding item. For example, item identifiers 226 may include information describing the type of interaction (such as viewed, purchased, timestamp for interaction, percentage of item visited, amount of time spent interacting with the item, etc.). Thus, item identifiers 226 not only identify a specific item with which user profile 218 interacted, but also include information specifically describing how user profile 218 interacted with the specific item.

[0049] Although item identifiers 214 and 226 are illustrated as being included in session data 108, in some implementations, the recommender system 104 is not aware of any item identifiers or item interaction data for a given user profile. For example, in an example where user profile 206 is a cold user profile, session data 108 for session 204 may include only the user identifier, profile attributes 210, and attribute values ​​212, or a combination thereof. However, even without receiving information corresponding to item identifier 214, the recommender system may still use feature similarity matrix 118 to generate feature recommendations 232 for user profile 206.

[0050] To generate the feature similarity matrix 118, the recommendation system 104 applies the word embedding model 228 to the conversation data 108. The word embedding model 228 is configured to implement natural language processing to generate feature vector representations for various information included in the conversation data 108. The word embedding model 228 can be configured in various ways, such as a modified global vector for the word representation model, which is described below with respect to Figure 3 and 4 Discussed in further detail. From the feature vector representation, the recommender system generates a feature similarity matrix 118. In implementations, the generation of the feature similarity matrix 118 is performed using feature vector representation dot product comparisons, feature vector representation arithmetic, and the like.

[0051] The recommender system 104 is configured to provide feature recommendations 232 based on new session data 230. The new session data 230 represents information describing a session for a user profile that was not previously described in the session data 108 used to generate the feature similarity matrix 118. For example, the new session data 230 may include information describing a user identifier of the user profile. Alternatively or additionally, the new session data 230 may include information describing profile attributes of the user profile and corresponding attribute values ​​of the profile attributes. Alternatively or additionally, the new session data 230 may include information describing items and interactions between items and the user profile during the session.

[0052] For example, if the new session data 230 describes that the user profile is currently viewing item "A", the recommendation system 104 is configured to employ the feature similarity matrix 118 to identify similar items based on the features of the items to output feature recommendations 232 for the similar items. Similarly, the recommendation system 104 is configured to employ the feature similarity matrix 118 to identify items having relevance to the user profile based on the user profile identifier or the user profile attribute value.

[0053] Because the feature similarity matrix 118 includes information describing user profile identifiers, user profile attribute values, item identifiers, and item interaction data, any suitable features may be used to determine user profile-item relevance to generate feature recommendations 232. This enables the recommender system 104 to tailor feature recommendations 232 even for cold user profiles that are not associated with any item interaction information describing previous sessions of the cold user profile. For example, the recommender system 104 may identify that the user profile corresponds to Las Vegas and use the feature similarity matrix 118 to identify items having relevance to Las Vegas-affiliated user profiles to include in the feature recommendations 232.

[0054] In implementations where the feature similarity matrix 118 does not identify similar items for inclusion in the feature recommendations 232 for the new session data 230 , the recommender system 104 may provide a generic feature recommendation 232 or forgo providing a feature recommendation 232 .

[0055] Conversely, in scenarios where the recommender system 104 identifies multiple similar items, the recommender system 104 applies a weighting function to identify the items from the feature similarity matrix 118 that are most similar to the new session data 230 for inclusion in the feature recommendations 232. Additional factors such as item ratings, item tags, etc. may additionally be used by the recommender system 104 to determine the order of feature recommendations 232 for delivery to the user profile.

[0056] Figure 3 The depiction shows that the recommendation system 104 is based on Figure 2 Example implementation 300 of a word embedding model 228 applied to session data to generate a feature vector 314. In the illustrated example, the session data is represented as describing sessions 204 and 216, such as Figure 2 . In this example, session 204 includes information describing a user profile identifier 208 that specifies that session 204 corresponds to "User 1." Similarly, session 204 includes information describing a gender profile attribute 210 that specifies that the user profile for session 204 corresponds to the "Male" gender. Session 204 additionally includes a city profile attribute 302 that specifies that the user profile for session 204 corresponds to the "New York" city. Finally, session 204 includes an item identifier 214 that indicates that the user profile interacted with items "I1, I3, I5, and I7" during the session.

[0057] Although not illustrated, item identifier 214 may additionally include information describing interaction information between the user profile and individual items in items I1, I3, I5, and I7 for session 204. For example, the information specified by item identifier 214 may indicate that item I3 was viewed multiple times, item I5 was purchased, only 15% of item I7 was displayed by the computing device for the user profile, and so on. Thus, the information specified by item identifier 214 may specifically describe how the user profile interacted with various items during session 204, both explicitly and implicitly. Figure 4 Includes additional discussion of interaction information specified by the session data.

[0058] In the illustrated example, session 216 includes information describing a user profile identifier 220 that specifies that session 216 corresponds to "User 2." Similarly, session 216 includes information describing a gender profile attribute 222 that specifies that the user profile for session 216 corresponds to the "Female" gender. Session 216 additionally includes a city profile attribute 306 that specifies that the user profile for session 216 corresponds to the "New York" city. Finally, session 216 includes an item identifier 226 that indicates that during the session the user profile interacted with items "I1, I2, I3, I5, I7, and I8."

[0059] Although not shown, item identifier 226 may additionally include information describing interaction information between the user profile and individual items in items I1, I2, I3, I5, I7, and I8 for session 216. For example, the information specified by item identifier 226 may indicate that item I2 was viewed multiple times, item I1 was purchased, 100% of item I7 was displayed by the computing device for the user profile, etc. Thus, the information specified by item identifier 226 may specifically describe how the user profile interacted with various items during session 216, both explicitly and implicitly.

[0060] The conversation data from the conversations 204 and 216 are then fed to the word embedding model 228 to generate the feature vector 314. The word embedding model 228, as described herein, characterizes individual aspects of the conversation data as words and combines the various words to form sentences describing the respective conversations 204 and 216. For example, the word embedding model 228 generates a sentence 310 from the conversation 204. In the illustrated example, the sentence 310 includes the individual words "I1", "I3", "I5", "I7", "gender-male", "city-New York", and "user 1", which succinctly describe the conversation data of the conversation 204. Similarly, the sentence 312 includes the individual words "I1", "I2", "I3", "I7", "I8", "gender-female", "city-New York", and "user 2", which succinctly describe the conversation data of the conversation 216.

[0061] From these sentences, feature vectors 314 may be generated for each distinct word included in the sentences generated by the word embedding model 228. For example, a feature vector for "City - New York" may be generated to describe items having relevance to a user profile associated with New York City. In this manner, feature vectors 314 may be used to identify item relevance to a user profile that is not associated with any prior explicit or implicit interaction data. To account for implicit interaction data, the recommender system 104 is configured to generate multiple sentences for a single session.

[0062] Figure 4 Depicted is a diagram showing that a recommendation system based on Figure 2 4. Example implementation 400 of applying the word embedding model 228 to session data to generate a feature vector 422. In the example shown, the session data is represented as describing a session 402. The session 402 includes information describing a user profile identifier 404, which specifies that the session 402 corresponds to "User 3". Similarly, the session 402 includes information describing a gender profile attribute 406, which specifies that the user profile for the session 402 corresponds to the "female" gender. The session 402 additionally includes a city profile attribute 410, which specifies that the user profile for the session 402 corresponds to the "Las Vegas" city. Finally, the session 402 includes an item identifier 414, which indicates that the user profile interacts with the items "I1, I4, I7, and I3".

[0063] The information specified by the project identifier 414 is illustrated as including interaction information that describes the interaction between the corresponding project and the user profile for the session 402. For example, the project identifier includes three separate listings of the project "I4". In some implementations, this may indicate that the user profile viewed the project I4 at three separate times during the session 402, while the projects I1, I3, and I7 were each viewed only once during the session 402. Although described herein in terms of multiple views, the interaction information included in the project identifier 414 may include any type of information describing the explicit or implicit interaction between the user profile and the project. For example, the project identifier 414 may specify the number of project segments consumed by the computing device, the amount of time spent interacting with a particular project, and so on. Therefore, the information included in the project identifier 414 specifically describes how the user profile interacts with various projects during the session 404.

[0064] The recommendation system 104 then applies the word embedding model 228 to the conversation data from the conversation 402 to generate a feature vector 422. As described herein, the word embedding model 228 characterizes individual aspects of the conversation data as words and combines the various words to form sentences describing the conversation 402. For example, the recommendation system 104 can generate separate sentences for explicit interaction data and implicit interaction data. In the example shown, sentences 416, 418, and 420 are generated from a single conversation to consider implicit interaction data related to item I4. Sentence 416 includes individual words "I1", "I4", "I7", "I3", "gender-female", "city-Las Vegas", and "user 3", which describe the explicit interaction between the item and the user profile corresponding to user 3. In contrast, sentences 418 and 420 include individual words "I4", "gender-female", "city-Las Vegas", and "user 3".

[0065] Thus, sentences 418 and 420 represent implicit interaction data, such as item I4 being viewed more frequently than items I1, I3, and I7 during session 402. In this manner, the recommender system 104 is configured to take into account implicit interaction data that would not otherwise be conveyed by generating a single sentence for a session. From the sentences generated by the recommender system 104, correlations between various features, such as items and user profile attributes, can be determined as a function of their corresponding feature vectors 422. To take into account such implicit interaction data, the word embedding model 228 implements a modified objective function.

[0066] The word embedding model 228 represents a modified global vector (GloVe) model, which is a global log-bilinear regression model that is configured to use a word-word co-occurrence matrix together with a local context window to generate word embeddings in a low-dimensional space. These word embeddings can then be used by a computing device implementing the recommendation system 104 to generate feature vector representations of associated items, profile attribute values, interaction data, and the like. Although described herein in terms of using a word-to-word co-occurrence matrix in which conversation data is represented as words or sentences, an item-item co-occurrence matrix can also be used to directly generate item embeddings in a low-dimensional space.

[0067] The word embedding model 228 is therefore configured to implement a modified GloVe model and exploit statistical information by training only the non-zero elements of the word-to-word co-occurrence matrix. By training only the non-zero elements, rather than training the entire sparse matrix of individual context windows on a large corpus, the word embedding model 228 is configured to quickly generate feature vectors 314 for providing recommendations in real time for the current state of the user profile. The representation reasoning process employed by the word embedding model 228 considers the distance between words as a representation of similarity or the relevance between words. For example, items that are interacted with continuously in a single session may be more similar than items that are interacted with through different, non-continuous sessions. In addition to capturing the item session similarity, the word embedding model 228 is configured to take into account the implicit interactions between items and the user profile using a modified objective function of the GloVe model represented by Equation 1:

[0068]

[0069] In Equation 1, C ijRepresent a feature similarity matrix 118 that contains session data 108, where each row can correspond to a user profile attribute and each column corresponds to an item. Alternatively or additionally, the rows can correspond to items and the columns can correspond to user profiles or a combination thereof. In this way, the modified objective function of Equation 1 is able to take into account implicit interaction data describing implicit correlations between items and user profiles. Equation 1 additionally includes a confidence factor α that represents the confidence of the value of the implicit user profile and item correlation. For example, the higher the value of α, the greater the weight that will be given to the implicit user profile correlation. In some implementations, α represents a value of 0.1. However, any suitable value between 0 and 1 may be used in Equation 1.

[0070] Therefore, the recommendation system 104 generates a feature vector 314, which includes information describing a user profile identifier, a user profile attribute value, an item identifier, an explicit interaction between a user profile and an item, an implicit interaction between a user profile and an item, or a combination thereof. In this way, even if there is no previous interaction data stored for the user profile, the feature vector 314 can be used to identify items similar to the user profile, so that the recommendation system 104 can generate customized feature recommendations even for cold users. Although described as identifying items that may be of interest to a user profile, the techniques described herein can similarly be used to identify user profiles that may be relevant to an item based on information describing a particular item. In addition, because the recommendation system 104 embeds user profiles, items, and user profile attributes into a single vector space, it can be used to provide customized recommendations in real time across different domains.

[0071] Sample Program

[0072] The following discussion describes techniques that can be implemented using the previously described systems and devices. Aspects of each process can be implemented in hardware, firmware, software, or a combination thereof. These processes are shown as a collection of blocks that specify operations performed by one or more devices and are not necessarily limited to the order shown for performing the operations of the corresponding blocks. In portions of the following discussion, reference will be made to Figure 1-4 .

[0073] Figure 5 A process 500 in an example implementation of feature word embedding for global vector recommendation is depicted. Data describing a user profile, attributes of the user profile, items, and interactions between the user profile and the items is received (block 502). For example, a computing device implementing the recommendation system 104 receives session data 108 from the communication system 110. As described herein and Figure 2-4As shown in , the received data is organized into individual sessions for corresponding user profiles by the recommendation system. In some implementations, the session data may include information describing the user profile, user profile attributes, items, or interactions between the user profile and items, or a combination thereof. For example, the received session data may describe that the user profile corresponds to "User 3", the gender attribute for the user profile is "Female", the city attribute for the user profile is "Las Vegas", and the user profile interacts with items "I1, I4, I7, and I3".

[0074] In response to receiving the conversation data, at least one sentence is generated, the at least one sentence including words that identify a user profile, user profile attributes, items, and interactions between the user profile and the items (box 504). For example, a first sentence may be generated from the exemplary received conversation data, which reads "I1, I4, I7, I3, gender-female, city-Las Vegas, user 3". In some implementations, the sentence represents a word embedding description of explicit interaction data for a particular conversation. Additionally, a second sentence and a third sentence may be generated from the example received conversation data to read "I4, gender-female, city-Las Vegas, user 3". In this example, the second sentence and the third sentence represent implicit interaction data that would not otherwise be quantified by traditional techniques that simply describe item-item conversation correspondences. Additionally, each generated sentence includes features that describe the user profile as well as the items and corresponding interactions of a single sentence. Therefore, the sentences generated by the recommendation system 104 can be used to identify feature similarities across similar and different feature categories.

[0075] Subsequently, at least one feature vector representation is determined based on the at least one generated sentence (block 506). In implementation, the feature vector representation is generated by a word embedding model such as the word embedding model 228 of the recommendation system 104. For example, the word embedding model 228 can implement a GloVe model with a modified objective function, as described above with respect to Equation 1. In this manner, the feature vector representation includes information describing the user profile, the item, and the corresponding explicit and implicit interactions between the user profile and the item.

[0076] Finally, a feature similarity matrix is ​​output based on at least one feature vector representation (block 508). In implementation, the feature similarity matrix 118 is output by the recommendation system 104 based on a function of the determined feature vector representation, such as feature vectors 314 and 422, as shown in FIG. Figure 3 and Figure 4 For example, dot products of feature vectors may be calculated to determine feature similarities between various features described in the feature similarity matrix 118. In this manner, features having higher dot product values ​​from their corresponding feature vectors are associated with a higher degree of similarity than features having lower dot product values.

[0077] Therefore, the feature similarity matrix and the corresponding feature vector representation can be stored such as in Figure 1 The feature vector representation and feature similarity matrix may be indexed into a search engine even for cold user profiles for real-time delivery of customized recommendations in a distributed computing environment.

[0078] Figure 6 A process 600 in an example implementation of feature word embedding for global vector recommendation is depicted. Data describing a user profile is received (block 602). For example, a computing device implementing the recommendation system 104 receives new session data 230 describing a user profile for a session via a network 112. For example, the received data may describe an identifier of the user profile or at least one user profile attribute value or a combination thereof. In some implementations, such as Figure 1 As shown, data describing a user profile is received by the recommender system 104 from the service manager system 102 via the network 112 as part of a query for a website 114 or an item 116 .

[0079] Data describing the interactions between the user profile and the items is also received (block 604). This data describing the interactions between the user profile and the items is optionally received, as indicated by the arrow circumventing block 604. For example, in the case of a cold user profile, no data describing the interactions between the cold user profile and the items is available and is therefore not received by the recommender system 104. For example, the data describing the interactions may be received by the recommender system 104 via the network 112 as part of the new session data 230.

[0080] Alternatively or additionally, data describing the interaction may be received from storage 120, such as Figure 1 As shown. In implementations, the data describing the interaction between the user profile and the item may refer to historical interactions from previous sessions between the user profile and the item. Additionally or alternatively, the data describing the interaction between the user profile and the item may refer to items accessed by the user profile during an ongoing session. Thus, the recommendation system 104 is configured to provide recommendations that reflect the current state of the user profile in real time.

[0081] In response to receiving data describing a user profile and optionally receiving data describing interactions between the user profile and items, at least one item recommendation for the user profile is determined based on the feature similarity matrix (block 606). For example, the recommendation system 104 can employ the feature similarity matrix 118 to generate feature recommendations 232 for the user profile based on the new session data 230. The feature similarity matrix 118 can be used by the recommendation system 104 to generate feature recommendations 232 for the user profile based on the new session data 230. Figure 1-5 Describe the technique to generate.

[0082] In response to determining multiple item recommendations for the user profile based on the feature similarity matrix, the item recommendations are ranked (block 608). The recommendation system 104 may use any suitable metric to determine the ranking of the multiple recommendations. For example, the multiple feature recommendations 232 may be ranked based on timestamps corresponding to interactions between the user profile and the item, based on explicit interaction data such as item rankings, or based on implicit interaction data such as the amount of time spent interacting with the item, etc. In some implementations, the recommendation system 104 performs the ranking by adjusting the alpha value used in the modified objective function described with respect to Equation 1. In the case where the recommendation system 104 generates only a single item recommendation, the single item recommendation is considered the highest ranked recommendation.

[0083] In response to ranking the recommendations, item recommendations are output based on the rankings (block 610). For example, the recommender system 104 may transmit the one or more feature recommendations 232 to the communication system 110 via the network 112 for output at a computing device corresponding to one of the user profiles 106. In this manner, the recommender system 104 is configured to generate recommendations in real time based on vector representations that describe user profiles, user profile attributes, items, explicit interactions, and implicit interactions in a general framework.

[0084] Therefore, the recommendation system 104 is configured to output recommendations based on a function describing a vector representation of features representing user profiles, items, and corresponding interactions. By describing explicit interactions, implicit interactions, and user profile attributes in a common framework, the recommendation system 104 can output customized recommendations for cold user profiles that have not previously interfaced with the recommendation system 104, so that different cold user profiles receive different recommendations. In addition, the recommendation system 104 takes into account additional features that are ignored by traditional recommendation techniques, thereby providing recommendations with a high degree of relevance for a given user profile.

[0085] Example Systems and Devices

[0086] Figure 7 An example system, generally at 700, is shown including an example computing device 702, which is representative of one or more computing systems and / or devices that may implement the various techniques described herein. This is illustrated by the inclusion of the recommendation system 104. The computing device 702 may be, for example, a server of a service provider, a device associated with a client (e.g., a client device), a system on a chip, and / or any other suitable computing device or computing system.

[0087] The exemplary computing device 702 shown in the figure includes a processing system 704, one or more computer-readable media 706, and one or more I / O interfaces 708 coupled to each other in communication. Although not shown, the computing device 702 may also include a system bus or other data and command transmission systems that couple various components to each other. The system bus may include any one or combination of different bus structures, such as a memory bus or memory controller, a peripheral bus, a universal serial bus, and / or a processor or local bus utilizing any of a variety of bus architectures. Various other examples are also contemplated, such as control and data lines.

[0088] Processing system 704 represents the functionality of performing one or more operations using hardware. Thus, processing system 704 is illustrated as including hardware elements 710, which may be configured as processors, functional blocks, and the like. This may include implementation of hardware as a dedicated integrated circuit or other logic device formed using one or more semiconductors. Hardware element 710 is not limited by the material from which it is formed or the processing mechanism employed therein. For example, a processor may include semiconductors and / or transistors (e.g., electronic integrated circuits (ICs)). In this case, the processor executable instructions may be electronic executable instructions.

[0089] Computer-readable storage media 706 is illustrated as including memory / storage 712. Memory / storage 712 represents memory / storage capacity associated with one or more computer-readable media. Memory / storage 712 may include volatile media (such as random access memory (RAM)) and / or non-volatile media (such as read-only memory (ROM), flash memory, optical disks, magnetic disks, etc.). Memory / storage 712 may include fixed media (e.g., RAM, ROM, fixed hard drives, etc.) and removable media (e.g., flash memory, removable hard drives, optical disks, etc.). Computer-readable media 706 may be configured in various other ways as further described below.

[0090] The input / output interface 708 represents functionality that allows a user to input commands and information to the computing device 702, and also allows information to be presented to the user and / or other components or devices using various input / output devices. Examples of input devices include a keyboard, a cursor control device (e.g., a mouse), a microphone, a scanner, touch functionality (e.g., a capacitive or other sensor configured to detect physical touch), a camera (e.g., employing visible or invisible wavelengths such as infrared frequencies to recognize motion as gestures that do not involve touch), and the like. Examples of output devices include a display device (e.g., a monitor or projector), a speaker, a printer, a network card, a tactile response device, and the like. Thus, the computing device 702 can be configured in various ways as further described below to support user interaction.

[0091] The various techniques herein can be described in the general context of software, hardware elements, or program modules. Typically, these modules include routines, programs, objects, elements, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The terms "module," "function," and "component" as used herein generally represent software, firmware, hardware, or a combination thereof. The features of the techniques described herein are platform-independent, which means that these techniques can be implemented on a variety of commercial computing platforms with a variety of processors.

[0092] An implementation of the described systems, modules, and techniques may be stored on or transmitted via some form of computer-readable media. Computer-readable media may include various media that may be accessed by computing device 702. By way of example and not limitation, computer-readable media may include “computer-readable storage media” and “computer-readable signal media.”

[0093] In contrast to a simple signal transmission, carrier wave, or signal itself, "computer-readable storage medium" may refer to a medium and / or device that enables persistent and / or non-temporary storage of information. Therefore, a computer-readable storage medium refers to a non-signal bearing medium. Computer-readable storage media include hardware such as volatile and non-volatile, removable and non-removable media and / or storage devices, which are implemented in a method or technology suitable for the storage of information such as computer-readable instructions, data structures, program modules, logic elements / circuits, or other data. Examples of computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical storage, hard disk, cassette, magnetic tape, disk storage or other magnetic storage device, or other storage device, tangible medium or article suitable for storing the desired information and accessible by a computer.

[0094] "Computer-readable signal media" may refer to signal-bearing media that is configured to send instructions to the hardware of the computing device 802, such as via a network. Signal media may typically embody computer-readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave, data signal, or other transport mechanism. Signal media also include any information delivery media. The term "modulated data signal" refers to a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media include: wired media, such as a wired network or direct-wired connection; and wireless media, such as acoustic, RF, infrared, and other wireless media.

[0095] As previously described, hardware elements 810 and computer-readable media 806 represent modules, programmable device logic, and / or fixed device logic implemented in hardware that may be employed in some embodiments to implement at least some aspects of the techniques described herein, such as for executing one or more instructions. Hardware may include components of an integrated circuit or system on a chip, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a complex programmable logic device (CPLD), and other implementations of silicon or other hardware. In this context, hardware may operate as: a processing device that executes instructions and / or program tasks defined by logic embodied in the hardware; and hardware for storing instructions for execution, such as the computer-readable storage medium described previously.

[0096] The aforementioned combination may also be used to implement the various techniques described herein. Therefore, software, hardware or executable modules may be implemented as one or more instructions and / or logic implemented on some form of computer-readable storage medium and / or by one or more hardware elements 710. Computing device 702 may be configured to implement specific instructions and / or functions corresponding to software and / or hardware modules. Therefore, the implementation of a module that may be executed as software by computing device 702 may be implemented in hardware at least in part, for example, by the use of computer-readable storage medium and / or hardware elements 710 of processing system 704. Instructions and / or functions may be performed / operated by one or more articles (e.g., one or more computing devices 702 and / or processing system 704) to implement the techniques, modules and examples described herein.

[0097] The techniques described herein can be supported by various configurations of computing device 702 and are not limited to the specific examples of the techniques described herein. The functionality can also be implemented in whole or in part through the use of a distributed system, such as on a "cloud" 714 via a platform 716, as described below.

[0098] Cloud 714 includes and / or represents a platform 716 of resources 718. Platform 716 abstracts the underlying functionality of the hardware (e.g., servers) and software resources of cloud 714. Resources 718 may include applications and / or data that may be utilized when a computer process is executed on a server remote from computing device 702. Resources 718 may also include services provided over the Internet and / or through a subscriber network such as a cellular or Wi-Fi network.

[0099] The platform 716 may abstract resources and functionality to connect the computing device 702 with other computing devices. The platform 716 may also be used to abstract scaling of resources to provide a level of scaling that corresponds to the demand encountered for resources 718 implemented via the platform 716. Thus, in an interconnected device embodiment, the implementation of the functionality described herein may be distributed throughout the system 700. For example, functionality may be implemented in part on the computing device 702 and via the platform 716 that abstracts the functionality of the cloud 714.

[0100] in conclusion

[0101] Although the invention has been described in language specific to structural features and / or methodological acts, it should be understood that the invention defined in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as example forms of implementing the claimed invention.

Claims

1. A method implemented by at least one computing device in a digital media environment for providing recommendations based on feature vector representations, the method comprising: receiving, by the at least one computing device, session data for a plurality of sessions, the session data for each session of the plurality of sessions comprising data describing at least one attribute for a user profile, at least one item, and at least one interaction between the user profile and the at least one item during the session; defining, by the at least one computing device, each of the plurality of conversations as a sentence comprising a plurality of terms, each term of the plurality of terms corresponding to at least one of: the at least one attribute of the user profile, the at least one item, or the at least one interaction between the user profile and the at least one item; determining, by the at least one computing device, feature vector representations for the at least one attribute represented as words in the plurality of sentences, the at least one item, and the at least one interaction, and generating a feature similarity matrix from the feature vector representations; receiving, by the at least one computing device, data describing a cold user profile requesting access to an item; generating, by the at least one computing device, item recommendations for the cold user profile using the data describing the cold user profile and the feature similarity matrix; as well as The item recommendation is outputted by the at least one computing device.

2. The method of claim 1 , wherein the feature vector representation is determined based on a word embedding model, the word embedding model comprising a global log-bilinear regression model configured to generate word embeddings using a word-to-word co-occurrence matrix and a local context window.

3. The method of claim 2, wherein the word embedding model includes an objective function that includes fields for data describing implicit interactions between items and user profiles.

4. The method of claim 1, wherein generating the feature similarity matrix comprises: The dot product of the feature vector representations is calculated, and the dot product can be used to determine the similarity between different features. The method of claim 4 , wherein the different features include features describing attributes of a user profile and features describing an item.

6. A method according to claim 1, wherein the at least one attribute of the user profile describes one or more of the following: an identifier for the user profile, a city associated with the user profile, a gender associated with the user profile, an occupation associated with the user profile, an industry associated with the user profile, a country associated with the user profile, or an education level associated with the user profile.

7. The method of claim 1 , wherein defining each of the plurality of conversations as a sentence comprises: Each of the at least one attribute for the user profile, the at least one item, and the interaction between the user profile and the at least one item is associated with each of the plurality of terms using natural language processing.

8. The method according to claim 1, further comprising: Data describing subsequent interactions between the cold user profile and items is received by the at least one computing device, and additional item recommendations are generated based on the feature similarity matrix and the data describing the subsequent interactions between the cold user profile and the items.

9. A method implemented by at least one computing device in a digital media environment for receiving recommendations based on feature vector representations, the method comprising: sending, by the at least one computing device, data describing at least one user profile attribute along with a request for at least one item; receiving, by the at least one computing device, the at least one item identified based on the at least one user profile attribute and a feature similarity matrix generated based on a plurality of feature vector representations, each of the plurality of feature vector representations generated based on a sentence of words, each of the words representing a user profile identifier, a user profile attribute, an item, and an interaction between the user profile and the item; as well as The at least one item is output by the at least one computing device.

10. The method of claim 9, wherein a first sentence in the sentence of words describes an explicit interaction between the user profile and the item, and a second sentence in the sentence of words describes an implicit interaction between the user profile and the item.

11. The method of claim 10, wherein the explicit interaction between the user profile and the item comprises one or more of: a purchase of the item, or a rating assigned to the item.

12. A method according to claim 10, wherein the implicit interaction between the user profile and the item includes one or more of the following items: the number of views of the item during a session, the amount of interaction time between the user profile and the item, or the percentage of visits to the item accessed by the user profile.

13. The method of claim 9, wherein each of the feature vector representations is determined by applying a word embedding model to a sentence of the corresponding word.

14. The method of claim 13, wherein the word embedding model includes an objective function including fields for data describing implicit interactions between items and user profiles.

15. A system in a digital media environment for facilitating recommendations based on feature vector representations, comprising: at least one processor; as well as At least one module implemented at least in part in hardware of a computing device and operable to provide a recommendation by performing operations comprising: receiving session data for a plurality of sessions, the session data for each session of the plurality of sessions comprising data describing at least one attribute of a user profile, at least one item, and at least one interaction between the user profile and the at least one item during the session; defining each of the plurality of conversations as a sentence including a plurality of words, each word of the plurality of words corresponding to at least one of: the at least one attribute of the user profile, the at least one item, or the at least one interaction between the user profile and the at least one item; determining a feature vector representation for the at least one attribute represented as words in the plurality of sentences, the at least one item, and the at least one interaction, and generating a feature similarity matrix from the feature vector representation; receiving data describing a cold user profile different from the plurality of user profiles; determining at least one item recommendation for the cold user profile based on the data describing the cold user profile and the feature similarity matrix; and The at least one item recommendation is output to the cold user profile.

16. The system of claim 15, wherein the data describing the cold user profile comprises: At least one of an identifier for the cold user profile or an attribute describing a characteristic of the cold user profile.

17. The system of claim 15, wherein the data describing the cold user profile does not include information describing interactions between the cold user profile and items.

18. The system of claim 15, wherein the feature vector representation is determined using a word embedding model, the word embedding model comprising a global log-bilinear regression model configured to generate word embeddings using a word-to-word co-occurrence matrix and a local context window.

19. The system of claim 15, wherein generating the feature similarity matrix comprises: The dot product of the feature vector representations is calculated, and the dot product can be used to determine the similarity between different features.

20. The system of claim 15, wherein defining each of the plurality of conversations as a sentence comprises: Each of the at least one attribute for the user profile, the at least one item, and the interaction between the user profile and the at least one item is associated with a term from the plurality of terms using natural language processing.