Method, device, intelligent device and storage medium for processing user recommendations

Abstract

An embodiment of the present application provides a method, apparatus, smart device and storage medium for processing recommendations for users, wherein the method includes: obtaining an object-associated feature set of a target user identifier; using an interest category feature set to perform interest prediction processing on a first feature set determined based on the object-associated feature set to obtain a reference interest category of the target user identifier; using a second feature set determined based on the object-associated feature set to obtain a candidate recommendation set corresponding to the target user identifier, wherein the second feature set is obtained by filtering the object-associated feature set based on the determined interest category features of the target user identifier; determining a recommendation list for the target user identifier based on the reference interest category and the candidate recommendation set, thereby better determining recommended objects for the user.

Description

Technical Field

[0001] The present application relates to the field of computer application technology, and in particular to a method, apparatus, intelligent device, and storage medium for processing user recommendations. Background Art

[0002] In recent years, with the development of internet technology, intelligent recommendation technology has been widely used in an increasing number of internet products. Intelligent recommendation can help users find interesting information and videos more quickly, without having to spend too much time searching for them.

[0003] Existing recommendations typically determine recommendations based on user actions. For example, based on a video a user has watched, a recommendation might be made to that video, or one or more videos within that video's category. Similarly, based on a product a user has purchased, a recommendation might be made to that product, or one or more products within that product's category. However, how to more intelligently and comprehensively screen and confirm recommendations has become a hot research topic. Summary of the Invention

[0004] The embodiments of the present application provide a method, apparatus, smart device, and storage medium for processing user recommendations, which can better determine recommendation objects for users.

[0005] In one aspect, an embodiment of the present application provides a method for processing recommendations for a user, the method comprising:

[0006] Obtaining a set of object-related features identified by a target user;

[0007] Using the interest category feature set, performing interest prediction processing on the first feature set determined according to the object association feature set to obtain a reference interest category identified by the target user;

[0008] Obtaining a candidate recommendation set corresponding to the target user identifier by using a second feature set determined according to the object-associated feature set, wherein the second feature set is obtained by filtering the object-associated feature set according to the determined interest category features of the target user identifier;

[0009] Based on the reference interest category and the candidate recommendation set, a recommendation list for the target user identification is determined.

[0010] On the other hand, an embodiment of the present application provides a device for processing recommendations for a user, the device comprising:

[0011] An acquisition unit, configured to acquire an object-related feature set identified by a target user;

[0012] a processing unit configured to perform interest prediction processing on a first feature set determined according to the object-associated feature set using the interest category feature set to obtain a reference interest category identified by the target user;

[0013] The acquiring unit is further configured to acquire a candidate recommendation set corresponding to the target user identifier by using a second feature set determined based on the object-associated feature set, wherein the second feature set is obtained by filtering the object-associated feature set based on the determined interest category features of the target user identifier;

[0014] The processing unit is further configured to determine a recommendation list for the target user identifier based on the reference interest category and the candidate recommendation set.

[0015] Correspondingly, an embodiment of the present application further provides an intelligent device, comprising: a storage device and a processor; the storage device is used to store a computer program; the processor is used to execute the computer program to implement the above method.

[0016] Correspondingly, an embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed, the above method will be implemented.

[0017] The present application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a smart device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the smart device to perform the above-described method.

[0018] In an embodiment of the present application, after obtaining the object-associated feature set of the target user identifier, the computer is set to, on the one hand, perform interest prediction processing on the first feature set determined according to the object-associated feature set using the interest category feature set to obtain the reference interest category of the target user identifier; on the other hand, obtain the candidate recommendation set corresponding to the target user identifier using the second feature set determined according to the object-associated feature set; and then determine the recommendation list of the target user identifier based on the reference interest category and the candidate recommendation set. By determining the recommendation list of the target user identifier through the above two aspects, it can be ensured that the recommendation list generated when the interest exploration is completed includes the user's potential points of interest, rather than just points of interest that have been explored multiple times or in the user profile, thereby better determining the recommended objects for the user and improving the effectiveness and accuracy of interest recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1a This is a flowchart of a recommended processing solution for users provided in an embodiment of the present application;

[0021] Figure 1b This is a schematic diagram of the architecture of an interest recommendation system provided by an embodiment of the present application;

[0022] Figure 2 This is a flowchart of a method for processing user recommendations provided by an embodiment of the present application;

[0023] Figure 3a This is a schematic diagram of a positive user behavior sequence provided by an embodiment of the present application;

[0024] Figure 3b This is a schematic diagram of a reverse user behavior sequence provided by an embodiment of the present application;

[0025] Figure 4 This is a flowchart of a method for processing user recommendations provided by an embodiment of the present application;

[0026] Figure 5a This is a structural diagram of an interest exploration model provided by an embodiment of the present application;

[0027] Figure 5b This is a schematic diagram of the structure of another interest exploration model provided in an embodiment of the present application;

[0028] Figure 6 This is a flow chart of a model training method provided in an embodiment of the present application;

[0029] Figure 7a is a schematic diagram of an activation function provided in an embodiment of the present application;

[0030] Figure 7b This is a schematic diagram of analysis of results after using a method for recommending processing to a user provided in an embodiment of the present application;

[0031] Figure 8 This is a structural diagram of a device for processing user recommendations provided by an embodiment of the present application;

[0032] Figure 9 This is a schematic diagram of the structure of a smart device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0034] The embodiment of the present application provides a user recommendation processing solution, which can be applied to various recommendation applications or recommendation systems. It can improve the efficiency of interest exploration through interest prediction and interest candidate set recall, thereby better determining the recommended objects for users and improving the effectiveness and accuracy of interest recommendations. Figure 1a As shown, interest prediction and interest candidate set recall are started in parallel in the recommendation processing scheme for users; on the one hand, during interest prediction, object-related features such as user-side features of the target user, item-side features associated with the target user, and contextual features of the item features can be pulled in real time, and then these object-related features are processed for interest prediction using the interest category feature set to be predicted (hereinafter referred to as the interest category feature set) to obtain the reference interest category of the target user; on the other hand, interest category candidate set recall can also be performed at the same time, and the object-related features that have been pulled above can be filtered using the filtering strategy and the target user's determined interest category features, and then the candidate list to be inserted corresponding to the target user (hereinafter referred to as the candidate recommendation set) can be obtained based on the filtered object-related features. Furthermore, after obtaining the reference interest category and the candidate recommendation set, the candidate recommendation set can be categorized according to the reference interest category, and the recommended object (item selection) can be obtained from the candidate recommendation set, and the recommended object can be added to the target user's recommendation list. By using the interest category feature set to predict the interest of the target user's object association features, and filtering the above-pulled object association features through the target user's determined interest category features, the recommendation list generated when the interest exploration is completed can include the user's potential interest points, rather than just the interest points that have been explored multiple times or in the user portrait, thereby improving the efficiency and accuracy of interest exploration, thereby better determining the recommended objects for the user and improving the effectiveness and accuracy of interest recommendations.

[0035] In one embodiment, Figure 1aIn the process, the interest prediction processing of the above-mentioned object association features using the interest category feature set can be realized through the interest exploration model loaded online. Specifically, the interest exploration model can be based on the serialized deep learning model. By constructing a sequence feature extraction structure and using the attention mechanism in the interest exploration modeling to capture the correlation between interest sequences, serialized interest exploration can be realized, thereby significantly improving the exploration efficiency. Among them, constructing a sequence feature extraction structure is to extract user behavior features to form a user behavior feature sequence. At the same time, focusing on the interest exploration of the target user, the training sample weights are adaptively adjusted by defining the exploration degree formula to obtain weighted samples. The formula is used to reduce the weight of samples with the same interest points as the existing interest points. Then, the interest exploration model is trained according to the weighted samples, so that the learning objectives of the interest exploration model are tilted towards different interest points, further improving the exploration efficiency and better determining the recommended objects for users.

[0036] Based on the recommendation processing solution for users provided by the embodiment of this application, the embodiment of this application provides an interest recommendation system, see Figure 1b , the interest recommendation system may include at least one user terminal 101, at least one service device 102 and at least one recommendation processing device 103. Among them, the service device 102 and the recommendation processing device 103 can be terminal devices or servers, and the embodiments of this application are not limited. The user terminal 101 and the service device 102 can be directly or indirectly connected via wired or wireless communication. The service device 102 and the recommendation processing device 103 can be directly or indirectly connected via wired or wireless communication. The user terminal 101 is used to display various news information, videos, etc.; the service device 102 is used to provide news information, videos and other multimedia information to users of the user terminal 101; the recommendation processing device 103 is used to determine the recommended objects recommended to the user terminal 101. The above-mentioned user terminal device 101 can be a smart phone, tablet computer, laptop computer, desktop computer, etc.; the above-mentioned server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), as well as big data and artificial intelligence platforms.

[0037] In a specific implementation, when a user watches a video or browses information through the user terminal 101, the recommendation processing device 103 can obtain multimedia data such as the video watched or information browsed by the user through the service device 102, and obtain the user's object-related feature set from the multimedia data such as the video watched or information browsed by the user. The object-related features in the object-related feature set are then filtered to obtain a first feature set. Then, the first feature set is subjected to interest prediction processing using the interest category feature set to obtain a reference interest category of the target user identifier. The object-related feature set is then filtered using the determined interest category features of the user identifier to obtain a second feature set. Based on the second feature set, a candidate recommendation set corresponding to the target user identifier can be obtained. Then, based on the reference interest category and the candidate recommendation set, a recommendation list for the target user identifier is determined. The recommendation processing device 103 can then send the candidate recommendation set to the service device 102, and the service device 102 then provides the user terminal 101 with the recommended objects in the candidate recommendation set, such as information, videos, commodities, and other Internet objects that can be recommended to the user.

[0038] In one embodiment, the aforementioned target user's object-associated feature set and the recommended list of the finalized target user identifier can both be stored in a blockchain, thereby preventing tampering with the target user's object-associated feature set and the recommended list of the finalized target user identifier. Blockchain is a novel application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. It is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of the information (to prevent counterfeiting) and generate the next block.

[0039] Through the above-mentioned user recommendation processing scheme and interest recommendation system, the interest category feature set is used to predict the interest of the determined first feature set, and the above-mentioned object-related feature set is filtered through the target user's determined interest category features. It can be ensured that the recommendation list generated when the interest exploration is completed includes the user's potential interest points, rather than just the interest points that have been explored multiple times or in the user portrait, so as to better determine the recommendation objects for the user and improve the effectiveness and accuracy of interest recommendations.

[0040] Based on the above-mentioned user recommendation processing scheme and interest recommendation system, the embodiment of the present application provides a method for processing user recommendations. The method described in the embodiment of the present application can be executed by an intelligent device, which can be the recommendation processing device 103 in the above-mentioned interest recommendation system. The recommendation processing device 103 can be a dedicated server or some Internet application server. The Internet application server can not only execute the relevant steps of the embodiment of the present application, but also provide other services. Figure 2 , Figure 2 The provided method for processing user recommendations includes the following steps S201-S204.

[0041] S201: Obtain a set of object-related features for a target user identifier. The target user identifier is a unique identifier used to identify the target user. The target user identifier can be the target user's user account, username, nickname, ID card, etc. The object-related feature set can include one or more object-related features, each of which reflects the target user's interest categories. Object-related features can be one or more of user-side features, item-side features, and contextual features. User-side features can include basic user features, user behavior sequence features, user profile features, and so on. Basic user features can include gender, age, and region; user behavior sequence features can include positive and negative behavior sequence features; and user profile features can include user-preferred interest categories, authors, and tags. Item-side features can include category features for various items. For example, if an item is a cat, the corresponding category for the item cat is animals. Contextual features refer to features associated with positive or negative user behavior features. For example, if a user purchases an item, the positive user behavior feature is the purchase of the item, and the contextual features can include the time and location of the purchase.

[0042] It is understood that the collection and processing of user data, user characteristics, and other information related to this application should be strictly in accordance with the requirements of relevant national laws and regulations in actual application, and the informed consent or separate consent of the user and other information subjects should be obtained. The subsequent use and processing of user data, user characteristics, and other information should be carried out within the scope of authorization of laws and regulations and the information subjects. In addition, when the various embodiments of this application are applied to specific products or technologies, the collection, use, and processing of relevant data, characteristics, and other information should comply with the requirements of national laws and regulations, and comply with the principles of legality, legitimacy, and necessity, and should not involve the acquisition of data types prohibited or restricted by laws and regulations.

[0043] The smart device can obtain the user data of the target user identifier from the local space or the blockchain network, and then obtain the object-related feature set from the user data of the target user identifier. For example, the user data of the target user identifier obtained by the smart device includes likes for sports and likes for fashion; the smart device can obtain user behavior sequence features from the user data. The user behavior sequence features include likes for sports and likes for fashion. In one embodiment, since the main goal of the embodiment of the present application is to explore interests, interest categories other than the existing interest categories can be obtained by jumping out of the existing interest categories for recall through a suitable interest exploration and recall algorithm cluster. Specifically, the smart device can use a suitable interest exploration and recall algorithm cluster to obtain the object-related feature set of the target user identifier. Among them, the interest exploration and recall algorithm cluster may include: CF (collaborative filtering), CF second-order jump exploration, and the like.

[0044] In one embodiment, the smart device can preprocess the object-associated feature set to obtain a first feature set, and execute step S202. The first feature set may include: a user behavior feature set identified by the target user and a user auxiliary identification feature set; different from the general sequence DNN (Deep Neural Networks) model that only uses positive sequences, if negative feedback is not considered, the interest exploration model cannot recognize the state during the user's continuous negative feedback, which will cause the interest exploration to be stagnant. Therefore, in an embodiment of the present application, the user behavior feature set may include positive user behavior sequence features and negative user behavior sequence features, so as to capture the impact of positive and negative feedback on user interests at the same time. The positive user behavior sequence is a sequence composed of behaviors with positive feedback of the target user (it can also be understood as a positive user interest sequence), and the positive user behavior sequence is a sequence generated by arranging multiple positive user behaviors in chronological order; the positive user behavior sequence features may include positive user behavior features, and the position features of the positive user behavior features in the positive user behavior sequence features. Multiple positive user behavior features can be expressed as C 11 、C 12 ...C 1n , the location features of multiple positive user behavior features can be expressed as P 11 , P 12 ...P 1n For example, if the target user likes a movie at 12 o'clock, likes a sports activity at 12:10, and likes a pet at 12:11, these are all positive feedback behaviors. The positive user behavior sequence is as follows: Figure 3a As shown, the positive user behavior sequence includes liking movies, liking sports, and liking pets; liking movies is at position 1 in the positive user behavior sequence, i.e., C 11 =Movies, P 11=1; the like campaign is at position 2 in the positive user behavior sequence, i.e. C 12 =Movies, P 12 =1; Liking the pet is at position 3 in the positive user behavior sequence, i.e. C 13 =Movies, P 13 =3.

[0045] Negative user behavior sequence refers to a sequence of behaviors that generate negative feedback from the target user (it can also be understood as a negative user interest sequence). A negative user behavior sequence is a sequence of multiple negative user behaviors arranged in chronological order. Negative user behavior sequence features include negative user behavior features and their position features in the negative behavior sequence features. Multiple negative user behavior features can be expressed as C 21 、C 22 ...C 2n , the position features of multiple negative user behavior features can be expressed as P 21 , P 22 ...P 2n For example, the target user swipes away parent-child at 12 o'clock and fashion at 12:10, which are both negative feedback behaviors. The negative user behavior sequence is as follows: Figure 3b As shown; the negative user behavior sequence includes swiping away parent-child and swiping away fashion; swiping away parent-child is at position 1 in the negative user behavior sequence, that is, C 21 = Parent-child, P 21 =1; when swiping away, it is still at position 2 in the negative user behavior sequence, that is, C 22 =Fashion, P 22 = 2. The user auxiliary identification feature set may include one or more of context features, user basic class features and user portrait features in user-side features, and item-side features.

[0046] The preprocessing can include either or both of: feature selection on the object-related feature set; and discretization on the object-related feature set. The first feature set obtained through feature selection can provide a basis for enhancing the generalization capability of the model; the first feature set obtained through discretization can facilitate subsequent use of the model for interest exploration.

[0047] (1) The specific implementation method of the smart device selecting features from the object-related feature set may include: the smart device screening the object-related feature set based on the determined interest category features of the target user identifier, and performing deduplication processing on the object-related feature set. The smart device screening the object-related feature set based on the determined interest category features of the target user identifier means: removing object-related features in the object-related feature set that are similar or identical to the determined interest category features. Specifically, the smart device may use an inter-feature correlation evaluation algorithm to calculate the similarity between the object-related features in the object-related feature set and the determined interest category features, and then remove the object-related features that meet the similarity conditions. The inter-feature correlation evaluation algorithm may be a Pearson correlation coefficient evaluation. For example, the determined interest category feature identified by the target user is type A cat; the object association feature set includes type A cat and type B cat; the smart device can respectively calculate the similarity between type A cat and type B cat in the object association feature set and type A cat, and the smart device determines that the similarity between type B cat and type A cat in the object association feature set is greater than 90%, and the similarity between type A cat and type A cat in the object association feature set is also greater than 90%; type A cat and type B cat in the object association feature set are removed. The smart device performs deduplication processing on the object association feature set, which means: there are at least two identical or similar object association features in the object association feature set, and then the smart device deduplicates the at least two identical or similar object association features, that is, only retains one of the at least two object association features. Specifically, the smart device can calculate the similarity between each pair of object association features in the object association feature set, and remove any of the pair of object association features that meet the similarity conditions. For example, the object association feature set includes two type A cats; the smart device can calculate the similarity between the two type A cats. If the similarity between the two type A cats is greater than 99%, the smart device removes duplicates from any of the two type A cats and retains only one type A cat.

[0048] Optionally, when performing feature selection on the object-associated feature set, the smart device may also perform engineering processing on the object-associated feature set. The engineering processing method may include equal-interval and equal-frequency bucketing, missing value processing, etc.

[0049] (2) The specific implementation method of the smart device performing discrete processing on the object-related feature set is: hashing each object-related feature in the object-related feature set to obtain a hash identifier for each object-related feature. For example, the object-related feature set includes male and age 30; the smart device can hash the male to obtain the hash identifier of the male; the age is divided into 0-19 years old and 20-39 years old; the smart device can determine that the age of 30 years old belongs to the age of 20-39 years old, and then the smart device hashes the age of 20-39 years old to obtain the hash identifier corresponding to the age of 30 years old.

[0050] In one embodiment, the smart device may filter the object-related feature set based on the determined interest category features of the target user identifier to obtain a second feature set, and then execute step S203. The second feature set may include user-side features, item-side features, and contextual features, among others. The smart device filtering the object-related feature set based on the determined interest category features of the target user identifier means removing object-related features from the object-related feature set that are similar or identical to the determined interest category features.

[0051] S202: Using the interest category feature set, perform interest prediction processing on the first feature set determined according to the object association feature set to obtain the reference interest category of the target user identifier. The interest categories included in the interest category feature set may refer to the interest categories to be predicted for the target user identifier, the number of reference interest categories is one or more, and the reference interest category may be one or more belonging to the interest category feature set. In a specific implementation, for the target user corresponding to the target user identifier, the smart device may first obtain the interest category feature set about the target user from the local space or the blockchain network, such as the interest category feature set may include interest category pets, interest category ball games, etc.; then use the interest category feature set to perform interest prediction processing on the first feature set to obtain one or more reference interest categories of the target user identifier.

[0052] In one embodiment, since the interest category feature set contains interest categories to be predicted for the target user, the specific implementation method for performing interest prediction processing on the first feature set using the interest category feature set is as follows: the smart device determines the interest category features corresponding to the first feature set, then compares the interest category features corresponding to the first feature set with the interest category feature set, and obtains a reference interest category identified by the target user based on the comparison result. Specifically, if the comparison result indicates that the target interest category features corresponding to the first feature set are similar or identical to the interest category features in the interest category feature set, then the target interest category features are used as the reference interest category identified by the target user.

[0053] In another embodiment, in order to improve the accuracy and efficiency of interest prediction, the reference interest category can be obtained by prediction processing through an interest exploration model. Specifically, the interest category feature set is used to perform interest prediction processing on the first feature set determined according to the object association feature set to obtain the reference interest category identified by the target user. The specific implementation method is: the smart device can call the interest exploration model to perform interest prediction processing on the interest category feature set and the first feature set to obtain the reference interest category identified by the target user.

[0054] S203: Obtain a candidate recommendation set corresponding to the target user identifier using a second feature set determined based on the object-associated feature set. The second feature set is obtained by filtering the object-associated feature set based on the determined interest category features of the target user identifier. The candidate recommendation set may include candidate interest categories, or may include recommended objects corresponding to the candidate interest categories. The recommended objects involved in the embodiments of the present application may include articles, videos, and the like, as well as mobile phones, cats, dogs, balls, and the like, without limitation in the embodiments of the present application.

[0055] In one embodiment, the smart device can analyze the second feature set to obtain the interest category corresponding to each feature in the second feature set, then obtain the recommended object corresponding to the interest category corresponding to each feature, and add the recommended object corresponding to the interest category corresponding to each feature to the candidate recommendation set corresponding to the target user identifier.

[0056] S204: Determine a recommendation list for the target user ID based on the reference interest category and the candidate recommendation set. The smart device may determine a target interest category from one or more reference interest categories, then obtain recommended objects corresponding to the target interest category from the candidate recommendation set, and add the recommended objects corresponding to the target interest category to the target user ID's recommendation list, so that the smart device can subsequently recommend the recommended objects in the recommendation list to the target user ID. For example, if the determined target interest category is pets, recommended objects corresponding to pets may be obtained from the candidate recommendation set, such as cats, dogs, and so on. Cats and dogs are then added to the target user ID's recommendation list.

[0057] In one embodiment, in step S202, when obtaining the reference interest category, an interest score corresponding to the reference interest category may also be obtained. The interest score is used to represent the target user's interest level in the reference interest category. The smart device may determine a recommendation list for the target user based on the interest score corresponding to the reference interest category and the candidate recommendation set.

[0058] As an example, the smart device may determine the target interest category with the highest interest score based on the interest scores corresponding to the reference interest categories; then determine one or more recommended objects corresponding to the target interest category from the candidate recommendation set; and then add the one or more recommended objects corresponding to the target interest category to the recommendation list of the target user identifier. As another implementation, the smart device may determine multiple target interest categories that meet a score threshold from multiple reference interest categories; then obtain one or more recommended objects corresponding to each of the multiple target interest categories that meet the score threshold from the candidate recommendation set; and add these to the recommendation list of the target user identifier.

[0059] In one embodiment, after determining the recommendation list of the target user identifier, the smart device can send the recommendation list to any node in the blockchain network, so that any node can verify the legitimacy of the recommendation list, and after the verification is passed, generate a block including the recommendation list and send it to other nodes for consensus. When the consensus of other nodes is completed, the block is saved in the blockchain network to prevent the data in the block from being tampered with.

[0060] In an embodiment of the present application, a computer is provided to obtain an object-associated feature set identified by a target user; perform interest prediction processing on a first feature set determined according to the object-associated feature set using an interest category feature set to obtain a reference interest category identified by the target user; obtain a candidate recommendation set corresponding to the target user using a second feature set determined according to the object-associated feature set, the second feature set being obtained by filtering the object-associated feature set according to the determined interest category features of the target user; and determine a recommendation list for the target user based on the reference interest category and the candidate recommendation set. By using the interest category feature set to perform interest prediction on the user's object-associated features, and filtering the above-mentioned object-associated features using the target user's determined interest category features, it is ensured that the recommendation list generated when the interest exploration is completed includes the user's potential points of interest, rather than just points of interest that have been explored multiple times or in the user's profile, thereby better determining recommended objects for the user and improving the effectiveness and accuracy of interest recommendations.

[0061] Based on the above-mentioned recommendation processing scheme and interest recommendation system for users, the embodiment of the present application provides a method for recommending users. The method described in the embodiment of the present application can be executed by an intelligent device, which can be the recommendation processing device 103 in the above-mentioned interest recommendation system. The recommendation processing device 103 can be a dedicated server or some Internet application server. The Internet application server can not only execute the relevant steps of the embodiment of the present application, but also provide other services. Figure 4 , Figure 4The provided method for processing user recommendations includes the following steps S401-S404:

[0062] S401: Obtain an object-related feature set of a target user identifier.

[0063] S402: Call the interest exploration model to perform interest prediction processing on the interest category feature set and the first feature set determined according to the object association feature set to obtain the reference interest category identified by the target user. As can be seen from the above, the first feature set may include the user behavior feature set and the user auxiliary identification feature set identified by the target user; wherein, the input of the interest exploration model may include the first feature set, and the output is one or more reference interest categories. The above interest exploration model can be as follows Figure 5a As shown, the interest exploration model can include an input layer, a first model network module, a second model network module, a splicing layer, a third model network module and an output layer from bottom to top. Among them, the input layer can be used to input the user behavior feature set and the user auxiliary identification feature set of the target user identifier. In the embodiment of the present application, the first feature set is discretized on the smart device. The input layer can also be used to input the hash identifier corresponding to each user behavior feature included in the user behavior feature set, and the hash identifier corresponding to each user auxiliary identification feature included in the user auxiliary identification feature set, thereby improving the efficiency of interest prediction processing based on the first feature set determined by the object association feature set. The first model network module can be used to perform feature vector analysis processing on the input user behavior feature set and interest category feature set; the second model network module can be used to perform feature cross processing on the input user auxiliary identification feature set and interest category feature set; the splicing layer can be used to splice the feature vector output by the first model network module and the feature vector output by the second model network module; the third model network module can be used to analyze the spliced feature vector; and the output layer can be used to output the user's reference interest category.

[0064] In one embodiment, the smart device calls the interest exploration model, performs interest prediction processing on the interest category feature set and the first feature set determined according to the object association feature set, and obtains the reference interest category identified by the target user. The specific implementation method is: the smart device uses the first model network module in the interest exploration model to perform feature vector analysis processing on the user behavior feature set and the interest category feature set to obtain the user interest behavior sequence feature vector; and uses the second model network module in the interest exploration model to perform feature cross-processing on the user auxiliary identification feature set and the interest category feature set to obtain the auxiliary identification feature vector; after obtaining the user interest behavior sequence feature vector and the auxiliary identification feature vector, the smart device uses the splicing layer in the interest exploration model to splice the user interest behavior sequence feature vector and the auxiliary identification feature vector, and then uses the third model network module in the interest exploration model to analyze the feature vector obtained after splicing to obtain the reference interest category identified by the target user.

[0065] In one embodiment, the above interest exploration model can also be Figure 5b As shown, in Figure 5bIn the interest exploration model, the first model network module may include: a first embedding layer 51 and an attention mechanism layer, the attention mechanism layer includes a first attention module 53 and a second attention module 54; wherein, the first embedding layer 51 may be an Embedding embedding layer, and the first embedding layer 51 refers to mapping high-dimensional sparse features to low-dimensional semantic space in deep learning. In an embodiment of the present application, the first embedding layer 51 can be used to perform low-dimensional processing on the input user behavior feature set and interest category feature set to obtain a low-dimensional implicit vector of the features in the user behavior feature set and a low-dimensional implicit vector of the features in the interest category feature set. In one embodiment, the features in the user behavior feature set and the features in the interest category feature set can also be used as inputs to the first embedding layer 51 with hash identifiers. The attention mechanism layer (attention layer) is used to perform attention analysis on the feature vector obtained after processing by the embedding layer. At this time, the smart device uses the first model network module in the interest exploration model to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain the user interest behavior sequence feature vector. The specific implementation method is as follows: the smart device uses the first embedding layer 51 to perform feature analysis on the user behavior feature set to obtain the user behavior feature embedding vector; uses the first embedding layer 52 to perform feature analysis on the interest category feature set to obtain the interest category feature embedding vector; uses the attention mechanism layer to perform attention analysis on the user behavior feature embedding vector and the interest category feature embedding vector to obtain the interest weight value; then, based on the user behavior feature embedding vector and the interest weight value, the user interest behavior sequence feature vector can be obtained. In one embodiment, the output of each embedding layer involved in this application can be a corresponding low-dimensional implicit vector, mapping the high-dimensional sparse features of the corresponding input to a low-dimensional semantic space.

[0066] As can be seen from the above, the user behavior feature set includes positive user behavior sequence features and negative user behavior sequence features. The sequence length corresponding to the positive user behavior sequence features and the sequence length corresponding to the negative user behavior sequence features can be set according to the interest exploration requirements. When taking the positive user behavior sequence features and the negative user behavior sequence as input, the position feature of each user behavior feature in the user behavior sequence features can also be introduced to express the impact of the behavior position on the future interest of the target user. In this case, the positive user behavior sequence features include multiple positive user behavior features and the position feature of each positive user behavior feature in the positive behavior sequence features.

[0067] Next, we will use a single positive user behavior feature and the corresponding location feature, and a single negative user behavior feature and the corresponding location feature to explain the modules involved in the entire interest exploration model.

[0068] In one embodiment, the input of the first embedding layer 51 may include positive user behavior features and their positional features, as well as negative user behavior features and their positional features. The smart device then uses the first embedding layer 51 to perform low-dimensional processing on the positive user behavior features to obtain a positive user behavior feature embedding vector. The smart device uses the first embedding layer 51 to perform low-dimensional processing on the positional features of the positive user behavior features to obtain a positional feature vector of the positive user behavior features. The smart device uses the first embedding layer 51 to perform low-dimensional processing on the negative user behavior features to obtain a negative user behavior feature embedding vector. The smart device uses the first embedding layer 51 to perform low-dimensional processing on the positional features of the negative user behavior features to obtain a positional feature vector of the negative user behavior features. At this point, the user behavior feature embedding vector output by the first embedding layer includes: the positive user behavior feature embedding vector and the positional feature vector of the positive user behavior features, the negative user behavior feature embedding vector and the positional feature vector of the negative user behavior features.

[0069] For example, in Figure 5b In the first embedding layer 51, the input includes positive user behavior features C 11 and positive user behavior characteristics C 11 Position feature P 11 ; Negative user behavior characteristics C 21 and negative user behavior characteristics C 21 Position feature P 21 Then the smart device can use the first embedding layer 51 to identify the positive user behavior feature C 11 and the corresponding position feature P 11 Perform low-dimensional processing to obtain the positive user behavior feature embedding vector 502 and the corresponding position feature vector 501; then the smart device performs low-dimensional processing on the negative user behavior feature C 21 and the corresponding position feature P 21 Low-dimensional processing is performed to obtain a negative user behavior feature embedding vector 503 and a corresponding position feature vector 504.

[0070] Furthermore, the interest weight value includes a positive weight value and a negative weight value. After obtaining the positive user behavior feature embedding vector and the position feature vector of the positive user behavior feature, the negative user behavior feature embedding vector and the position feature vector of the negative user behavior feature, the smart device can use the positive user behavior feature embedding vector and the position feature vector of the positive user behavior feature, the negative user behavior feature embedding vector and the position feature vector of the negative user behavior feature as inputs of the attention mechanism layer. Then, the smart device uses the first attention module 53 in the attention mechanism layer to perform attention analysis on the positive user behavior feature embedding vector, the position feature vector of the positive user behavior feature, and the interest category feature embedding vector to obtain a positive interest weight value; and uses the second attention module 54 in the attention mechanism layer to perform attention analysis on the negative user behavior feature embedding vector, the position feature vector of the negative user behavior feature, and the interest category feature embedding vector to obtain a negative interest weight value. Among them, the attention mechanism layer is a model network layer based on the attention mechanism; since the attention mechanism layer in the first model network module is mainly used to model user behavior sequences and incorporate the position features corresponding to each user behavior, in the embodiment of the present application, the attention mechanism layer can also be an RNN (Recurrent Neural Network) algorithm or a Transformed algorithm (a sequence modeling algorithm).

[0071] Following the above example, Figure 5b In the embodiment, the smart device uses the first attention module 53 in the attention mechanism layer to perform attention analysis on the positive user behavior feature embedding vector 502 and the corresponding position feature vector 501, as well as the interest category feature embedding vector, to obtain a positive interest weight value 505 (such as the positive interest weight value 505 is 0.2); and uses the second attention module 54 in the attention mechanism layer to perform attention analysis on the negative user behavior feature embedding vector 503 and the corresponding position feature vector 504, as well as the interest category feature embedding vector, to obtain a negative interest weight value 506 (such as the positive interest weight value 505 is 0.3).

[0072] In one embodiment, the smart device may obtain a user interest behavior sequence feature vector based on the behavior feature embedding vector and the interest weight value by performing a product operation on the behavior feature embedding vector and the interest weight value to obtain the user interest behavior sequence feature vector. Specifically, the smart device may perform a product operation on the positive interest weight value and the positive user behavior feature embedding vector to obtain a positive behavior product quantity, and may perform a product operation on the negative interest weight value and the negative user behavior feature embedding vector to obtain a negative behavior product quantity. The computer device may then concatenate the positive behavior product quantity and the negative behavior product quantity to obtain the user interest behavior sequence feature vector.

[0073] In one embodiment, when there are multiple positive user behavior feature embedding vectors, each positive user behavior feature embedding vector corresponds to a positive interest weight value. The smart device can multiply each positive user behavior feature embedding vector with the corresponding positive interest weight value to obtain multiple positive behavior product quantities, and perform weighted pooling operations on the multiple positive user behavior superposition quantities to obtain a positive pooling vector; similarly, when there are multiple negative user behavior feature embedding vectors, each positive user behavior feature embedding vector corresponds to a negative interest weight value, and the smart device can multiply each negative user behavior feature embedding vector with the corresponding negative interest weight value to obtain multiple negative behavior product quantities, and perform weighted pooling operations on the multiple negative behavior product quantities to obtain a negative pooling vector. The positive pooling vector and the positive pooling vector are then spliced to obtain a user interest behavior sequence feature vector. The user behavior sequence features can be mapped to a fixed-length implicit space through weighted pooling.

[0074] For example, in Figure 5b In the middle, positive user behavior feature C 11 and the corresponding position feature P 11 After passing through the first embedding layer 51 and the first attention module 53 in the attention mechanism layer, the positive interest weight value 505 (for example, the positive interest weight value is 0.3) is obtained; then the positive user behavior feature C 1n and the corresponding position feature P 1n After passing through the first embedding layer 51 and the first attention module 53 in the attention mechanism layer, the positive interest weight value n is obtained (the positive interest weight value n is 0.2); the smart device converts the positive user behavior feature C 11 After the low-dimensional processing of the first embedding layer 51, the positive user behavior feature embedding vector 502 and the positive interest weight value 505 are multiplied to obtain the positive behavior product 1 (i.e., the positive user behavior feature embedding vector 502 multiplied by 0.3); then the smart device converts the positive user behavior feature C 1n After low-dimensional processing by the first embedding layer, the positive user behavior feature embedding vector n is multiplied by the positive interest weight value n to obtain the positive behavior product n (i.e., the positive user behavior feature embedding vector n multiplied by 0.2). The smart device then performs weighted pooling on the positive behavior product 1 and the positive behavior product n to obtain the positive pooling amount (i.e., the positive user behavior feature embedding vector 505 multiplied by 0.3 + ... the positive user behavior feature embedding vector n multiplied by 0.2).

[0075] Similarly, negative user behavior feature C 21 and the corresponding position feature P 21 After passing through the first embedding layer 51 and the second attention module 54 in the attention mechanism layer, the negative interest weight value 506 is obtained; then the negative behavior user feature C2n and the corresponding position feature P 2n After passing through the first embedding layer 51 and the second attention module 54 in the attention mechanism layer, the negative interest weight value n is obtained; then the smart device converts the negative user behavior feature C 21 After the low-dimensional processing of the first embedding layer 51, the negative user behavior feature embedding vector 503 and the negative interest weight value 506 are multiplied to obtain the negative behavior product 1; then the smart device converts the negative user behavior feature C 2n The negative user behavior feature embedding vector n after low-dimensional processing by the first embedding layer is multiplied by the negative interest weight value n to obtain the negative behavior product n; then the smart device performs weighted pooling on the negative behavior product 1 and the negative behavior product n to obtain the negative pooling amount; then the positive pooling amount and the negative pooling amount are spliced to obtain the user interest behavior sequence feature vector.

[0076] In one embodiment, Figure 5b In the interest exploration model, the second model network module may include: a second embedding layer 52 and a feature cross layer; the second embedding layer 52 may be used to perform low-dimensional processing on the input user-assisted identification feature set and the interest category feature set to obtain a low-dimensional implicit vector of the features in the user-assisted identification feature set and a low-dimensional implicit vector of the features in the interest category feature set; in one embodiment, the features in the user-assisted identification feature set and the interest category feature set features may also be hashed as input to the first embedding layer 51. The feature cross layer may be used to perform feature cross processing on the feature vector obtained after processing by the embedding layer. The above-mentioned feature cross layer may be an FM (factor Machine) feature cross layer, in which the second-order feature cross capability of the FM algorithm is used to perform a second-order cross on the user-assisted identification feature set and the interest category feature set, which may enrich the model feature combination and improve the model generalization capability; wherein, the FM algorithm is a machine learning algorithm based on matrix decomposition, which is designed to solve the feature combination problem in large-scale sparse matrices. The calculation formula of the FM algorithm is as follows:

[0077]

[0078] Among them, n represents the number of features, w0, w i represents the model parameters, x i and x j represents the features in the user-assisted identification feature set and the interest category feature set, and v represents the implicit vector.

[0079] It should be noted that the first embedding layer 51 and the second embedding layer 52 may be the same embedding layer or different embedding layers.

[0080] At this time, the auxiliary identification feature vector includes: user portrait features, user basic class features, and context features. The smart device uses the second model network module in the interest exploration model to perform feature cross-processing on the user auxiliary identification feature set and the interest category feature set. The specific implementation method for obtaining the auxiliary identification feature vector is as follows: the smart device uses the second embedding layer 52 in the interest exploration model to perform low-dimensional processing on the user auxiliary identification feature set to obtain a user auxiliary identification feature embedding vector. For example, the user auxiliary identification feature set includes user portrait features, user basic class features, and context features. The smart device uses the second embedding layer 52 to perform low-dimensional processing on the context features to obtain a context feature embedding vector. The smart device then uses the second embedding layer 52 to perform low-dimensional processing on the user portrait features to obtain a user portrait feature embedding vector. The smart device also uses the second embedding layer to perform low-dimensional processing on the user basic class features to obtain a user basic class feature embedding vector. The smart device then uses the second embedding layer 52 to perform low-dimensional processing on the interest category feature set to obtain an interest category feature embedding vector. The smart device then uses the user auxiliary identification feature embedding vector (such as the context feature embedding vector, the user portrait feature embedding vector, and the user basic class feature embedding vector) and the interest category feature embedding vector as the input of the feature cross layer. The smart device performs feature cross processing on the user auxiliary identification feature embedding vector and the interest category feature embedding vector to obtain the auxiliary identification feature vector.

[0081] In the embodiment of the present application, the vectors obtained by the low-dimensional processing of the first embedding layer 51 or the second embedding layer 52 are all low-dimensional implicit vectors.

[0082] In one embodiment, Figure 5b In the interest exploration model, the third model network module may include multiple activation functions, and the multiple activation functions may include a PReLu activation function and a softmax function. The PReLu function refers to a ReLu activation function with parameters. Through the PReLu activation function, the fitting ability of the interest exploration model can be improved and the risk of overfitting can be reduced without adding any additional parameters. In one embodiment, after the smart device splices the user interest behavior sequence feature vector and the auxiliary identification feature vector by the splicing layer in the interest exploration model, the smart device uses two activation functions PReLu and the activation function softmax to analyze the feature vector obtained after splicing to obtain one or more reference interest categories.

[0083] S403: Acquire a candidate recommendation set corresponding to the target user identifier using a second feature set determined according to the object-associated feature set, where the second feature set is obtained by filtering the object-associated feature set according to the determined interest category of the target user identifier.

[0084] In one embodiment, the smart device uses a second feature set determined based on the object-related feature set to obtain an initial candidate recommendation set corresponding to the target user identifier, which is equivalent to evaluating the user's potential interest in commodities, services, multimedia data such as audio and video data, etc. based on some of the user's features, and recording them as preliminary candidate recommendation objects in the preliminary candidate recommendation set. After obtaining the initial candidate recommendation set, the initial candidate recommendation set can be filtered using a filtering strategy as needed to obtain a candidate recommendation set; wherein, filtering the initial candidate recommendation set using a filtering strategy includes one or more of the following: (1) The recommendation system has exposed the same interest category to the target user N times, indicating that the target user has fully browsed the interest category and there is no need to explore the interest category again. For example, if pets are exposed to the target user 5 times, there is no need to explore the recommended objects corresponding to the pet category. In this case, filtering the initial candidate recommendation set using a filtering strategy can be: the smart device can delete the first interest category included in the initial candidate recommendation set, and the first interest category and the interest category that has been recommended to the target user identifier meet the first category consistency condition. wherein, the first category consistency condition can be set according to demand. The first category consistency condition means that the first interest category is similar or identical to an interest category that has been recommended to the target user identifier. Wherein, similar means that the similarity between the first interest category and the interest category that has been recommended to the target user identifier meets the similarity condition, and the similarity does not reach 100%.

[0085] (2) In order to prevent the predicted interest categories from being explored continuously, the predicted interest categories need to be deleted. For example, if the interest category pet has been predicted, the interest category pet needs to be deleted and other interest categories need to be selected for exploration. In this case, the smart device can delete the second interest category included in the initial candidate recommendation set. The second interest category satisfies the second category consistency condition with the interest category that has been predicted for the target user identifier. The second category consistency condition means that the second interest category is similar or identical to the interest category that has been predicted for the target user identifier. Among them, similar means that the similarity between the second interest category and the interest category that has been predicted for the target user identifier meets the similarity condition, and the similarity does not reach 100%. By deleting the interest category that has been predicted for the target user identifier, the distribution diversity of the items corresponding to the interest category can be improved, the distribution efficiency of the long-tail classification can be improved, and the user can be given more interest exploration attempts, which has a good effect on improving the discovery of recommendations.

[0086] (3) Since the interest categories that have already appeared in the user portrait corresponding to the target user identifier also have low exploration value, there is no need to explore them again. In this case, the smart device can delete the third interest category included in the initial candidate recommendation set, and the third interest category and the interest category that appears in the user portrait corresponding to the target user identifier meet the third category consistency condition. The three-category consistency condition means that the first interest category is similar to or the same as the interest category that appears in the user portrait. Among them, similar means that the similarity between the third interest category and the interest category that appears in the user portrait meets the similarity condition, and the similarity does not reach 100%. Among them, the above-mentioned first interest category, second interest category and third interest category can be the same or different.

[0087] S404: Determine a recommendation list for the target user identifier based on the reference interest category and the candidate recommendation set.

[0088] The specific implementation of step S404 can be found in the above embodiment. Figure 2 The specific implementation of step 204 in will not be repeated here.

[0089] In one embodiment, the smart device can also detect in real time whether the overall distribution of the target user's interest categories meets the expected conditions, so as to adjust the interest exploration model. For example, the target user's interest categories are pets and parenting, and the expected conditions are: pets account for 5% of the target user's total interest categories, and parenting accounts for 7% of the target user's total interest categories; the smart device detects in real time that pets account for 50% of the target user's total interest categories and parenting accounts for 7% of the target user's total interest categories. At this time, the smart device determines that the overall distribution of the target user's interest categories does not meet the expected conditions, and can output a prompt message to adjust the online model based on the prompt message.

[0090] In an embodiment of the present application, an interest exploration model is called to perform interest prediction processing on an interest category feature set and a first feature set determined based on an object-associated feature set, thereby obtaining a reference interest category for the target user ID. A second feature set determined based on the object-associated feature set is then used to obtain a candidate recommendation set corresponding to the target user ID. Based on the reference interest category and the candidate recommendation set, a recommendation list for the target user ID is then determined. By constructing a sequential feature extraction structure and utilizing the attention mechanism in the interest exploration model, the correlation between interest categories can be captured, enabling serialized interest exploration and effectively improving the efficiency and accuracy of interest exploration.

[0091] The present application also provides a model training method to obtain an interest exploration model. The architecture diagram of the interest exploration model can be found in the above Figure 5a and Figure 5bWhen performing model training, the model training can be performed on the smart device provided in the embodiment of the present application; or, the model training can be performed on a device other than the smart device in the embodiment of the present application, and then the interest exploration model can be loaded into the above-mentioned smart device. Figure 6 , Figure 6 The provided model training method includes the following steps S601-S602:

[0092] S601: Obtain an interest training sample set. The interest training sample set includes one or more interest sample pairs, each interest sample pair includes: object association training features, interest category training features and annotation data; wherein, the object association training features may include: one or more of user-side training features, item-side training features and context training features. The object association training features are obtained based on the user behavior features obtained in the first time period within the target time range. The annotation data may be an interest category, and the annotation data is determined based on the user behavior features obtained in the second time period within the target time range; in a specific implementation, the smart device may obtain the user characteristic behavior of the user in the first time period within the target time range in a certain recommendation system, and obtain the user behavior characteristics of the user in the second time period within the target time range; then the smart device obtains the object association training features from the user characteristic behavior obtained in the first time period, and obtains the annotation data from the user characteristic behavior obtained in the second time period. wherein, the second time period is after the first time period. The target time range, the first time period and the second time period can be set according to the model training requirements. For example, if a user watches a 20-minute video in a video software within the target time range of 11:00 to 11:20, the smart device can obtain the 10-minute video watched from 11:00 to 11:10 in the first period, and then the smart device can extract object-related training features from the 10-minute video watched by the user. Next, the smart device can obtain the 5-minute sports video watched by the user from 11:15 to 11:20 in the first period, and then the smart device can obtain labeled data from the 5-minute sports video, and the labeled data is sports. Among them, the interest category training features and the interest category features in the interest category feature set mentioned above can be the same.

[0093] The above-mentioned object association training features are obtained based on the user behavior features obtained in the first time period within the target time range, and the labeled data are determined based on the user behavior features obtained in the second time period within the target time range. On the one hand, the user's real user behavior features can be obtained, and the user's other interest categories at the next moment can be predicted based on the user's real user behavior features, thereby ensuring the accuracy of model training; on the other hand, the smart device directly obtains the user's user behavior features and can automatically generate interest sample pairs without the need for manual labeling by the user.

[0094] In one embodiment, the interest training sample set can be generated based on the user behavior characteristics of the same user within multiple target time ranges. Specifically, for any target time range of the target user, the smart device can obtain the user behavior characteristics of the first time period within any target time range from the user data (such as the information and videos watched by the user), and obtain the object association training features from the user behavior characteristics of the first time period; then the smart device obtains the user behavior characteristics of the second time period within any target time range, and obtains the annotation data from the user behavior characteristics of the second time period, and generates an interest sample pair by combining the object association training features, the annotation data, and the interest category training features; according to the above method of generating an interest sample pair within any target time range, the object association training features and annotation data of multiple time ranges can be obtained from the user data, thereby generating multiple interest sample pairs. For example, for user A, the smart device obtains the user behavior characteristics of the first time period within the target time range 1 from the user data, and obtains the object association training feature 1 from the user behavior characteristics of the first time period; then obtains the user behavior characteristics of the second time period within the target time range 1, and obtains the annotation data 1 from the user behavior characteristics of the second time period; then generates an interest sample pair 1 based on the object association training feature 1, the interest category training feature, and the annotation data 1. Then, for user A, the smart device obtains the user behavior features of the first period within the target time range 2 from the user data, and obtains object association training features 2 from the user behavior features of the first period. It then obtains the user behavior features of the second period within the target time range 2, and obtains labeled data 2 from the user behavior features of the second period. An interest sample pair 2 is then generated based on the object association training features 2, the interest category training features, and the labeled data 2. An interest training sample set is then generated based on interest sample pairs 1 and 2.

[0095] In another embodiment, the interest training sample set can be generated based on the user behavior characteristics of different users. The smart device can obtain the user behavior characteristics of different users in the first time period and the second time period within the target time range from a recommendation platform. The smart device can then generate interest sample pairs corresponding to different users in the same manner as described above for generating an interest sample pair within any target time range. For example, for user A, the smart device can obtain the user behavior characteristics of user A in the first period within the target time range from a recommendation platform, and obtain object-association training feature 1 from the user behavior characteristics of the first period; then obtain the user behavior characteristics of user A in the second period within the target time range 1, and obtain labeled data 1 from the user behavior characteristics of the second period; then generate interest sample pair 1 corresponding to user A based on object-association training feature 1, interest category training feature 1 and labeled data 1; for user B, the smart device can obtain the user behavior characteristics of user B in the first period within the target time range from a recommendation platform, and obtain object-association training feature 2 from the user behavior characteristics of the first period; then obtain the user behavior characteristics of user B in the second period within the target time range, and obtain labeled data 2 from the user behavior characteristics of the second period; then generate interest sample pair 2 corresponding to user B based on object-association training feature 2, interest category training feature 2 and labeled data 2.

[0096] It should be noted that when obtaining user behavior characteristics of different users within a target time range to generate different interest sample pairs, the target time ranges can be the same or different. For example, for user A, user behavior characteristics of user A within target time range 1 can be obtained from a recommendation platform; for user B, user behavior characteristics of user B within target time range 2 can be obtained from a recommendation platform.

[0097] In one embodiment, the interest training sample set can be generated based on the user behavior characteristics of the same user within the target time range on different recommendation platforms. For example, for user A, the smart device can obtain the user behavior characteristics within the target time range on recommendation platform 1, and then generate corresponding interest sample pairs in the same manner as generating an interest sample pair within any of the aforementioned target time ranges. Similarly, for user A, the smart device can obtain the user behavior characteristics within the target time range on recommendation platform 2, and then generate corresponding interest sample pairs in the same manner as generating an interest sample pair within any of the aforementioned target time ranges.

[0098] In one embodiment, since the same user logs into different recommendation platforms within different time ranges, the interest training sample set can be generated based on the user behavior characteristics of the same user within different target time ranges on different recommendation platforms. For example, for user A, the smart device can obtain the user behavior characteristics within target time range 1 on recommendation platform 1, and then generate corresponding interest sample pairs in the same manner as generating an interest sample pair within any of the aforementioned target time ranges. Similarly, for user A, the smart device can obtain the user behavior characteristics within target time range 2 on recommendation platform 2, and then generate corresponding interest sample pairs in the same manner as generating an interest sample pair within any of the aforementioned target time ranges.

[0099] It should be noted that, in the embodiments of the present application, the interest training sample set can be generated based on the user behavior characteristics of different users within different target time ranges on different recommendation platforms. The interest training sample set can also be generated based on the user behavior characteristics of different users within different target time ranges on the same recommendation platform. The embodiments of the present application are not limited to this.

[0100] Since the embodiment of this application is mainly about interest exploration, generally the full amount of samples will cause the model to fit the user's known interest categories at the same time and deviate from the training goal. Therefore, it is also designed to set different weight training features by calculating the exploration degree, so that the interest exploration model can be more focused on interest exploration through sample weights. Specifically, the smart device can calculate the learning weight of each interest sample pair based on the interest category features in the user portrait identified by the target user and the interest training sample set; then, when the smart device trains the model, it performs training optimization based on the interest sample pairs in the interest training sample set and the learning weights calculated for the interest sample pairs to obtain an interest exploration model.

[0101] Next, we will take the first interest sample pair in the interest training sample set to introduce how to calculate the learning weight of the first interest sample pair.

[0102] Generally speaking, the higher the exploration degree, the higher the learning weight, that is, the higher the learning weight, the more worthy of exploration the first interest sample pair is. The smart device sets the learning weight of the first interest sample pair by calculating the exploration degree of the first interest sample pair. First, the smart device can calculate the two-dimensional similarity matrix M of each interest category based on the offline collaborative filtering algorithm and perform normalization. The normalized similarity matrix is used to measure the similarity between interest categories. Among them, each interest category includes: interest category training features of interest sample pairs in the interest training sample set, and interest category features in user portraits. The smart device then calculates the exploration degree of the first interest sample pair according to the exploration degree calculation formula. Among them, the exploration degree calculation formula is as follows:

[0103]

[0104] in, Denotes the predicted interest category C for user u t The exploration degree of the exploratory index is interest C t and user interest categories {C1,C2,...,C N} is the weighted mean of the difference between k Represents the interest categories in the user's user profile; N represents the number of interest categories in the user profile.

[0105] After obtaining the exploration degree of the first interest sample pair, the learning weight of the first interest sample pair can be determined based on the principle that the higher the exploration degree, the higher the learning weight. For example, if the exploration degree is 10%-29%, the learning weight is 2; if the exploration degree is 30%-49%, the learning weight is 10; after obtaining that the exploration degree of the first interest sample pair is 32%, the learning weight of the first interest sample pair can be determined to be 10.

[0106] In one embodiment, the target interest has good exploratory properties and has a certain correlation with the user's interest categories. If the exploration degree is directly used to measure the sample weight, the model will tend to recommend interest categories with particularly high exploration degrees, resulting in a poor user experience. Therefore, the present application can also add an activation function after calculating the exploration degree. The activation function is designed as y = αx * e x , its function curve is as follows Figure 7a As shown in the figure, the horizontal axis represents the exploration degree, and the vertical axis represents the sample weight. Initially, the sample weight increases as the exploration degree increases. However, when the exploration degree exceeds the set threshold, the sample weight rapidly decreases, preventing the introduction of highly divergent interest categories. This activation function ensures that the sample weights remain within a certain range, ensuring the accuracy of interest exploration in the interest exploration model obtained after model training.

[0107] It should be noted that any interest sample pair in the interest training sample set can be calculated according to the learning weight calculation method of the first interest sample pair, which will not be repeated here.

[0108] In one embodiment, because the interest training sample set may contain interest category training features that are identical or similar to interest category features in the user profile identified by the target user, the interest training sample set is first processed, so that only different interest sample pairs are subsequently trained by calculating the exploration degree to set different weights. In a specific implementation, the smart device may first process the interest training sample set before calculating the exploration degree of the first interest sample pair to set the learning weight of the first interest sample pair. In a specific implementation, when the interest category training features included in the second interest sample pair in the interest training sample set meet a category similarity condition with the interest category features in the user profile identified by the target user, the second interest sample pair may be deleted or the exploration degree of the second interest sample pair may be set to 0, and the exploration degree calculation for the second interest sample pair may not be performed subsequently, thereby suppressing the model from fitting to the user's existing interests. The category similarity condition may be that the interest category training features included in the second interest sample pair are similar or identical to the interest category features in the user profile identified by the target user. Similarity refers to the interest category training features included in the second interest sample pair meeting the similarity condition with the interest category features in the user profile identified by the target user, but the similarity does not reach 100%.

[0109] S602: Perform model training using the interest training sample set to obtain an interest exploration model.

[0110] In one embodiment, in the process of training an interest exploration model based on an interest training sample set, the interest exploration model can also be trained and optimized based on the interest sample pairs in the interest training sample set and the learning weights calculated for the interest sample pairs; wherein, the learning weight of the first interest sample pair in the interest training sample set is calculated based on the exploration degree of the first interest sample pair, and the exploration degree of the first interest sample pair is determined based on the interest category features in the user portrait identified by the target user and the first interest sample pair.

[0111] In one embodiment, when performing model training, the interest sample pairs in the interest training sample set and the learning weights calculated for the interest sample pairs are input into the initial interest exploration model, wherein the object association training features in the interest sample pairs may include user behavior sequence training features, and the corresponding sequence length of the user behavior sequence training features may be set as required. The user behavior sequence training features may include multiple user behavior training features, and the position features of each user behavior training feature in the user behavior sequence training features. The multiple user behavior training features include positive user behavior training features and negative user behavior training features. Then, the attention mechanism layer of the initial interest exploration model can learn the learning weight corresponding to each user behavior training feature based on each user behavior training feature and the position features of each user behavior training feature in the user behavior sequence training features, so that when performing model training, the training loss of each interest training sample pair is multiplied by the corresponding learning weight, which can make the model pay more attention to the learning of this sample during training. The parameter learning formula in the attention mechanism layer is as follows:

[0112] w t =v t tanh(W c c t +W p p t +b)

[0113]

[0114] Among them, w t represents the learning weight of the t-th user behavior training feature; c t represents the t-th user behavior training feature; v t Is a learning parameter used to complete the operation of calculating the attention weight, p t W represents the position feature of the t-th user behavior training feature; c is the parameter corresponding to the t-th user behavior training feature, W p is the parameter corresponding to the position feature of the t-th user behavior training feature; T is the sequence length of the user behavior sequence training feature.

[0115] Then, during the training of the initial interest exploration model, the model loss value is calculated based on the loss function, the labeled data, and the learning weights; then, in the direction of reducing the model loss value, the parameters in the interest exploration model are updated in the reverse direction to obtain the interest exploration model. Among them, the parameters in the interest exploration model include the above-mentioned W c , W p and b.

[0116] In an embodiment of the present application, the smart device can obtain a set of interest training samples, and then perform model training based on the interest sample set to obtain an interest exploration model. The attention mechanism in the trained interest exploration model can capture the correlation between interest categories and realize serialized interest exploration. It can be ensured that the recommendation objects can be better determined based on the interest exploration model in the future, thereby achieving the effectiveness and accuracy of interest recommendations.

[0117] The interest exploration model and user recommendation processing method after model training provided in the embodiment of the present application can be naturally integrated into the recommendation system in parallel in the form of an interest recommendation module. The recommendation system can be a small world recommendation feed stream. From an engineering perspective, it only adds a small amount of time and has high operating efficiency, such as Figure 7b As shown, the horizontal axis represents interest categories, and the vertical axis represents exposure share. Compared to traditional UCB algorithms, this algorithm improves exploration efficiency by 13%, driving a 2.25% increase in average market time spent and a 2.5% increase in the proportion of new profile users. It also improves item-level category distribution efficiency, significantly increasing the share of weak long-tail categories.

[0118] See also Figure 8 , Figure 8 This is a schematic diagram of the structure of a device for processing user recommendations according to an embodiment of the present application. The device for processing user recommendations according to an embodiment of the present application can be set on a smart device, which can be the above-mentioned Figure 1b The recommendation processing device in the example may also be the user device or the recommendation device. The recommendation processing device for users includes the following units:

[0119] An acquisition unit 801 is configured to acquire an object-related feature set identified by a target user;

[0120] The processing unit 802 is configured to perform interest prediction processing on the first feature set determined according to the object-related feature set using the interest category feature set to obtain a reference interest category identified by the target user;

[0121] The acquisition unit 801 is further configured to acquire a candidate recommendation set corresponding to the target user identifier by using a second feature set determined based on the object-associated feature set, wherein the second feature set is obtained by filtering the object-associated feature set based on the determined interest category features of the target user identifier;

[0122] The processing unit 802 is further configured to determine a recommendation list for the target user identifier based on the reference interest category and the candidate recommendation set.

[0123] In one embodiment, the reference interest category is obtained by predicting an interest exploration model, the input of the interest exploration model includes the first feature set, and the output is one or more reference interest categories; the first feature set includes: a user behavior feature set and a user auxiliary identification feature set of the target user identifier;

[0124] The processing unit 802, when performing interest prediction processing on the first feature set determined according to the object-related feature set using the interest category feature set to obtain the reference interest category of the target user identifier, may be specifically configured to:

[0125] Using the first model network module in the interest exploration model to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain a user interest behavior sequence feature vector;

[0126] Using the second model network module in the interest exploration model, performing feature cross processing on the user auxiliary identification feature set and the interest category feature set to obtain an auxiliary identification feature vector;

[0127] The user interest behavior sequence feature vector and the auxiliary identification feature vector are spliced using the splicing layer in the interest exploration model, and the feature vector obtained after splicing is analyzed using the third model network module in the interest exploration model to obtain the reference interest category of the target user identifier.

[0128] In one embodiment, the first model network module includes a first embedding layer and an attention mechanism layer. When the processing unit 802 uses the first model network module in the interest exploration model to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain the user interest behavior sequence feature vector, it can be specifically used to:

[0129] Performing feature analysis on the user behavior feature set using the first embedding layer to obtain a user behavior feature embedding vector;

[0130] Performing feature analysis on the interest category feature set using the first embedding layer to obtain an interest category feature embedding vector;

[0131] Using the attention mechanism layer to perform attention analysis on the user behavior feature embedding vector and the interest category feature embedding vector to obtain an interest weight value;

[0132] A user interest behavior sequence feature vector is obtained according to the user behavior feature embedding vector and the interest weight value.

[0133] In one embodiment, the user behavior feature set includes a positive user behavior sequence feature and a negative user behavior sequence feature, wherein the positive user behavior sequence feature includes a positive user behavior feature and a position feature of the positive user behavior feature in the positive user behavior sequence feature; the negative user behavior sequence feature includes a negative user behavior feature and a position feature of the negative user behavior feature in the negative user behavior sequence feature;

[0134] The input of the first embedding layer includes positive user behavior features and position features of the positive user behavior features, and negative user behavior features and position features of the negative user behavior features;

[0135] The user behavior feature embedding vector outputted in the first embedding layer includes: a positive user behavior feature embedding vector and a position feature vector of the positive user behavior feature, a negative user behavior feature embedding vector and a position feature vector of the negative user behavior feature.

[0136] In one embodiment, the interest weight value includes a positive interest weight value and a negative interest weight value. When the processing unit 802 uses the attention mechanism layer to perform attention analysis on the user behavior feature embedding vector and the interest category feature embedding vector to obtain the interest weight value, it can be specifically used to:

[0137] Using the first attention module in the attention mechanism layer to perform attention analysis on the positive user behavior feature embedding vector, the position feature vector of the positive user behavior feature, and the interest category feature embedding vector to obtain the positive interest weight value;

[0138] The second attention module in the attention mechanism layer is used to perform attention analysis on the negative user behavior feature embedding vector, the position feature vector of the negative user behavior feature, and the interest category feature embedding vector to obtain the negative interest weight value.

[0139] In one embodiment, before performing interest prediction processing on the first feature set determined according to the object-associated feature set using the interest category feature set to obtain the reference interest category identified by the target user, the processing unit 802 is further configured to:

[0140] Preprocessing the object-associated feature set to obtain a first feature set, wherein the preprocessing includes: performing feature selection on the object-associated feature set; performing any one or both of discrete processing on the object-associated feature set;

[0141] The performing feature selection on the object-associated feature set includes:

[0142] Either one or both of screening the object-related feature set according to the determined interest category feature of the target user identifier and deduplicating the object-related feature set.

[0143] In one embodiment, when the acquisition unit 801 acquires the candidate recommendation set corresponding to the target user identifier using the second feature set determined according to the object-associated feature set, it may be specifically configured to:

[0144] Obtaining an initial candidate recommendation set corresponding to the target user identifier using a second feature set determined according to the object-associated feature set;

[0145] Filtering the initial candidate recommendation set using a filtering strategy to obtain the candidate recommendation set;

[0146] The filtering process of the initial candidate recommendation set by using a filtering strategy includes one or more of the following:

[0147] Deleting a first interest category included in the initial candidate recommendation set, where the first interest category satisfies a first category consistency condition with an interest category that has been recommended to the target user identifier;

[0148] Deleting a second interest category included in the initial candidate recommendation set, where the second interest category satisfies a second category consistency condition with an interest category that has been predicted for the target user identifier;

[0149] The third interest category included in the initial candidate recommendation set is deleted, and the third interest category and the interest category feature appearing in the user portrait corresponding to the target user identifier meet the third category consistency condition.

[0150] In one embodiment, the interest exploration model is trained based on an interest training sample set, and the interest training sample set includes interest sample pairs, and the interest sample pairs include: object association training features, interest category training features and annotation data; the object association training features are obtained based on user behavior characteristics obtained in a first time period within a target time range, and the annotation data is determined based on user behavior characteristics obtained in a second time period within the target time range, and the second time period is after the first time period.

[0151] In one embodiment, the interest exploration model is obtained by training and optimizing based on the interest sample pairs in the interest training sample set and the learning weights calculated for the interest sample pairs;

[0152] Among them, the learning weight of the first interest sample pair in the interest training sample set is calculated based on the exploration degree of the first interest sample pair, and the exploration degree is determined based on the interest category features in the user portrait of the target user identifier and the first interest sample pair.

[0153] In one embodiment, the processing unit 802 is further configured to:

[0154] When the interest category training feature included in the second interest sample pair in the interest training sample set and the interest category feature in the user portrait of the target user identifier meet a category similarity condition, the second interest sample pair is deleted.

[0155] According to one embodiment of the present application, Figure 2 or Figure 4 or Figure 6 Each step involved in the method shown can be performed by Figure 8 The recommendation processing device shown in FIG. 1 is executed by each unit in the user. For example, Figure 2 Step S201 shown is performed by Figure 8 The acquisition unit 801 shown in FIG is executed, and step S202 is performed by Figure 8 The processing unit 802 shown in FIG is executed, and step S203 is performed by Figure 8 The acquisition unit 801 shown in FIG is executed; step S204 is performed by Figure 8 802 shown in . Figure 4 Step S401 shown is performed by Figure 8 The acquisition unit 801 shown in FIG is executed, and step S402 is performed by Figure 8 The processing unit 802 shown in FIG is executed, and step S403 is performed by Figure 8 The acquisition unit 801 shown in FIG is executed; step S404 is performed by Figure 8 802 shown in . Figure 6 Step S601 shown is performed by Figure 8 The acquisition unit 801 shown in FIG is executed, and step S602 is performed by Figure 8 The processing unit 801 shown in FIG. 8 is used for execution.

[0156] According to another embodiment of the present application, Figure 8The various units in the recommendation processing device for users shown can be individually or all combined into one or several other units to constitute, or one (or some) of the units can be further divided into multiple functionally smaller units to constitute, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions. In actual applications, the function of a unit can also be implemented by multiple units, or the functions of multiple units can be implemented by one unit. In other embodiments of the present application, the recommendation processing device for users can also include other units. In actual applications, these functions can also be implemented with the assistance of other units, and can be implemented by the collaboration of multiple units.

[0157] The specific implementation of the processing unit 802 in the embodiment of the present invention can refer to the description of the relevant steps in the above embodiment, and produces the same technical effect, which is not repeated here.

[0158] Based on the above description of the method for processing user recommendations, this application also discloses a smart device. Figure 9 The smart device may include at least a processor 901, an input interface 902, an output interface 903, and a computer storage medium 904. The processor 901, input interface 902, output interface 903, and computer storage medium 904 in the smart device may be connected via a bus or other means.

[0159] The computer storage medium 904 is a memory device in the smart device, which is used to store programs and data. It is understandable that the computer storage medium 904 here can include the built-in storage medium of the smart device, and of course it can also include the extended storage medium supported by the smart device. The computer storage medium 904 provides a storage space, which stores the operating system of the smart device. In addition, one or more instructions suitable for being loaded and executed by the processor 901 are also stored in the storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer storage medium here can be a high-speed RAM memory; optionally, it can also be at least one computer storage medium away from the aforementioned processor. The processor can be called a central processing unit (CPU), which is the core and control center of the smart device, suitable for implementing one or more instructions, specifically loading and executing one or more instructions to realize the corresponding method flow or function.

[0160] In one embodiment, the processor 901 may load and execute one or more first instructions stored in a computer storage medium to implement the corresponding steps of the method in the embodiment of the method for processing user recommendations. In a specific implementation, the processor 901 loads the one or more first instructions in the computer storage medium and executes the following operations:

[0161] Obtaining a set of object-related features identified by a target user;

[0162] Using the interest category feature set, performing interest prediction processing on the first feature set determined according to the object association feature set to obtain a reference interest category identified by the target user;

[0163] Obtaining a candidate recommendation set corresponding to the target user identifier by using a second feature set determined according to the object-associated feature set, wherein the second feature set is obtained by filtering the object-associated feature set according to the determined interest category features of the target user identifier;

[0164] Based on the reference interest category and the candidate recommendation set, a recommendation list for the target user identification is determined.

[0165] In one embodiment, the reference interest category is obtained by predicting an interest exploration model, the input of the interest exploration model includes the first feature set, and the output is one or more reference interest categories; the first feature set includes: a user behavior feature set and a user auxiliary identification feature set of the target user identifier;

[0166] When performing interest prediction processing on the first feature set determined based on the object-associated feature set using the interest category feature set to obtain a reference interest category identified by the target user, the one or more computer programs in the computer storage medium are loaded by the processor 901 and execute the following steps:

[0167] Using the first model network module in the interest exploration model to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain a user interest behavior sequence feature vector;

[0168] Using the second model network module in the interest exploration model, performing feature cross processing on the user auxiliary identification feature set and the interest category feature set to obtain an auxiliary identification feature vector;

[0169] The user interest behavior sequence feature vector and the auxiliary identification feature vector are spliced using the splicing layer in the interest exploration model, and the feature vector obtained after splicing is analyzed using the third model network module in the interest exploration model to obtain the reference interest category of the target user identifier.

[0170] In one embodiment, the first model network module includes a first embedding layer and an attention mechanism layer. When the first model network module in the interest exploration model is used to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain a user interest behavior sequence feature vector, the one or more computer programs in the computer storage medium are loaded by the processor 901 and execute the following steps:

[0171] Performing feature analysis on the user behavior feature set using the first embedding layer to obtain a user behavior feature embedding vector;

[0172] Performing feature analysis on the interest category feature set using the first embedding layer to obtain an interest category feature embedding vector;

[0173] Using the attention mechanism layer to perform attention analysis on the user behavior feature embedding vector and the interest category feature embedding vector to obtain an interest weight value;

[0174] A user interest behavior sequence feature vector is obtained according to the user behavior feature embedding vector and the interest weight value.

[0175] In one embodiment, the user behavior feature set includes a positive user behavior sequence feature and a negative user behavior sequence feature, wherein the positive user behavior sequence feature includes a positive user behavior feature and a position feature of the positive user behavior feature in the positive user behavior sequence feature; the negative user behavior sequence feature includes a negative user behavior feature and a position feature of the negative user behavior feature in the negative user behavior sequence feature;

[0176] The input of the first embedding layer includes positive user behavior features and position features of the positive user behavior features, and negative user behavior features and position features of the negative user behavior features;

[0177] The user behavior feature embedding vector outputted in the first embedding layer includes: a positive user behavior feature embedding vector and a position feature vector of the positive user behavior feature, a negative user behavior feature embedding vector and a position feature vector of the negative user behavior feature.

[0178] In one embodiment, the interest weight value includes a positive interest weight value and a negative interest weight value. When the attention mechanism layer is used to perform attention analysis on the user behavior feature embedding vector and the interest category feature embedding vector to obtain the interest weight value, one or more computer programs in the computer storage medium are loaded by the processor 901 and execute the following steps:

[0179] Using the first attention module in the attention mechanism layer to perform attention analysis on the positive user behavior feature embedding vector, the position feature vector of the positive user behavior feature, and the interest category feature embedding vector to obtain the positive interest weight value;

[0180] The second attention module in the attention mechanism layer is used to perform attention analysis on the negative user behavior feature embedding vector, the position feature vector of the negative user behavior feature, and the interest category feature embedding vector to obtain the negative interest weight value.

[0181] In one embodiment, before performing interest prediction processing on the first feature set determined based on the object-associated feature set using the interest category feature set to obtain the reference interest category identified by the target user, the processor 901 loads one or more computer programs in the computer storage medium and executes the following steps:

[0182] Preprocessing the object-associated feature set to obtain a first feature set, wherein the preprocessing includes: performing feature selection on the object-associated feature set; performing any one or both of discrete processing on the object-associated feature set;

[0183] The performing feature selection on the object-associated feature set includes:

[0184] Either one or both of screening the object-related feature set according to the determined interest category feature of the target user identifier and deduplicating the object-related feature set.

[0185] In one embodiment, when obtaining a candidate recommendation set corresponding to the target user identifier using the second feature set determined based on the object-associated feature set, the processor 901 loads one or more computer programs in the computer storage medium and executes the following steps:

[0186] Obtaining an initial candidate recommendation set corresponding to the target user identifier using a second feature set determined according to the object-associated feature set;

[0187] Filtering the initial candidate recommendation set using a filtering strategy to obtain the candidate recommendation set;

[0188] The filtering process of the initial candidate recommendation set by using a filtering strategy includes one or more of the following:

[0189] Deleting a first interest category included in the initial candidate recommendation set, where the first interest category satisfies a first category consistency condition with an interest category that has been recommended to the target user identifier;

[0190] Deleting a second interest category included in the initial candidate recommendation set, where the second interest category satisfies a second category consistency condition with an interest category that has been predicted for the target user identifier;

[0191] The third interest category included in the initial candidate recommendation set is deleted, and the third interest category and the interest category feature appearing in the user portrait corresponding to the target user identifier meet the third category consistency condition.

[0192] In one embodiment, the interest exploration model is trained based on an interest training sample set, wherein the interest training sample set includes interest sample pairs, and the interest sample pairs include: object association training features, interest category training features, and annotation data;

[0193] The object association training features are obtained based on user behavior features obtained in a first period within a target time range, and the annotation data is determined based on user behavior features obtained in a second period within the target time range, where the second period is after the first period.

[0194] In one embodiment, the interest exploration model is obtained by training and optimizing based on the interest sample pairs in the interest training sample set and the learning weights calculated for the interest sample pairs;

[0195] Among them, the learning weight of the first interest sample pair in the interest training sample set is calculated based on the exploration degree of the first interest sample pair, and the exploration degree is determined based on the interest category features in the user portrait of the target user identifier and the first interest sample pair.

[0196] In one embodiment, one or more computer programs in the computer storage medium are loaded by the processor 901 and execute the following steps: when the interest category training features included in the second interest sample pair in the interest training sample set meet the category similarity condition with the interest category features in the user portrait identified by the target user, the second interest sample pair is deleted.

[0197] The specific implementation of each step executed by the processor 901 in the embodiment of the present invention can refer to the description of the relevant content in the aforementioned embodiment, and can also achieve the same technical effect, which will not be repeated here.

[0198] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. The processor runs the computer program, so that the smart device executes the method provided in the aforementioned embodiment.

[0199] The present application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a smart device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the smart device to perform the method provided in the aforementioned embodiment.

[0200] The steps in the method of the embodiment of the present application can be adjusted in order, combined and deleted according to actual needs.

[0201] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0202] The above disclosure is only a preferred embodiment of the present application, and certainly cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that implementing all or part of the processes of the above embodiment and making equivalent changes in accordance with the claims of this application still fall within the scope of the invention.

Claims

1. A method for processing user recommendations, characterized in that: include: Obtaining a set of object-related features identified by a target user; Using the interest category feature set, performing interest prediction processing on the first feature set determined according to the object association feature set to obtain a reference interest category identified by the target user; Obtaining a candidate recommendation set corresponding to the target user identifier by using a second feature set determined according to the object-associated feature set, wherein the second feature set is obtained by filtering the object-associated feature set according to the determined interest category features of the target user identifier; Determining a recommendation list for the target user identifier based on the reference interest category and the candidate recommendation set; The obtaining of a candidate recommendation set corresponding to the target user identifier by using a second feature set determined according to the object-associated feature set includes: Obtaining an initial candidate recommendation set corresponding to the target user identifier using a second feature set determined according to the object-associated feature set; Filtering the initial candidate recommendation set using a filtering strategy to obtain a candidate recommendation set; The filtering process of the initial candidate recommendation set by using a filtering strategy includes one or more of the following: Deleting a first interest category included in the initial candidate recommendation set, where the first interest category satisfies a first category consistency condition with an interest category that has been recommended to the target user identifier; Deleting a second interest category included in the initial candidate recommendation set, where the second interest category satisfies a second category consistency condition with an interest category that has been predicted for the target user identifier; The third interest category included in the initial candidate recommendation set is deleted, and the third interest category and the interest category feature appearing in the user portrait corresponding to the target user identifier meet the third category consistency condition.

2. The method according to claim 1, wherein The reference interest category is obtained by performing prediction processing on an interest exploration model, the input of the interest exploration model includes the first feature set, and the output is one or more reference interest categories; The first feature set includes: a user behavior feature set and a user auxiliary identification feature set of the target user identifier; The step of performing interest prediction processing on the first feature set determined according to the object-associated feature set using the interest category feature set to obtain a reference interest category identified by the target user includes: Using the first model network module in the interest exploration model to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain a user interest behavior sequence feature vector; Using the second model network module in the interest exploration model, performing feature cross processing on the user auxiliary identification feature set and the interest category feature set to obtain an auxiliary identification feature vector; The user interest behavior sequence feature vector and the auxiliary identification feature vector are spliced using the splicing layer in the interest exploration model, and the feature vector obtained after splicing is analyzed using the third model network module in the interest exploration model to obtain the reference interest category of the target user identifier.

3. The method according to claim 2, wherein The first model network module includes a first embedding layer and an attention mechanism layer. The first model network module in the interest exploration model is used to perform feature vector analysis on the user behavior feature set and the interest category feature set to obtain a user interest behavior sequence feature vector, including: Performing feature analysis on the user behavior feature set using the first embedding layer to obtain a user behavior feature embedding vector; Performing feature analysis on the interest category feature set using the first embedding layer to obtain an interest category feature embedding vector; Using the attention mechanism layer to perform attention analysis on the user behavior feature embedding vector and the interest category feature embedding vector to obtain an interest weight value; A user interest behavior sequence feature vector is obtained according to the user behavior feature embedding vector and the interest weight value.

4. The method according to claim 3, wherein The user behavior feature set includes a positive user behavior sequence feature and a negative user behavior sequence feature, wherein the positive user behavior sequence feature includes a positive user behavior feature and a position feature of the positive user behavior feature in the positive user behavior sequence feature; the negative user behavior sequence feature includes a negative user behavior feature and a position feature of the negative user behavior feature in the negative user behavior sequence feature; The input of the first embedding layer includes positive user behavior features and position features of the positive user behavior features, and negative user behavior features and position features of the negative user behavior features; The user behavior feature embedding vector outputted in the first embedding layer includes: a positive user behavior feature embedding vector and a position feature vector of the positive user behavior feature, a negative user behavior feature embedding vector and a position feature vector of the negative user behavior feature.

5. The method according to claim 4, wherein The interest weight value includes a positive interest weight value and a negative interest weight value, and the attention analysis of the user behavior feature embedding vector and the interest category feature embedding vector by the attention mechanism layer to obtain the interest weight value includes: Using the first attention module in the attention mechanism layer to perform attention analysis on the positive user behavior feature embedding vector, the position feature vector of the positive user behavior feature, and the interest category feature embedding vector to obtain the positive interest weight value; The second attention module in the attention mechanism layer is used to perform attention analysis on the negative user behavior feature embedding vector, the position feature vector of the negative user behavior feature, and the interest category feature embedding vector to obtain the negative interest weight value.

6. The method according to claim 1, wherein Before performing interest prediction processing on the first feature set determined according to the object-associated feature set using the interest category feature set to obtain the reference interest category identified by the target user, the method further includes: Preprocessing the object-associated feature set to obtain a first feature set, wherein the preprocessing includes: performing feature selection on the object-associated feature set; performing any one or both of discrete processing on the object-associated feature set; The performing feature selection on the object-associated feature set includes: Either one or both of screening the object-related feature set according to the determined interest category feature of the target user identifier and deduplicating the object-related feature set.

7. The method according to claim 1, wherein The reference interest category is obtained by predicting the interest exploration model; The interest exploration model is obtained by training based on an interest training sample set, wherein the interest training sample set includes interest sample pairs, and the interest sample pairs include: object association training features, interest category training features, and annotation data; The object association training features are obtained based on user behavior features obtained in a first period within a target time range, and the annotation data is determined based on user behavior features obtained in a second period within the target time range, where the second period is after the first period.

8. The method according to claim 7, wherein The interest exploration model is obtained by training and optimizing based on the interest sample pairs in the interest training sample set and the learning weights calculated for the interest sample pairs; Among them, the learning weight of the first interest sample pair in the interest training sample set is calculated based on the exploration degree of the first interest sample pair, and the exploration degree is determined based on the interest category features in the user portrait of the target user identifier and the first interest sample pair.

9. The method according to claim 7 or 8, wherein When the interest category training feature included in the second interest sample pair in the interest training sample set and the interest category feature in the user portrait of the target user identifier meet a category similarity condition, the second interest sample pair is deleted.

10. A device for processing user recommendations, characterized in that: include: An acquisition unit, configured to acquire an object-related feature set identified by a target user; a processing unit configured to perform interest prediction processing on a first feature set determined according to the object-associated feature set using the interest category feature set to obtain a reference interest category identified by the target user; The acquiring unit is further configured to acquire a candidate recommendation set corresponding to the target user identifier by using a second feature set determined based on the object-associated feature set, wherein the second feature set is obtained by filtering the object-associated feature set based on the determined interest category features of the target user identifier; The processing unit is further configured to determine a recommendation list for the target user identifier based on the reference interest category and the candidate recommendation set; The acquiring unit is configured to, when acquiring the candidate recommendation set corresponding to the target user identifier by using the second feature set determined according to the object-associated feature set: Obtaining an initial candidate recommendation set corresponding to the target user identifier using a second feature set determined according to the object-associated feature set; Filtering the initial candidate recommendation set using a filtering strategy to obtain the candidate recommendation set; The filtering process of the initial candidate recommendation set by using a filtering strategy includes one or more of the following: Deleting a first interest category included in the initial candidate recommendation set, where the first interest category satisfies a first category consistency condition with an interest category that has been recommended to the target user identifier; Deleting a second interest category included in the initial candidate recommendation set, where the second interest category satisfies a second category consistency condition with an interest category that has been predicted for the target user identifier; The third interest category included in the initial candidate recommendation set is deleted, and the third interest category and the interest category feature appearing in the user portrait corresponding to the target user identifier meet the third category consistency condition.

11. A smart device, characterized in that: including a storage device and a processor; The storage device is used to store computer programs; The processor is configured to execute the computer program to implement the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program. When the computer program is executed, the method according to any one of claims 1 to 9 is implemented.

13. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images