Class recommendation with implicit item feedback

Abstract

Technologies for providing category recommendations include a category recommendation system that provides recommended categories to a user based on implicit data (e.g., user-item interactions) and / or explicit data (e.g., queries, user information). The recommendations may be personalized or non-personalized (i.e., depending on whether user embedding is used), queried or non-queried (i.e., depending on whether query embedding is used), or personalized and queried (if both user embedding and query embedding are used). In either of these cases, there is an offline mode and a service mode. In the offline mode, category embedding is generated from an aggregation of item embedding associated with candidate categories. In the service mode, a candidate category is selected for display on the user device based on a similarity between the category embedding and either or both of the user embedding and the query embedding.

Description

Background Art

[0001] Some recommender systems provide recommendations to users based on input received from the user. Some inputs are explicit (e.g., ratings, selections of particular preferences), while other inputs are implicit (e.g., clicks, views, swipes, skips, etc.). For example, some video service providers recommend videos to users based on explicit feedback such as star ratings. Such video services may also provide recommendations based on implicit feedback such as viewing time. Summary of the invention

[0002] Implementations described herein involve providing recommended categories to users based on implicit data (e.g., user-item interactions) and / or explicit data (e.g., queries, user information). For example, a user whose search history includes recipes for burgers, chicken wings, and spareribs may receive recommendations for categories such as American recipes, pub food recipes, barbecue recipes, etc. based on previous user interaction data and / or queries that associate recipes for burgers, chicken wings, and spareribs with the terms "American", "pubfood", "BBQ", etc. The recommendations may be personalized or non-personalized (i.e., depending on whether user embeddings are used), queried or non-queried (i.e., depending on whether query embeddings are used), or personalized and queried (i.e., when both user embeddings and query embeddings are used). In any of these cases, there is an offline mode and a service mode. In the offline mode, a category embedding is generated from an aggregation of item embeddings associated with a candidate category. In the service mode, a candidate category is selected for display on a user device based on similarity between the category embedding and either or both of the user embedding and the query embedding.

[0003] In one general aspect, a computer-based method may include generating a plurality of category embeddings by: for each candidate category in a plurality of candidate categories, selecting a respective subset of item embeddings corresponding to the candidate category; and aggregating the respective subset of item embeddings to produce a category embedding for the candidate category. The method may also include selecting a set of categories based on similarity between category embeddings of categories in the set of categories and at least one of a user embedding representing a user or a query embedding representing a query. The method may also include providing the set of categories for display on a user device.

[0004] In another general aspect, a computer program product includes a non-transitory storage medium, the computer program product including code that, when executed by a processing circuit system of a computing device, causes the processing circuit system to perform a method. The method may include: generating a plurality of category embeddings by: for each candidate category in a plurality of candidate categories, selecting a respective subset of item embeddings corresponding to the candidate category; and aggregating the respective subset of item embeddings to produce a category embedding for the candidate category. The method may also include selecting a category set based on a similarity between the category embeddings of the categories in the category set and at least one of a user embedding representing a user or a query embedding representing a query. The method may also include providing the category set for display on a user device.

[0005] In another general aspect, an electronic device includes a memory and a control circuit system coupled to the memory. The control circuit system can be configured to generate a plurality of category embeddings by: for each candidate category in a plurality of candidate categories, by further being configured to select a corresponding subset of item embeddings corresponding to the candidate category; and aggregating the corresponding subset of item embeddings to produce a category embedding for the candidate category. The control circuit system can also be configured to select a category set based on a similarity between the category embeddings of the categories in the category set and at least one of a user embedding representing a user or a query embedding representing a query. The control circuit system can also be configured to provide the category set for display on a user device.

[0006] The details of one or more implementations are set forth in the accompanying drawings and the description below. Other features will be apparent from the description and drawings, and from the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1A is a diagram illustrating an example search engine in which the improved techniques described herein may be implemented.

[0008] Figure 1B is a diagram illustrating an example set of categories displayed to a user resulting from a search query.

[0009] Figure 2 is a diagram illustrating an example electronic environment in which the improved techniques described herein may be implemented.

[0010] Figure 3 is a flow chart illustrating an example offline mode of the category generation system.

[0011] Figure 4 is a flow chart illustrating an example service model of a category generation system.

[0012] Figure 5 is a flow chart of an example method of generating categories according to an example of the disclosed implementations.

[0013] Figure 6 is a diagram illustrating examples of computer devices and mobile computer devices that can be used to implement the described techniques.

[0014] Figure 7 is a diagram illustrating an example of a distributed computer device that can be used to implement the described techniques. DETAILED DESCRIPTION

[0015] Conventional recommendation systems provide recommended items based on user feedback. This can be achieved by mapping items and users into a latent space (embedding space) and then calculating a similarity measure between users and items in the latent space.

[0016] For example, when a user has a browsing history that includes clicking on a burger recipe and a ribs recipe, the search system may return other burger and ribs recipes, and depending on the browsing history may have some other closely related recipes, such as a beef taco recipe, a grilled sausage recipe, a sauce recipe, etc.

[0017] A technical problem with the conventional recommendation systems described above is that the recommended results are limited in scope. For example, even in the above example where the recommendation system provides recipes other than burger or ribs recipes, the options provided by the recommendation system only cover what the user might like given the item the user clicked. It is unlikely that the user will be prompted to explore other items that the user might like but cannot be determined from browsing history and / or other implicit and explicit feedback.

[0018] According to the implementation described herein, the technical solution to the above technical problem includes a category recommendation system that provides recommended categories to users based on implicit data (e.g., user-item interactions) and / or explicit data (e.g., queries, user information). For example, a user whose search history includes recipes for burgers and ribs may receive recommendations for categories such as American recipes, bar food recipes, barbecue recipes, etc. The recommendation may be personalized or non-personalized (i.e., depending on whether user embedding is used), queried or non-queried (i.e., depending on whether query embedding is used), or personalized and queried (i.e., when both user embedding and query embedding are used). In any of these cases, there is an offline mode and a service mode. In the offline mode, the category embedding is generated from the aggregation of the item embeddings associated with the candidate category. In the service mode, the candidate category is selected for display on the user device based on the similarity between the category embedding and either or both of the user embedding and the query embedding.

[0019] A technical advantage of the disclosed implementation is that, compared to conventional recommendation systems, the above-described category recommendation system expands recommendations to include entire families of items that are not provided using conventional recommendation systems. For example, providing entire families of items represented by categories of recipes such as American recipes, bar food recipes, and / or barbecue recipes broadens the scope of the recommendation system to feature new families of items that may be missed by conventional recommendation systems that only provide a few related, individual items.

[0020] Figure 1A An example environment 100 is depicted in which a user may interact with one or more computer-implemented search services. Example computer-implemented search services may include search services for email services, chat services, document sharing services, calendar sharing services, photo sharing services, video sharing services, shopping services, blogging services, microblogging services, social networking services, location (location-aware) services, check-in services, and / or rating and review services. Figure 1A In the example of , an Internet search service is depicted, which is discussed in further detail herein. However, it should be understood that implementations of the present disclosure may include one or more computer-implemented services, such as the examples discussed herein.

[0021] Continue to refer Figure 1A , search engine 120 provides search services. Example environment 100 includes a network 102, such as a local area network (LAN), a wide area network (WAN), the Internet, or a combination thereof, which connects websites 104, user devices 106, and search engines 120. In some examples, network 102 can be accessed through wired and / or wireless communication links. For example, a mobile computing device such as a smart phone can access the network using a cellular network. Environment 100 may include millions of websites 104 and user devices 106.

[0022] In some examples, website 104 is provided as website resources 105 associated with a domain name and hosted by one or more servers. An example website is a collection of web pages formatted in a suitable machine-readable language, such as Hypertext Markup Language (HTML), which may contain text, images, multimedia content, and programming elements such as scripts. Each website 104 is maintained by a publisher, such as an entity that manages and / or owns the website. Website resources 105 may be static or dynamic.

[0023] In some examples, website resources 105 are data provided via network 102 and associated with a resource address, such as a uniform resource locator (URL). In some examples, resources 105 that may be provided by website 104 include web pages, word processing documents, and portable document format (PDF) documents, images, videos, and feeds, and other appropriate digital content. Resources 105 may include content such as words, phrases, images, and sounds, and may include embedded information such as meta information and hyperlinks and / or embedded instructions such as scripts.

[0024] In some examples, user device 106 is an electronic device under the control of a user and capable of requesting and receiving resources 105 through network 102. Example user devices 106 include personal computers, mobile computing devices, such as smart phones, wearable devices, and / or tablet computing devices that can send and receive data through network 102. As used throughout this document, the term mobile computing device ("mobile device") refers to a user device configured to communicate through a mobile communication network. Smart phones (e.g., phones capable of communicating through the Internet) are examples of mobile devices, and wearable devices and other smart devices such as smart speakers are also examples of mobile devices. User device 106 typically includes a user application (e.g., a web browser) to facilitate sending and receiving data through network 102.

[0025] In some examples, to facilitate searching for resources 105, search engine 120 identifies resources 105 by crawling and indexing resources 105 provided on website 104. Data about resources 105 can be indexed based on the resources to which the data corresponds. Indexed and optionally cached copies of resources 105 are stored in search index 122, for example, as indexed resources 126.

[0026] The user device 106 submits a search query 109 to the search engine 120. In some examples, the user device 106 may include one or more input modalities. Example modalities may include a keyboard, a touch screen, and / or a microphone. For example, a user may use a keyboard and / or a touch screen to type a search query. As another example, a user may speak a search query, and the user's voice is captured by a microphone and processed by speech recognition to provide a search query.

[0027] In response to receiving a search query 109, a search engine 120 accesses a search index 122 to identify resources 105 related to the search query 109, such as resources 105 having at least a minimum specified relevance score for the search query 109. The search engine 120 identifies the resources 105, generates a search result display 111 including search results 112 identifying the resources 105, and returns the search result display 111 to the user device 106. In an example scenario, the search result display may include one or more web pages, such as one or more search result pages. In some examples, the web pages may be provided based on web documents that may be written in any suitable machine-readable language. However, it is contemplated that implementations of the present disclosure may include other suitable display types. For example, the search results may be provided in a display generated by an application executed on a computing device and / or in a display generated by an operating system, such as a mobile operating system. In some examples, the search results may be provided in any suitable form, such as JavaScript-html, plain text.

[0028] Search results 112 are data generated by search engine 120 that identify a website resource 105 that is responsive to a particular search query and include a link to the website resource 105. Example search results 112 may include a web page title, a snippet of text or a portion of an image extracted from a web page, and a URL of the web page. In some examples, the data provided in search results 112 may be retrieved from a resource data store. For example, search engine 120 may provide a search results display 111 that displays search results 112. In some examples, search results display 111 may be populated with information provided from the resource data store, such as a web page title, a snippet of text or a portion of an image extracted from a web page.

[0029] According to implementations of the present disclosure, the example environment 100 also includes a category recommendation system 130 that is communicatively coupled—e.g., directly coupled or coupled through a network such as network 102—to the search engine 120. In some examples, and in the case of multiple computer-implemented services, the category recommendation system 130 is communicatively coupled to a corresponding system that provides the functionality of the corresponding computer-implemented service. In some examples, as discussed in further detail herein, a system such as the search engine 120 can interact with the category recommendation system 130 to provide category results 113 displayed in a browser window running the search engine 120.

[0030] To enable the category recommendation system 130, the search engine 120 may access user embeddings 132, item embeddings 134, and query embeddings 136 stored on the corresponding storage. Each of the user embeddings 132 is a vector of values ​​representing different attributes and / or characteristics of the user, and is generated, for example, from corresponding user data (e.g., the user's location, preferences, etc. obtained with the user's consent). In some implementations, the user embeddings 132 are generated using a supervised learning engine such as a dual encoder in collaboration with the project. Each of the item embeddings 134 is a vector of values ​​representing various attributes and / or characteristics of the project, and is generated, for example, from item data / item interaction data (item descriptors and / or identifiers, interaction history of one or more users). In some implementations, the item embeddings 134 are generated using a supervised learning engine such as a dual encoder in collaboration with a user or query. Each of the query embeddings 136 is a vector representing the value of the query, and is generated from the corresponding query data (a representation of information such as text, images, audio, video, tactile, etc. provided as a query or in response to a query provided to the search engine). In some implementations, the query embedding 136 is generated using a supervised learning engine such as a dual encoder in conjunction with items.

[0031] The supervised machine learning algorithm maps user data, project data, and query data to their corresponding embeddings so that the corresponding embeddings typically have dimensions significantly smaller than the data. In some implementations, the mapping of user and project / project interaction data generates project embeddings 134 and user embeddings 132 into a common latent space. A latent space is an embedding space in which embedding vectors are arranged by a measure of similarity to each other. Mapping data vectors into their embedded latent spaces simplifies the selection of potentially compatible users and projects. In some implementations, the mapping of query and project / project interaction data generates project embeddings 134 and query embeddings 136 into a common latent space.

[0032] Figure 1B 1 is a diagram illustrating a user interface 150 in which an example category result set 113 is displayed on a user device 106. Figure 1B As shown, the user interface 150 is divided into two sections 160 and 170 .

[0033] Section 160 displays a search bar in which a user enters a query consisting of text, or in some implementations, images, audio, video, etc. Figure 1B As shown in Figure 1, the example query is the text “How to make a great burger”. A conventional search system may return results such as Figure 1BExample search results with website, star rating and short description are shown for titles "Perfect Burger Recipe" and "Steakhouse Burgers." Conventional search engines may also include links to websites that may be of interest to the user.

[0034] However, if Figure 1B As shown, section 170 includes a collection of categories titled "Related Searches". This is the output of search engine 120. Figure 1B As shown, the related searches are categories generated and selected by the category recommendation system 130. In this example, the displayed categories include "hamburger recipes", "ground beef recipes", "steak recipes", "barbeque recipes", "high protein recipes", "American food recipes" and "pub food recipes". Compared with using a conventional search engine, these categories provide users with a more diverse collection of items that may be of further interest. For example, when a user clicks on "ground beef recipes", the category may provide a new set of links including taco recipes, meat sauce sandwich recipes, cheese macaroni recipes, etc. When a user clicks on "High protein recipes", it may provide another new set of links including chicken recipes, tofu recipes and insect recipes. Therefore, the search engine 120 provides links to recipes that may have a certain similarity to the query (or user) but are unlikely to be presented by a conventional search engine.

[0035] Figure 2 is a diagram illustrating an example environment 200 in which the above-described technical solutions may be implemented. The example environment 200 includes a computer 220 configured to generate categories for display to a user device in response to a query from a user.

[0036] The computer 220 includes a network interface 222, a set of processing units 224, a memory 226, and a display interface 228. The network interface 222 includes, for example, an Ethernet adapter, a token ring adapter, etc., for converting electronic and / or optical signals received from the network into electronic form for use by the computer 220. The processing unit 224 set includes one or more processing chips and / or assemblies. The memory 226 includes both volatile memory (e.g., RAM) and non-volatile memory such as one or more ROMs, disk drives, solid-state drives, etc. The processing unit 224 set and the memory 226 together form a control circuit system that is configured and arranged to perform the various methods and functions described herein. The display interface 228 is configured to provide data to a display device for rendering and display to a user.

[0037] In some implementations, one or more of the components of computer 220 may be or may include a processor (e.g., processing unit 224) configured to process instructions stored in memory 226. Figure 2 Examples of such instructions depicted in FIG. 2 include a latent space manager 230, an offline mode manager 240, and a service mode manager 250. Figure 2 As shown, the memory 226 is configured to store various data, which are described with respect to the corresponding managers using such data.

[0038] The latent space manager 230 is configured to generate latent space data 232 from items, users, and queries. The latent space manager 230 is configured to form item embedding vectors 234, user embedding vectors 236, and query embedding vectors 238 based on the item, user, and query data available to the latent space manager 230. In some implementations, the latent space manager 230 receives the item, user, and query data through the network interface 222—i.e., over a network.

[0039] The item embedding vector 234 is a representation of an item that one or more users have interacted with. As previously described, in some implementations, the item embedding vector 234 is the result of a supervised machine learning engine that takes the item data as input. In some implementations, the item embedding vector 234 includes a vector of values ​​and has a dimension that is significantly smaller than the input item data. In such implementations, the supervised machine learning engine performs compression, for example, by calculating a similarity (distance) metric between pairs of item embedding vectors 234 so that similar items have similar embedding vectors. In some implementations, the item embedding vector 234 is an input item vector. In such implementations, the item embedding vector 234 has a binary value, each of which corresponds to, for example, an item descriptor (e.g., size, weight, purpose, etc.).

[0040] The user embedding vector 236 is a representation of the user. In some implementations, the user embedding vector 236 can be formed from user data including an explicit description of the user. Examples of explicit descriptions are those provided by the user, such as user settings in a browser. In some implementations, the user embedding vector 236 can be formed from user data including an implicit description of the user. An implicit description is a description that is not provided directly by the user but provided by computer instructions that perform observation / analysis on data accessed by the user operating the computer running these instructions. In other words, the implicit description of the user is a description derived from the user's interaction with the computing device or system. Examples of implicit descriptions include geographic location, interaction data including the number of clicks and mouse movement data (i.e., data generated by the user about the item by, for example, viewing, clicking, highlighting, etc.), the amount of time spent on the item in the active window, etc.

[0041] In some implementations, the user embedding vector 236 comprises a vector of values ​​and has a dimension significantly smaller than the input user data input into the supervised machine learning engine. In such implementations, the supervised machine learning engine performs compression, such as by calculating a similarity (distance) metric between pairs of user embedding vectors 236, so that similar users have similar embedding vectors. In some implementations, the user embedding vector 236 comprises an input user vector. In such implementations, the user embedding vector 238 has binary values, where each value corresponds to, for example, an explicit or implicit user descriptor.

[0042] Furthermore, in some implementations, the supervised machine learning engine maps the item embedding vector 234 and the user embedding vector 236 to the same latent space. Thus, when the item embedding vector 234 and the user embedding vector 236 are in the same latent space, there is a high likelihood that the user represented by the user embedding vector 236 produced the item represented by the item embedding vector 234. An advantage of mapping data across the same latent space is that it reduces the amount of processing that must be performed and simplifies the processing that must be performed.

[0043] The query embedding vector 238 is a representation of the query data input by the user. As previously described, in some implementations, the query embedding vector 238 is the result of a supervised machine learning engine with the query data as input. In some implementations, the query embedding vector 238 includes a vector of values ​​and has a dimension that is significantly smaller than the input query data. In such implementations, the supervised machine learning engine performs compression, for example, by calculating a similarity (distance) metric between pairs of query embedding vectors 238 so that similar queries have similar embedding vectors. In some implementations, the query embedding vector 238 is an input item vector. In such implementations, the query embedding vector 238 has binary values, each of which corresponds to, for example, an item descriptor (e.g., size, weight, purpose, etc.).

[0044] In addition, in some implementations, the supervised machine learning engine maps the item embedding vector 234 and the query embedding vector 238 to the same latent space. As previously discussed, a latent space is an embedding space in which embedding vectors are arranged according to a measure of similarity to each other. Therefore, when the item embedding vector 234 and the query embedding vector 238 are in the same latent space, there is a high probability that the query represented by the query embedding vector 238 produces the item represented by the item embedding vector 234.

[0045] The offline mode manager 240 is configured to generate offline model data 241, in particular category embedding data 244, from the candidate category data 242 and the item embedding subset data 243. The category embedding data 244 is then an input to the service mode manager 250, which provides category links displayed to the user in the browser window. Figure 3 Details of the offline mode manager 240 are illustrated.

[0046] Figure 3 3 is a flow chart 300 illustrating an example offline mode of a category generation system (eg, offline mode manager 240). Figure 3 As shown, the offline mode has a plurality of candidate categories 310 as input and a plurality of category embeddings 380 as output at 370 .

[0047] Multiple candidate categories 310 may include a list of possible categories (a set of possible categories, multiple possible categories) that can be presented to the user. The candidate categories may originate from a pool of categories that are defined and developed over time. In some implementations, each of the multiple candidate categories is represented by a text string. However, this is not restrictive, and in some implementations, each of the multiple candidate categories is represented by an image, an audio clip, a video clip, tactile feedback, etc. In some implementations, each of the multiple candidate categories 310 is defined by a manually provided category label. In some implementations, each of the multiple candidate categories 310 can be defined using a machine learning process.

[0048] At 320, the offline mode manager 240 is configured to provide each of the plurality of candidate categories 310 as a search string to a search engine (e.g., the search engine 120) and obtain a corresponding search result for the provided category. The search result corresponds to a candidate category in the plurality of candidate categories 310. That is, as a result of the search at 320, each of the candidate categories has a corresponding query-to-item mapping. For example, when the first candidate category "hamburger recipes" is entered into the search engine, the search engine returns a first list of items from the input "hamburger recipes". When the second candidate category "barbeque recipes" is entered into the search engine, the search engine returns a second list of items from the input "barbeque recipes". This process of entering a candidate category in the plurality of candidate categories 310 into the search engine and returning a corresponding list of search results is repeated for each of the plurality of candidate categories 310.

[0049] At 330, the offline mode manager 240 is configured to rank, for each of the candidate categories 310, a list of search results corresponding to the candidate category. In some implementations, the offline mode manager 240 performs the ranking according to an indicator of relevance to the candidate category (input query).

[0050] At 340, offline mode manager 240 is configured to select, for each of candidate categories 310, item embeddings of a subset of ranked items of the candidate category to form item embedding subset data 243. In some implementations, offline mode manager 240 generates item embedding vectors for the top N ranked items, where N is a specified number, such as 25, 50, 100, etc. In some implementations, item embedding vectors 234 form item embedding subset data 243. Figure 3 As shown, for each candidate category, the item embedding subset data 243 is in the form of item embeddings 350(1), 350(2), ..., 350(N).

[0051] At 360, the offline mode manager 240 is configured to aggregate the item embeddings 350(1), 350(2), ..., 350(N) for each of the candidate categories 310 to form an aggregated embedding. In some implementations, the aggregation includes summing. In some implementations, the aggregation includes forming an arithmetic mean of the item embedding vectors representing the item embeddings 350(1), 350(2), ..., 350(N). In some implementations, the aggregation includes summing or forming a median of the item embedding vectors representing the item embeddings 350(1), 350(2), ..., 350(N). In some implementations, the aggregation includes forming a normalized mean / sum / median of the item embedding vectors representing the item embeddings 350(1), 350(2), ..., 350(N), or in other words, normalizing the aggregated embedding. In some implementations, aggregation includes selecting the item embedding having the smallest distance to the query embedding vector based on the query that produces the list of items for each of the candidate categories 310, such as item embedding 350(J). For example, the distance can be based on a similarity metric such as an inner product or a cosine metric. In some implementations, aggregation includes forming a weighted average of the item embedding vectors representing item embeddings 350(1), 350(2), ..., 350(N).

[0052] At 370, the offline mode manager 240 is configured to form, for each of the candidate categories 310, a category embedding in a category embedding 380 for the candidate category from the aggregated embeddings. That is, in some implementations, the aggregated item embedding vectors of the candidate categories are represented in the category embedding 380, e.g., the corresponding category embeddings of the candidate categories.

[0053] like Figure 3 As shown, the output of the offline mode is the category embeddings in the category embeddings 380 corresponding to the corresponding candidate categories in the plurality of candidate categories 310 .

[0054] Back to Figure 2 , the service mode manager 250 is configured to generate the selected category data 259 by selecting a set of category embeddings 380 (i.e., category embedding data 244) so ​​as to display the category corresponding to the category embedding in the browser window displayed on the user device 106. How the category embedding set is selected is based on the service mode data 254. Figure 2 As shown, the service mode manager 250 includes an embedding manager 251 and a ranking manager 252 .

[0055] The service pattern data 254 includes an indication of whether the set of category embeddings is based on the user embedding vector 236 or the query embedding vector 238 or both. Figure 4The details of the service mode manager 250's selection of the category embedding set are described in detail in .

[0056] Figure 4 400 is a diagram illustrating an example service model of a category generation system. Figure 4 As shown, the embedding manager 251 takes one or both of user feature data 410 and query data 420 as input. The user feature data 410 represents the user who inputs the query represented by the query data 420. In some implementations, the user feature data 410 takes the form of a binary vector indicating the presence or absence of a binary user feature. In some implementations, the query data 420 takes the form of a text string. In some implementations, the query data 420 takes the form of an image, an audio clip, a video clip, or a tactile input.

[0057] like Figure 4 As shown, the embedding manager 251 is configured to input the user feature data 410 into the user embedding generator 430 to generate a latent user embedding vector representation 255 from the latent space where the user embedding vector 236 and the item embedding vector 234 are located. The latent space, also known as the latent feature space or the embedding space, is a representation of compressed data in a high-dimensional space (e.g., 127 or 256 dimensions) in which similar data points are closer in space. The latent user / item embedding can be generated using a user-item interaction and a dual encoder (also known as a dual tower model) architecture.

[0058] like Figure 4 As shown, the embedding manager 251 is configured to input the query data 420 into the query embedding generator 450 to generate a latent query embedding vector representation 256 from the latent space where the query embedding vector 238 and the item embedding vector 234 are located. The query embedding can be generated in different ways, including but not limited to: (1) aggregating the embeddings of items returned by the search engine for a particular query, (2) aggregating the embeddings of the words used in the query.

[0059] In some implementations where the service pattern data 254 indicates that the selected category is based only on the user feature data 410, the embedding manager 251 uses only the user feature data 410. Therefore, in such implementations, the latent query embedding vector representation 256 representation is not generated.

[0060] In an implementation where the service mode data 254 indicates that the selected category is based only on the user feature data 410, the ranking manager 252 inputs the potential user embedding vector representation 255 and each category embedding in the category embedding 380 into the similarity score engine 440 to generate a similarity score between the potential user embedding vector representation 255 and the category embedding in the category embedding 380 as similarity data 257. That is, if there are M category embeddings in the category embedding 380, then there are M similarity scores, each of which corresponds to a category embedding. The ranking manager 252 then selects a specified number K of the best / highest similarity scores at 480 and generates a ranking of the plurality of category embeddings by similarity score, i.e., the best similarity score set, as ranked category data 258. The selected category data 259 represents the categories of the top K similarity scores between the potential user embedding vector representation 255 and the category embedding 380.

[0061] In some implementations, the similarity score is based on a similarity measure. In some implementations, the similarity measure includes an inner product between the user embedding vector representation 255 and the category embedding. In this context, the inner product is the sum of the products of the corresponding vector elements. In some implementations, the similarity measure includes the cosine of the angle between the user embedding vector representation 255 and the category embedding. In some implementations, the top K similarity scores are selected based on a specified threshold, for example, similarity scores greater than a specified threshold. Other similar methods of determining similarity scores may be used.

[0062] In some implementations where the service mode data 254 indicates that the selected category is based only on the query data 420, the embedding manager 251 uses only the query data 420. Thus, in such implementations, the user embedding vector representation 255 is not generated.

[0063] In an implementation where the service mode data 254 indicates that the selected category is based only on the query data 420, the ranking manager 252 inputs the potential query embedding vector representation 256 and each category embedding in the category embedding 380 into the similarity score engine 440 to generate a similarity score between the potential query embedding vector representation 256 and the category embedding in the category embedding 380 as similarity data 257. That is, if there are M category embeddings in the category embedding 380, then there are M similarity scores, each of which corresponds to a category embedding. The ranking manager 252 then selects a specified number K of the highest similarity scores at 480 and generates a ranking of the plurality of category embeddings by similarity score, i.e., a set of the highest ranked similarity scores, as ranked category data 258. The selected category data 259 represents the categories of the top K similarity scores between the potential query embedding vector representation 256 and the category embedding 380.

[0064] In some implementations, the similarity score is based on a similarity measure. In some implementations, the similarity measure comprises an inner product between the potential query embedding vector representation 256 and the category embedding. In this context, the inner product is the sum of the products of the corresponding vector elements. In some implementations, the similarity measure comprises the cosine of the angle between the potential query embedding vector representation 256 and the category embedding.

[0065] In some implementations where the service pattern data 254 indicates that the selected category is based on both the user feature data 410 and the query data 420, the embedding manager 251 uses both the user feature data 410 and the query data 420. Thus, in such implementations, a user embedding vector representation 255 and a potential query embedding vector representation 256 are generated.

[0066] In implementations where the service pattern data 254 indicates that the selected category is based on both the user feature data 410 and the query data 420 , the ranking manager 252 may operate first on the latent query embedding vector representation 256 or first on the user embedding vector representation 255 .

[0067] In the former case, the ranking manager 252 inputs the potential query embedding vector representation 256 and each of the category embeddings in the plurality of category embeddings 380 into the similarity score engine 440 to generate a first similarity score between the potential query embedding vector representation 256 and the category embedding in the category embeddings 380 as similarity data 257. That is, if there are M category embeddings in the category embeddings 380, then there are M similarity scores, each of which corresponds to a category embedding. The ranking manager 252 then generates a temporary ranking of the plurality of category embeddings by the first similarity score at 460. The ranking manager 252 then selects the top K / highest ranked temporary category embeddings based on the similarity scores at 470. In addition, the ranking manager 252 re-ranks the top K category embeddings based on the second similarity scores between the potential user embedding vector representation 255 and the top K / highest ranked temporary category embeddings to generate ranked category data 258.

[0068] It is worth noting that the first similarity score and the second similarity score do not need to be calculated using the same similarity metric. For example, the first similarity score can be based on inner product, while the second similarity score is based on cosine, or vice versa.

[0069] In the latter case, the ranking manager 252 inputs the potential user embedding vector representation 255 and each of the category embeddings in the category embeddings 380 into the similarity score engine 440 to generate a first similarity score between the potential user embedding vector representation 255 and the category embedding in the category embeddings 380 as similarity data 257. That is, if there are M category embeddings in the category embeddings 380, then there are M similarity scores, each of which corresponds to a category embedding. The ranking manager 252 then generates a temporary ranking of the plurality of category embeddings by the first similarity score at 460. The ranking manager 252 then selects the top K / highest ranked temporary category embeddings based on the similarity scores at 470. In addition, the ranking manager 252 re-ranks the top K category embeddings based on a second similarity score between the potential query embedding vector representation 256 and the top K / highest ranked temporary category embeddings to generate ranked category data 258.

[0070] In all cases described above with respect to ranking manager 252 , the selected category data 259 is a list of categories to which the elements of ranked category data 258 correspond.

[0071] The components of the user device 106 (e.g., modules, processing unit 224) can be configured to operate based on one or more platforms (e.g., one or more similar or different platforms), which can include one or more types of hardware, software, firmware, operating systems, runtime libraries, etc. In some implementations, the components of the computer 220 can be configured to operate within a device cluster (e.g., a server farm). In such implementations, the functionality and processing of the components of the computer 220 can be distributed to several devices of the device cluster.

[0072] The components of computer 220 may be or include any type of hardware and / or software configured to process attributes. Figure 2 One or more portions of the components shown in the components of computer 220 in FIG. 2 may be or may include hardware-based modules (e.g., digital signal processors (DSPs), field programmable gate arrays (FPGAs), memory), firmware modules, and / or software-based modules (e.g., modules of computer code, sets of computer-readable instructions that can be executed at a computer). For example, in some implementations, one or more portions of the components of computer 220 may be or may include software modules configured to be executed by at least one processor (not shown). In some implementations, the functionality of the components may include being executed in conjunction with Figure 2 The modules and / or components shown in FIG. 1 are different modules and / or different components, including combining the functionality illustrated as two components into a single component.

[0073] Although not shown, in some implementations, the components (or portions thereof) of the computer 220 may be configured to operate within, for example, a data center (e.g., a cloud computing environment), a computer system, one or more server / host devices, etc. In some implementations, the components (or portions thereof) of the computer 220 may be configured to operate within a network. Therefore, the components (or portions thereof) of the computer 220 may be configured to function within various types of network environments that may include one or more devices and / or one or more server devices. For example, the network may be or may include a local area network (LAN), a wide area network (WAN), etc. The network may be or may include a wireless network and / or a wireless network implemented using, for example, a gateway device, a bridge, a switch, etc. The network may include one or more segments and / or may have portions based on various protocols such as the Internet Protocol (IP) and / or a proprietary protocol. The network may include at least a portion of the Internet.

[0074] In some implementations, one or more of the components of the search engine 120 may be or include a processor configured to process instructions stored in a memory, such as the latent space manager 230 (and / or a portion thereof), the offline mode manager 240 (and / or a portion thereof), and the service mode manager 250 (and / or a portion thereof).

[0075] In some implementations, memory 226 can be any type of memory, such as random access memory, disk drive memory, flash memory, etc. In some implementations, memory 226 can be implemented as more than one memory component associated with a component of computer 220 (e.g., more than one RAM component or disk drive memory). In some implementations, memory 226 can be a database memory. In some implementations, memory 226 can be or include non-local memory. For example, memory 226 can be or include memory shared by multiple devices (not shown). In some implementations, memory 226 can be associated with a server device (not shown) within a network and configured to serve components of computer 220. Memory 226 can be a non-transitory storage device or a temporary storage device that enables processing circuitry (e.g., processing unit 224) to perform operations such as those described above. Figure 5 The method 500 is described. Figure 2 As shown, the memory 226 is configured to store various data, including latent space data 232 , offline mode data 241 , and service mode data 254 .

[0076] Figure 5 is a flowchart depicting an example method 500 according to the improved technique described above. The method 500 may be performed by combining Figure 2The described software structure is executed by the computer 220, which resides in the memory 226 of the computer 220 and is collectively executed by the processing unit 224.

[0077] At 502, the offline mode manager 240 generates multiple category embeddings (i.e., category embeddings 380) by: for each candidate category among multiple candidate categories, selecting a corresponding subset of item embeddings corresponding to the candidate category (i.e., item embedding subset data 243) at 504; and aggregating the corresponding subsets of item embeddings to produce a category embedding for the candidate category at 506.

[0078] At 508, ranking manager 252 selects a set of categories (e.g., selected category data 259) based on similarities (e.g., similarity data 257) between category embeddings of categories of the set of categories and at least one of user embeddings representing users (e.g., latent user embedding vector representation 255) or query embeddings representing queries (e.g., latent query embedding vector representation 256).

[0079] At 510 , computer 220 provides the set of categories for display on a user device (eg, user device 106 ).

[0080] Figure 6 An example of a general purpose computer device 600 and a general purpose mobile computer device 650 that can be used with the techniques described herein is illustrated. The computer device 600 is Figure 1A The search engine 120 and / or category recommendation system 130 and Figure 2 An example configuration of computer 220 is shown.

[0081] like Figure 6 As shown, computing device 600 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. Computing device 650 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, and other similar computing devices. The components shown here, their connections and relationships, and their functions are intended to be examples only and are not intended to limit implementations of the inventions described and / or claimed in this document.

[0082] The computing device 600 includes a processor 602, a memory 604, a storage device 606, a high-speed interface 608 connected to the memory 604 and a high-speed expansion port 610, and a low-speed interface 612 connected to a low-speed bus 614 and the storage device 606. Each of the components 602, 604, 606, 608, 610 and 612 is interconnected using various buses and can be installed on a common motherboard or installed in other ways as appropriate. The processor 602 can process instructions for execution within the computing device 600 (including instructions stored in the memory 604 or on the storage device 606) to display graphical information of a GUI on an external input / output device such as a display 616 coupled to the high-speed interface 608. In other implementations, multiple processors and / or multiple buses and multiple memories and multiple types of memories can be used as appropriate. In addition, multiple computing devices 600 can be connected, each of which provides a portion of the necessary operations (for example, as a server group, a blade server group or a multi-processor system).

[0083] The memory 604 stores information within the computing device 600. In one implementation, the memory 604 is one or more volatile memory units. In another implementation, the memory 604 is one or more non-volatile memory units. The memory 604 may also be another form of computer-readable medium, such as a magnetic disk or optical disk.

[0084] The storage device 606 can provide mass storage for the computing device 600. In one implementation, the storage device 606 can be or include a computer-readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a magnetic tape device, a flash memory or another similar solid-state memory device, or an array of devices, including devices in a storage area network or other configuration. A computer program product can be tangibly embodied in an information carrier. The computer program product can also include instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer-readable medium or a machine-readable medium, such as the memory 604, the storage device 606, or a memory on the processor 602.

[0085] The high-speed controller 608 manages bandwidth-intensive operations of the computing device 600, while the low-speed controller 612 manages less bandwidth-intensive operations. Such allocation of functions is merely an example. In one implementation, the high-speed controller 608 is coupled to the memory 604, the display 616 (e.g., through a graphics processor or accelerator), and a high-speed expansion port 610 that can accept various expansion cards (not shown). In this implementation, the low-speed controller 612 is coupled to the storage device 606 and the low-speed expansion port 614. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet), may be coupled to one or more input / output devices, such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, for example, through a network adapter.

[0086] As shown in the figure, computing device 600 can be implemented in a variety of different forms. For example, it can be implemented as a standard server 620 or multiple times in a group of such servers. It can also be implemented as part of a rack server system 624. In addition, the computing device can be implemented in a personal computer such as a laptop computer 622. Alternatively, components from computing device 600 can be combined with other components in a mobile device (not shown) such as device 650. Each of such devices can contain one or more of computing devices 600, 650, and the entire system can be composed of multiple computing devices 600, 650 that communicate with each other.

[0087] The computing device 650 includes a processor 652, a memory 664, an input / output device such as a display 654, a communication interface 666, and a transceiver 668, among other components. The device 650 may also be provided with a storage device, such as a micro drive or other device, for providing additional storage. Each of the components 650, 652, 664, 654, 666, and 668 are interconnected using various buses, and several of these components may be mounted on a common motherboard or otherwise as appropriate.

[0088] Processor 652 can execute instructions within computing device 650, including instructions stored in memory 664. The processor can be implemented as a chipset including separate multiple analog and digital processor chips. For example, the processor can provide coordination of other components of device 650, such as control of a user interface, applications executed by device 650, and wireless communications performed by device 650.

[0089] The processor 652 can communicate with the user through a control interface 658 and a display interface 656 coupled to the display 654. The display 654 can be, for example, a TFT LCD (thin film transistor liquid crystal display) or an OLED (organic light emitting diode) display, or other suitable display technology. The display interface 656 may include suitable circuitry for driving the display 654 to present graphics and other information to the user. The control interface 658 can receive commands from the user and convert the commands for submission to the processor 652. In addition, an external interface 660 in communication with the processor 652 can be provided to enable the device 650 to communicate with adjacent areas of other devices. The external interface 660 can, for example, provide wired communication in some implementations, or provide wireless communication in other implementations, and multiple interfaces can also be used.

[0090] The memory 664 stores information within the computing device 650. The memory 664 may be implemented as one or more of one or more computer-readable media, one or more volatile memory units, or one or more non-volatile memory units. An expansion memory 674 may also be provided, and the expansion memory is connected to the device 650 via an expansion interface 672, which may include, for example, a SIMM (single in-line memory module) card interface. Such an expansion memory 674 may provide additional storage space for the device 650, or may also store applications or other information of the device 650. Specifically, the expansion memory 674 may include instructions for executing or supplementing the above-mentioned processes, and may also include security information. Therefore, for example, the expansion memory 674 may be provided as a security module for the device 650, and may be programmed with instructions that allow the secure use of the device 650. In addition, security applications and additional information may also be provided via a SIMM card, such as placing identification information on a SIMM card in an unbreakable manner.

[0091] The memory may include, for example, flash memory and / or NVRAM memory, as discussed below. In one implementation, a computer program product is tangibly embodied in an information carrier. The computer program product includes instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer-readable medium or machine-readable medium, such as memory 664, expansion memory 674, or memory on processor 652, which may be received, for example, via transceiver 668 or external interface 660.

[0092] The device 650 can communicate wirelessly via a communication interface 666, which may include digital signal processing circuitry, if necessary. The communication interface 666 may provide for communication in various modes or protocols, such as GSM voice calls, SMS, EMS or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, etc. Such communications may occur, for example, via a transceiver 668. In addition, short-range communications may occur, such as using Bluetooth, WiFi, or other such transceivers (not shown). In addition, a GPS (global positioning system) receiver module 670 may provide additional navigation and location-related wireless data to the device 650, which may be used by applications running on the device 650 as appropriate.

[0093] Device 650 may also communicate audibly using audio codec 660, which may receive verbal information from a user and convert it into usable digital information. Audio codec 660 may also generate audible sounds for the user, such as through a speaker, for example, in a handset of device 650. Such sounds may include sounds from voice telephone calls, may include recorded sounds (e.g., voice messages, music files, etc.), and may also include sounds generated by applications operating on device 650.

[0094] As shown in the figure, computing device 650 can be implemented in many different forms. For example, it can be implemented as a cellular phone 680. It can also be implemented as part of a smart phone 682, a personal digital assistant or other similar mobile device.

[0095] Figure 7 Examples of devices that can be used with the techniques described herein include Figure 2 220. Computing device 700 is intended to represent various example forms of large-scale data processing devices, such as servers, blade servers, data centers, mainframes, and other large-scale computing devices. Computing device 700 may be a distributed system with multiple processors, and may include network-attached storage nodes interconnected by one or more communication networks. The components shown here, their connections and relationships, and their functions are intended to be examples only and are not intended to limit implementations of the inventions described and / or claimed in this document.

[0096] The distributed computing system 70 may include any number of computing devices 780. The computing devices 780 may include servers or rack servers, mainframes, etc. communicating via a local or wide area network, dedicated optical links, modems, bridges, routers, switches, wired or wireless networks, etc.

[0097] In some implementations, each computing device may include multiple racks. For example, computing device 780a includes multiple racks 758a, 758b, ..., 758n. Each rack may include one or more processors, such as processors 752a, 752B, ..., 752n and 762a, 762b, ..., 762n. The processor may include a data processor, a network attached storage device, and other computer-controlled devices. In some implementations, one processor may operate as a master processor and control scheduling and data distribution tasks. The processors may be interconnected through one or more rack switches 758, and one or more racks may be connected through a switch 778. Switch 778 may handle communications between multiple connected computing devices 700.

[0098] Each rack may include memory such as memory 754 and memory 764 and storage such as 756 and 766. Storage 756 and 766 may provide mass storage and may include volatile or non-volatile storage such as a network attached disk, a floppy disk, a hard disk, an optical disk, a tape, a flash memory or other similar solid-state memory device, or an array of devices, including devices in a storage area network or other configuration. Storage 756 or 766 may be shared between multiple processors, multiple racks, or multiple computing devices, and may include a computer-readable medium storing instructions executable by one or more processors. Memory 754 and 764 may include, for example, one or more volatile memory units, one or more non-volatile memory units, and / or other forms of computer-readable media such as magnetic or optical disks, flash memory, cache, random access memory (RAM), read-only memory (ROM), and combinations thereof. Memory such as memory 754 may also be shared between processors 752a, 752b, ..., 752n. Data structures such as indexes may be stored, for example, across storage 756 and memory 754. The computing device 700 may include other components not shown, such as a controller, a bus, an input / output device, a communication module, etc.

[0099] An entire system such as environment 100 may be comprised of multiple computing devices 700 that communicate with each other. For example, device 780a may communicate with devices 780b, 780c, and 780d, and these may be collectively referred to as environment 100. As another example, Figure 2 The environment 200 may include a computing device 700 or multiple computing devices 700. Some of the computing devices may be located geographically close to each other, and other computing devices may be located geographically far away. The layout of the computing device 700 is only an example, and the system may adopt other layouts or configurations.

[0100] In some implementations, providing a candidate paragraph for display based on the accuracy score includes: comparing the accuracy score to an accuracy score threshold; in response to the accuracy score being greater than the accuracy score threshold, displaying the candidate paragraph on the display; and in response to the accuracy score being less than the accuracy score threshold, not displaying the candidate paragraph on the display.

[0101] In some implementations, selecting a set of categories includes: obtaining a query embedding generated from a query; generating a plurality of similarity scores by generating a corresponding similarity score based on a similarity measure between the query embedding and each of the plurality of category embeddings; and generating the set of categories by ranking the plurality of similarity scores. In some implementations, generating the corresponding similarity score includes generating an inner product between the query embedding and the category embedding. In some implementations, the query embedding is obtained from a latent space, the latent space also including item embeddings.

[0102] In some implementations, selecting a set of categories includes: obtaining a user embedding generated from user feature data and user-item interaction data; generating a plurality of similarity scores by generating a corresponding similarity score based on a similarity measure between the user embedding and each of a plurality of category embeddings; and generating the set of categories by ranking the plurality of similarity scores. In some implementations, generating the corresponding score includes generating an inner product between the user embedding and the category embedding. In some implementations, the user embedding is obtained from a latent space that also includes item embeddings.

[0103] In some implementations, selecting a category set includes: obtaining a query embedding and a user embedding; generating a plurality of first similarity scores based on a first similarity measure between the query embedding and each category embedding in a plurality of category embeddings; generating a temporary category embedding set by ranking a specified number of first similarity scores in the plurality of first similarity scores; generating a plurality of second similarity scores based on a second similarity measure between the query embedding and each temporary category embedding in the temporary category embedding set; and generating the category set by ranking the plurality of second similarity scores. In some implementations, the user embedding is generated from user feature data or user interaction data. In some implementations, generating a corresponding first similarity measure between the query embedding and a category embedding in the plurality of category embeddings includes generating an inner product between the query embedding and the category embedding. In some implementations, generating a corresponding second similarity measure between the user embedding and the category embedding in the temporary category embedding set includes generating an inner product between the user embedding and the category embedding.

[0104] In some implementations, selecting a category set includes: obtaining a query embedding and a user embedding; generating multiple first similarity scores based on a first similarity measure between the user embedding and each category embedding in a plurality of category embeddings; generating a temporary category embedding set by ranking the first similarity scores; generating multiple second similarity scores based on a second similarity measure between the query embedding and each temporary category embedding in the temporary category embedding set; and generating the category set by ranking the multiple second similarity scores. In some implementations, the user embedding is generated from user feature data or user interaction data. In some implementations, generating a corresponding first similarity measure between the user embedding and a category embedding in the plurality of category embeddings includes generating an inner product between the query embedding and the category embedding. In some implementations, generating a corresponding second similarity measure between the user embedding and a category embedding in the temporary category embedding set includes generating an inner product between the query embedding and the category embedding.

[0105] In some implementations, aggregating a respective subset of the item embeddings for each of the plurality of candidate categories includes generating a median of the item embeddings corresponding to the candidate category as the category embedding for the candidate category.

[0106] In some implementations, method 500 further includes: obtaining a plurality of items corresponding to a candidate category in a plurality of candidate categories; generating a list of items by selecting a set of the plurality of items corresponding to the candidate category; and performing a mapping from the list of items to item embeddings of the candidate category (i.e., via a supervised machine learning function generated, for example, via double encoding). In some implementations, obtaining a plurality of items corresponding to each of the plurality of candidate categories includes generating a plurality of items corresponding to the candidate category as an output of the search engine by providing a search string to the search engine, the search string being based on an identifier (e.g., alphanumeric) of the candidate category.

[0107] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system that includes at least one programmable processor that can be either special purpose or general purpose and can be coupled to receive data and instructions from and transmit data and instructions to a storage system, at least one input device, and at least one output device.

[0108] These computer programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and may be implemented in high-level procedural and / or object-oriented programming languages ​​and / or assembly / machine languages. As used herein, the terms "machine-readable medium," "computer-readable medium," refer to any computer program product, apparatus, and / or device (e.g., a disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0109] To provide interaction with a user, the systems and techniques described herein may be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, verbal, or tactile input.

[0110] The systems and techniques described herein may be implemented in a computing system that includes a back-end component (e.g., as a data server), or includes a middleware component (e.g., an application server), or includes a front-end component (e.g., a client computer with a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described herein), or any combination of such back-end, middleware, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), and the Internet.

[0111] A computing system may include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises from computer programs running on the respective computers and having a client-server relationship to each other.

[0112] A number of implementations have been described, however, it will be appreciated that various modifications may be made without departing from the spirit and scope of the specification.

[0113] It should also be understood that when an element is referred to as being on another element, connected to another element, electrically connected to another element, coupled to another element, or electrically coupled to another element, the element may be directly on another element, connected or coupled to another element, or one or more intermediate elements may be present. On the contrary, when an element is referred to as being directly on another element, directly connected to another element, or directly coupled to another element, there is no intermediate element. Although the terms "directly on...", "directly connected to..." or "directly coupled to..." may not be used throughout the detailed description, the element shown as being directly on...", directly connected or directly coupled may be referred to as such. The claims of the present application may be modified to narrate the example relationships described in the specification or shown in the accompanying drawings.

[0114] Although certain features of the described implementations have been illustrated as described herein, many modifications, substitutions, changes and equivalents will now occur to those skilled in the art. Therefore, it should be understood that the appended claims are intended to cover all such modifications and changes that fall within the scope of the implementations. It should be understood that they are presented only by way of example and not limitation, and various changes in form and detail may be made. Except for mutually exclusive combinations, any part of the apparatus and / or method described herein may be combined in any combination. The implementations described herein may include various combinations and / or sub-combinations of the functions, components and / or features of the different implementations described.

[0115] Furthermore, the logic flows depicted in the figures do not require the particular order shown or sequential order to achieve the desired results. In addition, other steps may be provided, or steps may be excluded from the described flows, and other components may be added to or removed from the described systems. Therefore, other implementations are within the scope of the following claims.

Claims

1. A method comprising: Generate multiple category embeddings by: for each candidate category in multiple candidate categories, select the corresponding subset of item embeddings corresponding to that candidate class, and aggregating the corresponding subsets of item embeddings to produce a category embedding for the candidate category; selecting the set of categories from the plurality of category embeddings based on similarity between category embeddings of categories in the set of categories and at least one of a user embedding representing a user or a query embedding representing a query; and The set of categories is provided for display on a user device.

2. The method of claim 1, wherein: The category set selected includes: obtaining the query embedding generated from the query; generating a plurality of similarity scores by generating corresponding similarity scores based on a similarity measure between the query embedding and each of the plurality of category embeddings; and The set of categories is generated by ranking the plurality of similarity scores and selecting at least some categories having the highest ranked similarity scores.

3. The method of claim 2, wherein: Generating the corresponding similarity scores includes: Generate an inner product between the query embedding and the category embedding.

4. The method of claim 2, wherein: The query embedding is obtained from a latent space which also includes the item embeddings.

5. The method of claim 1, wherein: The category set selected includes: Obtaining the user embedding generated from user feature data or user-item interaction data; generating a plurality of similarity scores by generating a corresponding similarity score based on a similarity measure between the user embedding and each of the plurality of category embeddings; and The set of categories is generated by ranking the plurality of similarity scores and selecting at least some categories having the highest ranked similarity scores.

6. The method of claim 5, wherein: Generating the corresponding similarity scores includes: Generate an inner product between the user embedding and the category embedding.

7. The method of claim 5, wherein: The user embeddings are obtained from a latent space which also includes the item embeddings.

8. The method of claim 1, wherein: The category set selected includes: Obtaining the query embedding and the user embedding; generating a plurality of first similarity scores based on a first similarity measure between the query embedding and each of the plurality of category embeddings; generating a temporary set of category embeddings by selecting a specified number of categories with the highest ranked first similarity scores; generating a plurality of second similarity scores based on a second similarity measure between the query embedding and each provisional category embedding in the set of provisional category embeddings; and The set of categories is generated by selecting the category with the highest ranked temporary category embedding.

9. The method of claim 8, wherein: The user embedding is generated from user feature data or user interaction data.

10. The method of claim 8, wherein: Generating a respective first similarity score between the query embedding and a category embedding of the plurality of category embeddings comprises: Generate an inner product between the query embedding and the category embedding.

11. The method of claim 8, wherein: Generating respective second similarity scores between the user embedding and the category embeddings in the set of temporary category embeddings comprises: Generate an inner product between the user embedding and the category embedding.

12. The method of claim 1, wherein: The category set selected includes: Obtaining the query embedding and the user embedding; generating a plurality of first similarity scores based on a first similarity measure between the user embedding and each of the plurality of category embeddings; generating a temporary set of category embeddings by selecting a specified number of categories with the highest ranked first similarity scores; generating a plurality of second similarity scores based on a second similarity measure between the query embedding and each provisional category embedding in the set of provisional category embeddings; and The set of categories is generated by selecting the category with the highest ranked category embedding.

13. The method of claim 12, wherein: The user embedding is generated from user feature data or user interaction data.

14. The method of claim 1, wherein: Aggregating the respective subsets of item embeddings of candidate categories among the plurality of candidate categories comprises: A median of the item embeddings corresponding to the candidate categories is generated as the category embedding of the candidate category.

15. The method of claim 1, further comprising: obtaining a plurality of items corresponding to candidate categories among the plurality of candidate categories; generating a list of items by selecting a set of the plurality of items corresponding to the candidate categories; as well as A mapping is performed from the list of items to the item embeddings of the candidate categories.

16. The method of claim 15, wherein: Obtaining the plurality of items corresponding to each of the plurality of candidate categories comprises: The plurality of items corresponding to the candidate category are produced as an output of the search engine by providing a search string to the search engine, the search string being based on an identifier of the candidate category.

17. A computer program product comprising a non-transitory storage medium, the computer program product comprising code which, when executed by a processing circuit system on which a search engine is configured to execute, causes the processing circuit system to perform a method comprising: Generate multiple category embeddings by: for each candidate category in multiple candidate categories, select the corresponding subset of item embeddings corresponding to that candidate class, and aggregating the corresponding subsets of item embeddings to produce a category embedding for the candidate category; selecting the set of categories based on similarity between category embeddings of categories in the set of categories and at least one of a user embedding representing a user or a query embedding representing a query; as well as The set of categories is provided for display on a user device.

18. A device, comprising: Memory; as well as and control circuitry coupled to the memory, the control circuitry being configured to: Generate multiple category embeddings by: for each candidate category in multiple candidate categories, select the corresponding subset of item embeddings corresponding to that candidate class, and aggregating the corresponding subsets of item embeddings to produce a category embedding for the candidate category; selecting the set of categories based on similarity between category embeddings of categories in the set of categories and at least one of a user embedding representing a user or a query embedding representing a query; as well as The set of categories is provided for display on a user device.

19. The apparatus of claim 18, wherein: The control circuitry configured to select the set of categories is further configured to: obtaining the query embedding generated from the query; generating a plurality of similarity scores by generating corresponding similarity scores based on a similarity measure between the query embedding and each of the plurality of category embeddings; as well as The set of categories is generated by ranking the plurality of similarity scores.

20. The apparatus of claim 18, wherein: The control circuitry configured to select the set of categories is further configured to: Obtaining the user embedding generated from user feature data or user-item interaction data; generating a plurality of similarity scores by generating corresponding similarity scores based on a similarity measure between the user embedding and each of the plurality of category embeddings; as well as The set of categories is generated by ranking the plurality of similarity scores.

Images