Item recommendation method and apparatus
By constructing user profiles and utilizing SENet's dual-tower model, the problems of data sparsity and uneven sample distribution in item recommendation systems are solved, improving the accuracy and efficiency of recommendations and enabling full mining of user-item feature interactions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA MOBILE COMM LTD RES INST
- Filing Date
- 2021-12-20
- Publication Date
- 2026-06-12
AI Technical Summary
Existing item recommendation systems suffer from massive sparse data and uneven sample distribution, resulting in low learning efficiency, poor network scalability, insufficient mining of user-item feature interaction information, and low recommendation accuracy.
By acquiring user personal attribute information, item tag attribute information, and user item interaction information, user profiles are constructed, and users are divided into groups with different activity levels. The SENet dual-tower model is used for training, and a similarity metric function is used to measure the similarity between user and item embeddings to recommend items similar to the user.
It alleviates the problems of data sparsity and uneven sample distribution, improves network training efficiency and scalability, fully explores the feature interactions between users and items, and improves the accuracy of recommendations.
Smart Images
Figure CN116304285B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence, and in particular to a method and apparatus for recommending items. Background Technology
[0002] Item recommendation systems are a type of information filtering system that recommends items of interest to users by analyzing their historical behavior data and item attribute data. With the development of deep learning, collaborative filtering recommendation methods based on dual-tower network models have received widespread attention. The main idea is to input user feature information and item-related feature information into a dual-tower model, train the network, and obtain items of interest to the user based on the distance between the user and the item in the embedding feature space. Therefore, the selection of user and item features and the optimization of the user-item embedding are crucial.
[0003] (1) Considering the massive sparse data phenomenon in the item recommendation system, when using user click data for network training, the uneven distribution of samples and the long-tailed low-frequency noise data will affect the learning of the network, reduce the learning efficiency and network scalability.
[0004] (2) The deep network model based on dual towers showed good performance in the context of big data. The model was trained by building user tower and item tower separately, and the item embeddings that are close to the user in the embedding space were recommended to the user. However, this method learns the features of users and items separately, which makes the interaction features between user and item features insufficient and the recommendation accuracy low.
[0005] The existing item recommendation method has the following shortcomings:
[0006] 1. The problem of massive sparse data and uneven sample distribution reduces learning efficiency and network scalability.
[0007] 2. The network learning process, which separates user and item features, is insufficient to uncover the interaction features between user and item features, resulting in low recommendation accuracy. Summary of the Invention
[0008] The technical problem to be solved by the present invention is to provide a method and apparatus for recommending items with high accuracy.
[0009] To address the aforementioned technical problems, embodiments of the present invention provide the following technical solutions:
[0010] On the one hand, a method for recommending items is provided, including:
[0011] Obtain user personal attribute information, item tag attribute information, and user item interaction information;
[0012] A user profile is determined based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information. The user profile includes the association between the user and the tag.
[0013] Based on the user profile, users are divided into at least one user group, and each user group includes multiple types of users, with different types of users interacting with items at different frequency ranges.
[0014] Construct a dual-tower model based on SENet, and use the interaction information of the multiple types of users and items to construct the training set and test set of the dual-tower model;
[0015] The dual-tower model is trained using the training and test sets.
[0016] Using the trained dual-tower model, items that are similar to the user in the embedding space are used as recommended items.
[0017] In some embodiments, determining a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information includes:
[0018] The user item interaction information is matched with the user personal attribute information, so that each user item interaction information carries the user personal attribute information.
[0019] The user item interaction information is matched with the item's tag attribute information, so that each user item interaction information carries the item's tag attribute information;
[0020] Based on user-item interaction information carrying user personal attribute information and user-item interaction information carrying item tag attribute information, obtain the item tag information of each user with the highest interaction frequency and interest, and each user's personal attribute information, establish the association relationship between users and tags, and generate the user profile.
[0021] In some embodiments, dividing users into at least one user group based on the user profile includes:
[0022] User feature information is obtained using the user profile, and the user feature information is then feature-encoded.
[0023] Users are divided into K user groups using the k-means clustering algorithm and encoded user feature information, where K is a positive integer and is set according to the size of the user group in the dataset.
[0024] In some embodiments, each of the user groups is divided into the following four categories of users:
[0025] Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period.
[0026] Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold;
[0027] Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0.
[0028] New users will have 0 interactions with items within a preset time period.
[0029] In some embodiments, constructing the training and testing sets of the dual-tower model using the multi-type user-item interaction information includes:
[0030] The user characteristics and item characteristics of 90% of the active users and moderately active users in each user group are used to form a training set;
[0031] The user and item characteristics of 10% of the active and moderately active users in each user group are used as the validation set, and the user and item characteristics of the inactive users in each user group are used as the test set.
[0032] In some embodiments, obtaining items in the embedding space that are close to the user includes:
[0033] The training samples in the training set are passed through the SENet layer, and then through a dual-tower model consisting of two DNNs with shared parameters to obtain user embeddings and item embeddings.
[0034] The similarity metric function is used to measure the similarity θ between user and item embeddings. The formula is:
[0035]
[0036] Where u is the embedding of user features and t is the embedding of item features.
[0037] In some embodiments, after using the trained dual-tower model to obtain items similar to the user in the embedding space as recommended items, the method further includes:
[0038] Users are divided into at least one user group using the aforementioned personal attribute information;
[0039] Identify M active users within the same user group who are similar to the new user, where M is an integer greater than 1;
[0040] The dual-tower model is used to predict recommended items that the M active users are interested in;
[0041] The M recommended items are weighted and summed to obtain the items recommended to new users.
[0042] In some embodiments, after using the trained dual-tower model to obtain items similar to the user in the embedding space as recommended items, the method further includes:
[0043] The user characteristics and interactive item characteristics of low-activity users are input into the dual-tower model to obtain the prediction results;
[0044] Users are divided into at least one user group using the aforementioned personal attribute information;
[0045] Identify M active users within the same user group who are similar to the new user, where M is an integer greater than 1;
[0046] The dual-tower model is used to predict recommended items that the M active users are interested in;
[0047] The prediction result is weighted and added to the M recommended items to obtain the items recommended to inactive users.
[0048] This invention also provides an item recommendation device, comprising:
[0049] The acquisition module is used to acquire user personal attribute information, item tag attribute information, and user item interaction information;
[0050] The first processing module is used to determine a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information. The user profile includes the association between the user and the tag.
[0051] The second processing module is used to divide users into at least one user group according to the user profile. Each user group includes multiple types of users, and the interaction frequency of different types of users with items is in different ranges.
[0052] The third processing module is used to construct a dual-tower model based on SENet, and to construct the training set and test set of the dual-tower model using the interaction information of the multiple types of users and items.
[0053] A training module is used to train the dual-tower model using the training set and the test set;
[0054] The prediction module is used to obtain items similar to the user in the embedding space as recommended items using the trained dual-tower model.
[0055] In some embodiments, the first processing module is specifically used to match the user item interaction information with the user personal attribute information, so that each user item interaction information carries the user personal attribute information; match the user item interaction information with the item tag attribute information, so that each user item interaction information carries the item tag attribute information; based on the user item interaction information carrying the user personal attribute information and the user item interaction information carrying the item tag attribute information, obtain the item tag information that each user interacts with most frequently and is interested in and the personal attribute information of each user, establish the association relationship between users and tags, and generate the user profile.
[0056] In some embodiments, the second processing module is specifically used to obtain user feature information using the user profile, perform feature encoding on the user feature information, and divide the users into K user groups using the k-means clustering algorithm and the encoded user feature information, where K is a positive integer and is set according to the size of the number of users in the dataset.
[0057] In some embodiments, each of the user groups is divided into the following four categories of users:
[0058] Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period.
[0059] Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold;
[0060] Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0.
[0061] New users will have 0 interactions with items within a preset time period.
[0062] In some embodiments, the third processing module is specifically used to form a training set of user features and item features of 90% of the active and moderately active users in each user group; to form a validation set of user features and item features of 10% of the active and moderately active users in each user group; and to form a test set of user features and item features of low-activity users in each user group.
[0063] In some embodiments, the prediction module is specifically used to pass the training samples in the training set through an SENet layer, and then through a dual-tower model composed of two DNNs with shared parameters to obtain user embeddings and item embeddings; a similarity metric function is used to measure the similarity θ between user and item embeddings, as shown in the formula:
[0064]
[0065] Where u is the embedding of user features and t is the embedding of item features.
[0066] In some embodiments, the prediction module is further configured to: divide the user into at least one user group using the user's personal attribute information; identify M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and weight and sum the M recommended items to obtain the items recommended to the new user.
[0067] In some embodiments, the prediction module is further configured to input the user characteristics and interactive item characteristics of low-activity users into the dual-tower model to obtain prediction results; divide users into at least one user group using the user's personal attribute information; identify M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and add the prediction results to the M recommended items in a weighted sum to obtain items recommended to low-activity users.
[0068] This invention also provides an item recommendation device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, it implements the item recommendation method as described above.
[0069] In some embodiments, the processor is specifically configured to match the user item interaction information with the user personal attribute information, so that each user item interaction information carries the user personal attribute information; match the user item interaction information with the item tag attribute information, so that each user item interaction information carries the item tag attribute information; based on the user item interaction information carrying the user personal attribute information and the user item interaction information carrying the item tag attribute information, obtain the item tag information that each user interacts with most frequently and is interested in, and the personal attribute information of each user, establish the association relationship between users and tags, and generate the user profile.
[0070] In some embodiments, the processor is specifically used to obtain user feature information using the user profile, perform feature encoding on the user feature information, and divide the users into K user groups using the k-means clustering algorithm and the encoded user feature information, where K is a positive integer and is set according to the size of the number of users in the dataset.
[0071] In some embodiments, each of the user groups is divided into the following four categories of users:
[0072] Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period.
[0073] Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold;
[0074] Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0.
[0075] New users will have 0 interactions with items within a preset time period.
[0076] In some embodiments, the processor is specifically configured to form a training set using user features and item features of 90% of the active and moderately active users in each user group; to use the user features and item features of 10% of the active and moderately active users in each user group as a validation set; and to use the user features and item features of the inactive users in each user group as a test set.
[0077] In some embodiments, the processor is specifically used to pass the training samples in the training set through an SENet layer, and then through a dual-tower model composed of two DNNs with shared parameters to obtain user embeddings and item embeddings; a similarity metric function is used to measure the similarity θ between user and item embeddings, as shown in the formula:
[0078]
[0079] Where u is the embedding of user features and t is the embedding of item features.
[0080] In some embodiments, the processor is further configured to: divide the user into at least one user group using the user's personal attribute information; identify M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and weight and sum the M recommended items to obtain items recommended to the new user.
[0081] In some embodiments, the processor is further configured to input the user characteristics and interactive item characteristics of inactive users into the dual-tower model to obtain prediction results; divide users into at least one user group using the user's personal attribute information; identify M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and add the prediction results to the M recommended items in a weighted sum to obtain items recommended to inactive users.
[0082] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the item recommendation method described above.
[0083] The embodiments of the present invention have the following beneficial effects:
[0084] In the above scheme, users are classified according to user profiles and user activity, and the training and test sets of the dual-tower model are constructed using the interaction information between different types of users and items. This alleviates the problem of data sparsity and the impact of low-frequency noise data with uneven sample distribution and long tail. It is efficient and improves the training efficiency and scalability of the network. In addition, it can fully explore the feature interactions between users and items, thereby improving the accuracy of recommendations. Attached Figure Description
[0085] Figures 1-3 This is a flowchart illustrating the item recommendation method according to an embodiment of the present invention;
[0086] Figure 4 This is a schematic diagram of the item recommendation device according to an embodiment of the present invention;
[0087] Figure 5 This is a schematic diagram of the composition of the item recommendation device according to an embodiment of the present invention. Detailed Implementation
[0088] To make the technical problems, technical solutions and advantages of the embodiments of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0089] This invention provides a method for recommending items, such as... Figure 1 As shown, it includes:
[0090] Step 101: Obtain user personal attribute information, item tag attribute information, and user item interaction information;
[0091] Step 102: Determine a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information. The user profile includes the association between the user and the tags.
[0092] Step 103: Divide users into at least one user group based on the user profile. Each user group includes multiple types of users, and the interaction frequency of different types of users with items is in different ranges.
[0093] Step 104: Construct a dual-tower model based on SENet, and use the interaction information of the multiple types of users and items to construct the training set and test set of the dual-tower model;
[0094] Step 105: Train the dual-tower model using the training set and test set;
[0095] Step 106: Using the trained dual-tower model, obtain items in the embedding space that are similar to the user as recommended items.
[0096] In this embodiment, users are classified according to their profiles and activity levels. The training and testing sets of the dual-tower model are constructed using the interaction information between different user types and items. This alleviates the problem of data sparsity and the impact of low-frequency noise data with uneven sample distribution and long tails. It is highly efficient and improves the training efficiency and scalability of the network. In addition, it can fully explore the feature interactions between users and items, thereby improving the accuracy of recommendations.
[0097] In some embodiments, determining a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information includes:
[0098] The user item interaction information is matched with the user personal attribute information, so that each user item interaction information carries the user personal attribute information.
[0099] The user item interaction information is matched with the item's tag attribute information, so that each user item interaction information carries the item's tag attribute information;
[0100] Based on user-item interaction information carrying user personal attribute information and user-item interaction information carrying item tag attribute information, obtain the item tag information of each user with the highest interaction frequency and interest, and each user's personal attribute information, establish the association relationship between users and tags, and generate the user profile.
[0101] In some embodiments, dividing users into at least one user group based on the user profile includes:
[0102] User feature information is obtained using the user profile, and the user feature information is then feature-encoded.
[0103] Users are divided into K user groups using the k-means clustering algorithm and encoded user feature information, where K is a positive integer and is set according to the size of the user group in the dataset.
[0104] In some embodiments, each of the user groups is divided into the following four categories of users:
[0105] Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period.
[0106] Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold;
[0107] Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0.
[0108] New users will have 0 interactions with items within a preset time period.
[0109] In the above embodiments, the user group is divided into four categories. In practical applications, the user group is not limited to four categories; it can also be divided into two, three, five, or more categories. The preset time period, the first preset threshold, and the second preset threshold can be set as needed.
[0110] In some embodiments, constructing the training and testing sets of the dual-tower model using the various user-item interaction information includes:
[0111] The user characteristics and item characteristics of 90% of the active users and moderately active users in each user group are used to form a training set;
[0112] The user and item characteristics of 10% of the active and moderately active users in each user group are used as the validation set, and the user and item characteristics of the inactive users in each user group are used as the test set.
[0113] In this embodiment, users with similar interests are grouped into the same group based on user profiles. Active and moderately active users from each group are incorporated into the training of the SENet-based dual-tower attention network model to achieve the task of predicting user clicks on items, mitigating the impact of data sparsity and uneven sample distribution. The SENet-based attention mechanism plays a crucial role in the feature intersection of User Embedding and Item Embedding, better representing the feature interaction information between users and items and improving recommendation performance.
[0114] In some embodiments, obtaining items in the embedding space that are close to the user includes:
[0115] The training samples in the training set are passed through the SENet layer, and then through a dual-tower model consisting of two DNNs with shared parameters to obtain user embeddings and item embeddings.
[0116] The similarity metric function is used to measure the similarity θ between user and item embeddings. The formula is:
[0117]
[0118] Where u is the embedding of user features and t is the embedding of item features.
[0119] In some embodiments, after using the trained dual-tower model to obtain items similar to the user in the embedding space as recommended items, the method further includes:
[0120] Users are divided into at least one user group using the aforementioned personal attribute information;
[0121] Identify M active users within the same user group who are similar to the new user, where M is an integer greater than 1;
[0122] The dual-tower model is used to predict recommended items that the M active users are interested in;
[0123] The M recommended items are weighted and summed to obtain the items recommended to new users.
[0124] In some embodiments, after using the trained dual-tower model to obtain items similar to the user in the embedding space as recommended items, the method further includes:
[0125] The user characteristics and interactive item characteristics of low-activity users are input into the dual-tower model to obtain the prediction results;
[0126] Users are divided into at least one user group using the aforementioned personal attribute information;
[0127] Identify M active users within the same user group who are similar to the new user, where M is an integer greater than 1;
[0128] The dual-tower model is used to predict recommended items that the M active users are interested in;
[0129] The prediction result is weighted and added to the M recommended items to obtain the items recommended to inactive users.
[0130] like Figure 2 and Figure 3 As shown, in one specific embodiment, the item recommendation method includes the following steps:
[0131] Step 1: Obtain relevant user personal attribute information, item tag attribute information, and user item interaction information. The user attribute information is U = [U1, U2, ..., U...]. i ,…,U N ], u i This represents the attribute information of a user; the item attribute information is T = [T1, T2, ..., T]. i ,…,T N ], t i The attribute information of an item; user item interaction information UT = [UT1, UT2, ..., UT2] i ,…,UT N ] for, uti Let N be a set of users and items that generate interactions. N is the number of items that generate user interactions.
[0132] Step 2: Analyze the information obtained in Step 1 to obtain a user profile;
[0133] Specifically, user item interaction information is matched with user personal attribute information, so that each interaction carries user-specific personal attribute information; item tag attribute information is matched with user item interaction information, so that each interaction adds item-specific tag attribute information. Based on the interaction information with added item-specific tag attribute information and user-specific personal attribute information, the most frequently interacted and interested item tag information and each user-specific personal attribute information are obtained, thereby establishing the association characteristics between users and tags and generating user profiles.
[0134] Step 3: Use clustering algorithms to group users;
[0135] Based on the user profile generated in step 2 above, user feature information is obtained and feature encoding is performed. The k-means clustering algorithm is used to divide the users into K groups, where K is set according to the size of the number of users in the dataset.
[0136] Step 4: Based on the user groups generated in Step 3 above, divide the users in each group into the following four categories:
[0137] (1) Active users: Users who interact with items frequently, specifically those who interact with items more than 50 times per week.
[0138] UT i >50
[0139] (2) Medium-active users: Users who interact with items at a moderate frequency, specifically those who interact with items approximately 20 to 50 times per week.
[0140] 20 < UT i <50
[0141] (3) Low-activity users: Users who interact with items less frequently, specifically defined as those who interact with items less than 20 times per week.
[0142] UT i <20
[0143] (4) New users: Users who have not interacted with the system, i.e.:
[0144] UT i =0
[0145] Step 5: Build a training set and a test set for the model based on the user interaction information with items in each group;
[0146] Specifically, the user and item features of 90% of the active and moderately active users in each group can be used as the training set; the user and item features of 10% of the active and moderately active users in each group can be used as the validation set; and the user and item features of the inactive users in each group can be used as the test set.
[0147] Step 6: Input the training set constructed in Step 5 as follows Figure 3 The dual-tower model based on SENet shown predicts items of interest to users based on items that are close to them in the embedding space. Here, DNN is a deep neural network, Item Embedding is the item embedding, and User Embedding is the user embedding.
[0148] First, the training samples in the training set are passed through an SENet layer that dynamically suppresses low-frequency invalid features within user or item data. Then, they are passed through a dual-tower model consisting of two DNNs that can share parameters, ultimately obtaining user embeddings and item embeddings, such as... Figure 2 As shown.
[0149] The similarity measure function is used to measure the similarity between user and item embeddings. The formula is as follows:
[0150]
[0151] Where u is the embedding of user features, and t is the embedding similarity of item features.
[0152] In addition, this embodiment uses different schemes to recommend new users and inactive users:
[0153] (1) New User Recommendation:
[0154] Users are divided into multiple groups using user attribute information. Within each new user's group, approximately M similar active users are identified. Figure 3 The model shown predicts recommended items from M similar active users. The recommended items from the M active users are weighted, and the K items with the larger weights are recommended to new users, thereby alleviating the problem of cold start. The weights of the recommended items can be set according to actual needs.
[0155] This embodiment utilizes models and neighboring user information to optimize the user cold start problem, enabling personalized recommendations for new users.
[0156] (2) Recommendations for inactive users:
[0157] Considering that inactive users have limited interaction information for items, this embodiment adopts a semi-supervised strategy: first, the user characteristics and interactive item characteristics of inactive users are input... Figure 3 The model shown yields prediction results; users are divided into multiple groups using user attribute information, and within the group of inactive users, approximately M similar active users are obtained. Figure 3 The model shown predicts recommended items for M similar active users. The recommended item lists of the M active users are weighted and summed with the prediction results to obtain the final result.
[0158] In this embodiment, users are segmented based on user profiles and activity levels. Active and moderately active users are extracted for network model training, mitigating data sparsity and the impact of long-tailed low-frequency noise due to uneven sample distribution. This approach is highly efficient, improving network training efficiency and scalability. Furthermore, this embodiment utilizes high-frequency user and item features from active and moderately active users, combined with a dual-tower model based on feature-enhanced branch networks, to strengthen the learning of high-frequency features and fully explore the feature interactions between users and items, thereby improving recommendation accuracy.
[0159] This invention also provides an item recommendation device, such as... Figure 4 As shown, it includes:
[0160] The acquisition module 11 is used to acquire user personal attribute information, item tag attribute information, and user item interaction information;
[0161] The first processing module 12 is used to determine a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information. The user profile includes the association between the user and the tag.
[0162] The second processing module 13 is used to divide users into at least one user group according to the user profile. Each user group includes multiple types of users, and the interaction frequency of different types of users with items is in different ranges.
[0163] The third processing module 14 is used to construct a dual-tower model based on SENet, and to construct the training set and test set of the dual-tower model using the interaction information of the multiple types of users and items.
[0164] Training module 15 is used to train the dual-tower model using the training set and the test set;
[0165] The prediction module 16 is used to obtain items similar to the user in the embedding space as recommended items using the trained dual-tower model.
[0166] In some embodiments, the first processing module 12 is specifically used to match the user item interaction information with the user personal attribute information, so that each user item interaction information carries the user personal attribute information; match the user item interaction information with the item tag attribute information, so that each user item interaction information carries the item tag attribute information; based on the user item interaction information carrying the user personal attribute information and the user item interaction information carrying the item tag attribute information, obtain the item tag information that each user interacts with most frequently and is interested in and the personal attribute information of each user, establish the association relationship between users and tags, and generate the user profile.
[0167] In some embodiments, the second processing module 13 is specifically used to obtain user feature information using the user profile, perform feature encoding on the user feature information, and divide the users into K user groups using the k-means clustering algorithm and the encoded user feature information, where K is a positive integer and is set according to the size of the number of users in the dataset.
[0168] In some embodiments, each of the user groups is divided into the following four categories of users:
[0169] Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period.
[0170] Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold;
[0171] Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0.
[0172] New users will have 0 interactions with items within a preset time period.
[0173] In some embodiments, the third processing module 14 is specifically used to form a training set of user features and item features of 90% of the active and moderately active users in each user group; to use the user features and item features of 10% of the active and moderately active users in each user group as a validation set; and to use the user features and item features of the inactive users in each user group as a test set.
[0174] In some embodiments, the prediction module 16 is specifically used to pass the training samples in the training set through an SENet layer, and then through a dual-tower model composed of two DNNs with shared parameters to obtain user embeddings and item embeddings; and to use a similarity metric function to measure the similarity θ between user and item embeddings, as shown in the formula:
[0175]
[0176] Where u is the embedding of user features and t is the embedding of item features.
[0177] In some embodiments, the prediction module 16 is further configured to: divide the user into at least one user group using the user's personal attribute information; identify M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and weight and sum the M recommended items to obtain the items recommended to the new user.
[0178] In some embodiments, the prediction module 16 is further configured to input the user characteristics and interactive item characteristics of low-activity users into the dual-tower model to obtain prediction results; divide users into at least one user group using the user's personal attribute information; determine M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and add the prediction results to the M recommended items in a weighted sum to obtain items recommended to low-activity users.
[0179] This invention also provides an item recommendation device, such as... Figure 5 As shown, it includes a memory 21, a processor 22, and a computer program stored on the memory 21 and executable on the processor 22; when the processor 22 executes the program, it implements the item recommendation method as described above.
[0180] In some embodiments, the processor 22 is specifically configured to match the user item interaction information with the user personal attribute information, so that each user item interaction information carries the user personal attribute information; match the user item interaction information with the item tag attribute information, so that each user item interaction information carries the item tag attribute information; based on the user item interaction information carrying the user personal attribute information and the user item interaction information carrying the item tag attribute information, obtain the item tag information that each user interacts with most frequently and is interested in and the personal attribute information of each user, establish the association relationship between users and tags, and generate the user profile.
[0181] In some embodiments, the processor 22 is specifically used to obtain user feature information using the user profile, perform feature encoding on the user feature information, and divide the users into K user groups using the k-means clustering algorithm and the encoded user feature information, where K is a positive integer and is set according to the size of the number of users in the dataset.
[0182] In some embodiments, each of the user groups is divided into the following four categories of users:
[0183] Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period.
[0184] Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold;
[0185] Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0.
[0186] New users will have 0 interactions with items within a preset time period.
[0187] In some embodiments, the processor 22 is specifically configured to form a training set using the user features and item features of 90% of the active and moderately active users in each user group; to use the user features and item features of 10% of the active and moderately active users in each user group as a validation set; and to use the user features and item features of the inactive users in each user group as a test set.
[0188] In some embodiments, the processor 22 is specifically used to pass the training samples in the training set through an SENet layer, and then through a dual-tower model composed of two DNNs with shared parameters to obtain user embeddings and item embeddings; and to use a similarity metric function to measure the similarity θ between user and item embeddings, as shown in the formula:
[0189]
[0190] Where u is the embedding of user features and t is the embedding of item features.
[0191] In some embodiments, the processor 22 is further configured to: divide the user into at least one user group using the user's personal attribute information; identify M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and weight and sum the M recommended items to obtain items recommended to the new user.
[0192] In some embodiments, the processor 22 is further configured to input the user characteristics and interactive item characteristics of low-activity users into the dual-tower model to obtain prediction results; divide users into at least one user group using the user's personal attribute information; determine M active users similar to the new user within the same user group, where M is an integer greater than 1; predict recommended items that the M active users are interested in using the dual-tower model; and add the prediction results to the M recommended items in a weighted sum to obtain items recommended to low-activity users.
[0193] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the item recommendation method described above.
[0194] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage, or any other non-transferable medium that can be used to store information accessible to the computer-readable terminal device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0195] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for recommending items, characterized in that, include: Obtain user personal attribute information, item tag attribute information, and user item interaction information; A user profile is determined based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information. The user profile includes the association between the user and the tag. Based on the user profile, users are divided into at least one user group, and each user group includes multiple types of users, with different types of users interacting with items at different frequency ranges. Construct a dual-tower model based on SENet, and use the interaction information of the multiple types of users and items to construct the training set and test set of the dual-tower model; The dual-tower model is trained using the training and test sets. Using the trained dual-tower model, items that are similar to the user in the embedding space are used as recommended items; Each of these user groups is divided into the following four categories of users: Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period. Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold; Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0. New users will have 0 interactions with items within a preset time period; The construction of the training and testing sets for the dual-tower model using the interaction information of the various types of users and items includes: The user characteristics and item characteristics of 90% of the active users and moderately active users in each user group are used to form a training set; The user and item characteristics of 10% of the active users and medium-active users in each user group are used as the validation set, and the user and item characteristics of low-active users in each user group are used as the test set. Among them, the items in the Embedding space that are similar to the user include: The training samples in the training set are passed through the SENet layer, and then through a dual-tower model consisting of two DNNs with shared parameters to obtain user embeddings and item embeddings. The similarity metric function is used to measure the similarity θ between user and item embeddings. The formula is: in Embedding for user characteristics Embedding for item characteristics.
2. The item recommendation method according to claim 1, characterized in that, Determining a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information includes: The user item interaction information is matched with the user personal attribute information, so that each user item interaction information carries the user personal attribute information. The user item interaction information is matched with the item's tag attribute information, so that each user item interaction information carries the item's tag attribute information; Based on user-item interaction information carrying user personal attribute information and user-item interaction information carrying item tag attribute information, obtain the item tag information of each user with the highest interaction frequency and interest, and each user's personal attribute information, establish the association relationship between users and tags, and generate the user profile.
3. The item recommendation method according to claim 1, characterized in that, Based on the user profile, users are divided into at least one user group, including: User feature information is obtained using the user profile, and the user feature information is then feature-encoded. Users are divided into K user groups using the k-means clustering algorithm and encoded user feature information, where K is a positive integer and is set according to the size of the user group in the dataset.
4. The item recommendation method according to claim 1, characterized in that, After using the trained dual-tower model to obtain items similar to the user in the embedding space as recommended items, the method further includes: Users are divided into at least one user group using the aforementioned personal attribute information; Identify M active users within the same user group who are similar to the new user, where M is an integer greater than 1; The dual-tower model is used to predict recommended items that the M active users are interested in; The M recommended items are weighted and summed to obtain the items recommended to new users.
5. The item recommendation method according to claim 1, characterized in that, After using the trained dual-tower model to obtain items similar to the user in the embedding space as recommended items, the method further includes: The user characteristics and interactive item characteristics of low-activity users are input into the dual-tower model to obtain the prediction results; Users are divided into at least one user group using the aforementioned personal attribute information; Identify M active users within the same user group who are similar to the new user, where M is an integer greater than 1; The dual-tower model is used to predict recommended items that the M active users are interested in; The prediction result is weighted and added to the M recommended items to obtain the items recommended to inactive users.
6. An item recommendation device, characterized in that, include: The acquisition module is used to acquire user personal attribute information, item tag attribute information, and user item interaction information; The first processing module is used to determine a user profile based on the user's personal attribute information, the item's tag attribute information, and the user's item interaction information. The user profile includes the association between the user and the tag. The second processing module is used to divide users into at least one user group according to the user profile. Each user group includes multiple types of users, and the interaction frequency of different types of users with items is in different ranges. The third processing module is used to construct a dual-tower model based on SENet, and to construct the training set and test set of the dual-tower model using the interaction information of the multiple types of users and items. A training module is used to train the dual-tower model using the training set and the test set; The prediction module is used to use the trained dual-tower model to obtain items that are similar to the user in the embedding space as recommended items; Each of these user groups is divided into the following four categories of users: Active users are defined as those who interact with items more than a first preset threshold number of times within a preset time period. Among active users, the number of times they interact with items within a preset time period is less than or equal to the first preset threshold and greater than the second preset threshold; Low-activity users are defined as those whose number of interactions with items within a preset time period is less than or equal to the second preset threshold, which is greater than 0. New users will have 0 interactions with items within a preset time period; Specifically, the third processing module is used to form a training set by combining the user characteristics and item characteristics of 90% of the active users and moderately active users in each user group; to use the user characteristics and item characteristics of 10% of the active users and moderately active users in each user group as a validation set; and to use the user characteristics and item characteristics of the inactive users in each user group as a test set. Specifically, the prediction module processes the training samples in the training set through an SENet layer, and then through a dual-tower model composed of two DNNs with shared parameters to obtain user embeddings and item embeddings. A similarity metric function, θ, is used to measure the similarity between user and item embeddings, as shown in the formula: in Embedding for user characteristics Embedding for item characteristics.
7. An item recommendation device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the item recommendation method as described in any one of claims 1-5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the item recommendation method as described in any one of claims 1-5.