Information recommendation method and device, storage medium and computer device

Abstract

Embodiments of the present application disclose an information recommendation method and device, a storage medium and a computer device. The method comprises: extracting long-term interest features of a user according to user portrait information; extracting short-term interest features of the user according to click sequence information in user historical behavior data; calculating an item attribute preference vector of the user according to item attribute information in the user historical behavior data; performing fusion processing on the long-term interest features, the short-term interest features and item attribute preference information of the user to obtain a user vector; optimizing a plurality of targets in a target model according to the user vector to obtain an optimized target model; processing item information in an item information library through the optimized target model to generate recall information corresponding to a target user, and performing recommendation based on the recall information, so that the model can better reflect the interest preferences of the user, thereby improving the recommendation accuracy of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an information recommendation method and device, a storage medium and a computer device. BACKGROUND

[0002] A recommendation system is one of the important applications in the field of artificial intelligence, which can help users discover item information that may be of interest to the users in an information overload environment, and push the item information to users interested in the item.

[0003] In a complex recommendation system, there are usually many objectives to be optimized, resulting in a large deviation between the recommended content of the recommendation system and the user's interest preferences. SUMMARY

[0004] The embodiments of the present application provide an information recommendation method and device, a storage medium and a computer device, which can aggregate user portrait information, click sequence information and item attribute information into a user vector representing user interest preferences, and simultaneously optimize multiple objectives in a target model based on the user vector, so that the model can better reflect the user's interest preferences, thereby improving the recommendation accuracy of the model.

[0005] In a first aspect, an information recommendation method is provided, the method comprising:

[0006] extracting long-term interest features of a user according to user portrait information;

[0007] extracting short-term interest features of the user according to click sequence information in user historical behavior data;

[0008] calculating an item attribute preference vector of the user according to item attribute information in the user historical behavior data;

[0009] fusing the long-term interest features, the short-term interest features and the item attribute preference information of the user to obtain a user vector;

[0010] optimizing multiple objectives in a target model according to the user vector to obtain an optimized target model;

[0011] processing item information in an item information library through the optimized target model to generate recall information corresponding to a target user, and recommending based on the recall information.

[0012] In a second aspect, an information recommendation device is provided, the device comprising:

[0013] a first extraction unit configured to extract long-term interest features of a user according to user portrait information;

[0014] The second extraction unit is used to extract the user's short-term interest features based on the click sequence information in the user's historical behavior data;

[0015] The calculation unit is used to calculate the user's item attribute preference vector based on the item attribute information in the user's historical behavior data;

[0016] The fusion unit is used to fuse the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector.

[0017] An optimization unit is used to optimize multiple targets in the target model based on the user vector to obtain an optimized target model.

[0018] The recommendation unit is used to process the project information in the project information database through the optimized target model to generate recall information corresponding to the target user, and to make recommendations based on the recall information.

[0019] Thirdly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program adapted for loading by a processor to perform the steps in the information recommendation method as described in any of the above embodiments.

[0020] Fourthly, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the processor executing steps in the information recommendation method as described in any of the above embodiments by calling the computer program stored in the memory.

[0021] This application provides an information recommendation method, apparatus, storage medium, and computer device. The method extracts long-term interest features of users based on user profile information; extracts short-term interest features based on click sequence information in user historical behavior data; calculates a user's item attribute preference vector based on item attribute information in user historical behavior data; fuses the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector; optimizes multiple objectives in a target model based on the user vector to obtain an optimized target model; processes item information in an item information database using the optimized target model to generate recall information corresponding to the target user, and makes recommendations based on the recall information. This application aggregates user profile information, click sequence information, and item attribute information into a user vector representing user interests and preferences, and simultaneously optimizes multiple objectives in the target model based on the user vector, making the model better reflect user interests and preferences, thereby improving the model's recommendation accuracy. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram illustrating an application scenario of the recommendation system provided in an embodiment of this application.

[0024] Figure 2 This is a flowchart illustrating the information recommendation method provided in the embodiments of this application.

[0025] Figure 3 A first frame schematic diagram of the target model provided in the embodiments of this application.

[0026] Figure 4 A second frame schematic diagram of the target model provided in the embodiments of this application.

[0027] Figure 5 This is a schematic diagram of the information recommendation device provided in the embodiments of this application.

[0028] Figure 6 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] This application provides an information recommendation method, apparatus, computer device, and storage medium. Specifically, the information recommendation method of this application can be executed by a computer device, which can be a terminal or a server. The terminal can be a smartphone, tablet, laptop, smart TV, smart speaker, wearable smart device, personal computer (PC), etc. The terminal can also include a client, which can be a video client, browser client, or instant messaging client, etc. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0031] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0032] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0033] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0034] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence, enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, this field involves natural language—the language people use in daily life—and thus has a close relationship with linguistics. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0035] Item: This could be an article, a video project, or an audio project, etc.

[0036] Collaborative filtering (CF) uses the preferences of a group of like-minded people with shared experiences to recommend information that users may be interested in. Individuals respond to information to a certain extent (such as rating) through a collaborative mechanism and record it to achieve the purpose of filtering and help others filter information. The responses are not limited to those that are particularly interested; recording information that is particularly uninteresting is also quite important.

[0037] Multi-task learning: By training on several tasks simultaneously, multiple tasks influence each other, and multiple tasks share a single structure. The parameters within this structure are affected by all tasks during optimization. Thus, when all tasks converge, the structure is equivalent to a fusion of all tasks, and therefore, multi-task learning generally has better generalization ability than single-task learning.

[0038] ItemCF: Calculates item similarity by mining co-occurrence information of items, and then uses item similarity for recommendation and filtering.

[0039] Item2Vec: Assigns a dense vector to each item, which, in contrast to one-hot data representation, preserves the semantic dimension information between items.

[0040] Recurrent Neural Network (RNN): A multilayered feedback neural network, it is an artificial neural network in which nodes are connected in a directed loop. The internal state of an RNN network can exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can utilize their internal memory to process input sequences of arbitrary temporal order, making them more suitable for tasks such as handwriting recognition without segmentation and speech recognition.

[0041] UserCF: Mines user similarity information and recommends items liked by similar users.

[0042] LSTM stands for Long Short-Term Memory, a type of temporal recurrent neural network suitable for processing and predicting important events with relatively long intervals and delays in time series.

[0043] Gated Recurrent Unit (GRU): By introducing a reset gate and an update gate, the propagation of historical information can be controlled.

[0044] Expert Network: A network structure designed for different tasks in multi-task learning. It is usually composed of multiple fully connected networks and is used to extract features for different tasks.

[0045] Gate: Used to control the probability distribution of input weights for different expert networks, and to combine different expert networks with different weights, which are then input into different objective optimization towers.

[0046] Cosine similarity, also known as cosine similarity, assesses the similarity between two vectors by calculating the cosine of the angle between them. Cosine similarity plots the vectors in a vector space, such as the most common two-dimensional space, based on their coordinate values.

[0047] Target users: Users currently using the recommendation system.

[0048] User personas, also known as user roles, are an effective tool for characterizing users and connecting their needs with design direction. They are widely used across various fields, and in practice, user personas often use simple, relatable language to connect user attributes, behaviors, and expectations, serving as virtual representatives of actual users.

[0049] Recall: Retrieving relevant documents from a document library, such as making a preliminary selection of products to recommend to a user.

[0050] In a complex recommender system, there are often many objectives to optimize, such as click-through rate (CTR), share rate, and playback completion rate. To optimize CTR, Collaborative Filtering (CF) uses ItemCF and UserCF to calculate user similarity and item similarity based on co-occurrence between users and items, then provides personalized recommendations. Co-occurrence generally refers to the phenomenon where information described by the same or different types of features appears together in a document. Features include external and internal features of the document, such as title, author, keywords, and institution. In recommender systems, there are interactive behaviors between users and items, and each interaction is recorded as a co-occurrence. Interaction behaviors can be categorized into various types, such as clicks, shares, and playback. Another method for calculating item similarity is Item2vec, which trains semantic relevance between items for personalized recommendations. Other methods use sequence modeling (RNN, GRU, etc.) to predict user clicks at the next time step, improving CTR. These methods can be used to improve user share rate and playback completion rate.

[0051] For example, in the recommendation system of application A, the system focuses on different objectives. It typically trains multiple sub-networks to optimize for different objectives, allowing these training tasks to be trained independently. After training all sub-networks, different prediction values ​​can be generated. These different prediction values ​​are then weighted and combined to calculate the user's interest score for each item. Finally, the top K (Top K) candidate items are obtained by ranking them. For instance, in the recommendation system of application A, the click-through rate and share rate are not on the same order of magnitude; users usually have a higher click-through rate than a share rate. Therefore, in a certain objective recommendation scenario, there is a data sparsity problem: the number of samples used to train the click-through rate is far greater than the number of samples used to train the share rate. This data sparsity problem causes the user's final interest score to be biased towards the click-through rate model, when in reality, the user's share rate is a better reflection of their interests.

[0052] Independent training for optimizing multiple objectives is not realistic. These objectives are often correlated; for example, click-through rate and share rate are usually positively correlated, as are video completion rate. Training these objectives separately severs the correlation between these metrics, which is inconsistent with real-world scenarios. Furthermore, separate training introduces another problem: sample selection bias. In a target recommendation scenario, user interest scores are typically calculated for all candidate items. However, model training is performed on a small subset of the training set, and the samples used to train share rate are even smaller than those used to train click rate. Model prediction, on the other hand, occurs across the entire sample space, leading to sample selection bias. In other words, the distributions of the training and test sets are different, violating the machine learning assumption that samples should be independent and identically distributed.

[0053] Therefore, this application provides an information recommendation method. It designs a basic target model, aggregates user profile information, click sequence information, and item attribute information into a user vector, optimizes multiple targets in the target model, and recommends information to target users based on the optimized target model. Specifically, user profile information is modeled as the user's long-term interest features; click sequence information is modeled as the user's short-term interest features; and item attribute information is used to model the user's item attribute preferences. Compared to training multiple targets separately, this application optimizes multiple targets simultaneously, employing a multi-gate Mixture-of-Experts (MMoE) module to combine the influence of different expert networks on different target towers. For example, different target towers may include sharing targets and playback targets.

[0054] The MMoE module extracts features from different directions using multiple expert networks, which are generally composed of DNN structures. Embedding structures are shared at the bottom of each expert network, allowing each network to have different impacts on different tasks. Above the expert networks are towers with different optimization objectives, used for multi-task objective learning and optimization.

[0055] The MMoE module has n different expert networks and m optimization objectives, thus having m different gates. For example, if the optimization objectives are playback and sharing objectives (where playback indicates whether a user plays the project and sharing indicates whether a user shares the project), then m = 2, the same as the number of optimization objectives. Alternatively, n = 4, where n represents hyperparameters, typically determined by offline performance and online latency constraints. Or, if the optimization objectives are playback, sharing, and retention objectives (where playback indicates whether a user plays the project, sharing indicates whether sharing brings in new users), then m = 3, the same as the number of optimization objectives.

[0056] The score of the kth Gate is calculated as follows (1):

[0057] g k(x) =softmax(W gk(x) (1);

[0058] Among them, g k(x) W is the output of the k-th gate, which follows a softmax probability distribution. gk(x) The intermediate parameters represent the model structure; k(x) represents the k-th gate, and k corresponds to the number of optimization objectives; the MMoE module is a layer of the overall target model interface, and x refers to the input provided to the MMoE by the lower layer of the target model, such as x being a user vector.

[0059] Here, softmax is the normalization exponential function, also known as the softmax function, which is a generalization of the logistic function. The softmax function can compress a K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ(z), such that each element is in the range (0,1) and the sum of all elements is 1. The softmax function is often used in multi-class classification problems.

[0060] The input to the tower corresponding to the k-th optimization objective can be expressed as the following formula (2):

[0061]

[0062] Where k represents the k-th task; n represents the n expert networks; g k (x) i f represents the voting weights of different expert networks for an optimization objective; i (x) represents the suggestions of different expert networks for an optimization objective; the weighted sum of the above two is the final suggestion of multiple expert networks for an optimization objective.

[0063] The weight selection in the MMoE module is different, and a Gate model is provided for each task. For different tasks, the output of a specific k-th Gate represents the probability of different expert networks being selected. The weighted sum of multiple expert networks is used to obtain f. k (x) is then output to a specific Tower model for the final output.

[0064] The output of the kth optimization objective is given by the following formula (3):

[0065] y k = k (f k (x)) (3);

[0066] Among them, h k This is the lost structure of the k-th optimization target tower. Under the control of multiple expert networks and multiple gates, different targets can be optimized simultaneously. Furthermore, through the sharing of the bottom structure, the embodiments of this application have significantly fewer parameters compared to networks trained separately. The bottom structure is the layer below the MMoE module layer, serving as the input to the overall MMoE.

[0067] This application provides an information recommendation method, which can be executed by a terminal or a server, or by both a terminal and a server. This application uses the example of an information recommendation method executed by a server to illustrate the method.

[0068] An information recommendation method includes: extracting long-term interest features of users based on user profile information; extracting short-term interest features of users based on click sequence information in user historical behavior data; calculating a user's item attribute preference vector based on item attribute information in the user's historical behavior data; fusing the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector; optimizing multiple targets in a target model based on the user vector to obtain an optimized target model; processing item information in an item information database using the optimized target model to generate recall information corresponding to the target user, and making recommendations based on the recall information.

[0069] See Figure 1 , Figure 1 This is a schematic diagram illustrating an application scenario of the recommendation system provided in this application embodiment. The recommendation system includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal, and the mobile terminal can be at least one of a mobile phone, tablet computer, or laptop computer. The server 120 can be implemented using a standalone server or a server cluster composed of multiple servers.

[0070] Server 120 can acquire user profile information, click sequence information, and item attribute information. Based on the user profile information, it extracts the user's long-term interest features; based on the click sequence information in the user's historical behavior data, it extracts the user's short-term interest features; and based on the item attribute information in the user's historical behavior data, it calculates the user's item attribute preference vector. Then, it fuses the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector. Next, it optimizes multiple objectives in the target model based on the user vector to obtain an optimized target model. Finally, it processes item information in the item information database using the optimized target model to generate recall information corresponding to the target user, and makes recommendations based on the recall information. Server 120 sends the recall information to the terminal 110 corresponding to the target user.

[0071] It should be noted that the above application scenario is only an example. In some embodiments, the steps of the above information recommendation method can also be executed by the terminal 110. For example, the terminal 110 can directly use the configured information recommendation device to run a pre-trained target model, obtain user profile information, click sequence information, and item attribute information through the target model, extract the user's long-term interest features based on the user profile information, extract the user's short-term interest features based on the click sequence information in the user's historical behavior data, and calculate the user's item attribute preference vector based on the item attribute information in the user's historical behavior data. Then, the user's long-term interest features, short-term interest features, and item attribute preference information are fused to obtain a user vector. Then, multiple targets in the target model are optimized based on the user vector to obtain an optimized target model. The optimized target model is then used to process the item information in the item information database to generate recall information corresponding to the target user, and recommendations are made based on the recall information.

[0072] In some embodiments, an application (client) that supports content push function is installed and running on the terminal 110. When the terminal 110 runs the application, a user interface for displaying push information is displayed on the screen of the terminal 110. The terminal can select the recall information corresponding to the target user from the project information in the project information database by performing the steps of the above information recommendation method, and push the recall information corresponding to the target user to the application logged in by the target user, and display it to the currently logged-in target user through the user interface. The target user can browse and view the pushed information through the user interface.

[0073] For example, this recommendation system can be applied to application scenarios such as text information recommendation applications, video recommendation applications, and image recommendation applications.

[0074] For example, this recommendation system can be applied to text recommendation applications, that is, to determine the recommended text from the recalled text information. For example, after a user opens the recommendation APP, the terminal 110 automatically generates a request for text information recommendation for the target user, and processes it according to the user profile information, click sequence information and item attribute information corresponding to the target user to obtain the user vector of the target user. Then, it determines the recalled text information that is similar to the user vector of the target user from the item information database, and then determines the recommended text information that matches the user's interests and preferences from the recalled text information, and returns it to the recommendation APP so that the user can get accurate recommended text information.

[0075] For example, this recommendation system can be applied to video recommendation applications. It determines recommended video information from recalled video information. For instance, after a user opens the recommendation app, the terminal 200 automatically generates a video recommendation request for the target user. It processes the user profile information, click sequence information, and item attribute information corresponding to the target user to obtain the user vector of the target user. It then determines recalled video information similar to the user vector of the target user from the item information database. Finally, it determines recommended video information that matches the user's interests and preferences from the recalled video information and returns it to the recommendation app so that the user can receive accurate recommended video information.

[0076] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the priority of the embodiments.

[0077] Please see Figures 2 to 4 , Figure 2 This is a flowchart illustrating the information recommendation method provided in the embodiments of this application. Figures 3 to 4 These are schematic diagrams illustrating the framework of the target model provided in the embodiments of this application. The specific process of this information recommendation method can be as follows:

[0078] Step 101: Extract the user's long-term interest features based on the user profile information.

[0079] like Figure 3 As shown, the constructed target model is a multi-objective optimization recall model based on MMoE. Specifically, an MMoE module is introduced based on the aggregation of user multi-interest vectors, and training is performed simultaneously by optimizing the user's playback objective (whether to play the item) and sharing objective (whether to share the item). Optimization of the multi-objective task is performed based on user vectors.

[0080] For example, in the recommendation system of application A, predicting the user's playback goal and sharing goal has always been a very important optimization objective, so playback goal and sharing are selected for optimization.

[0081] First, user historical behavior data input and shared data embedding are performed. User historical behavior data input includes user profile inputs, user behavior history inputs (item seq inputs), and user behavior history item attribute inputs (item attr inputs). Shared data embedding includes item attribute embeddings (attr embeddings), item embeddings (item embeddings), user behavior embeddings (action embeddings), and context embeddings (context embeddings). The input feature for both user historical behavior data input and shared data embedding is ID information.

[0082] Then, the user's historical behavior data input and the corresponding ID information of the shared data embedding are transformed into dense vectors through a Various Embedded Lookup Layer. For example, an array data structure is used to model the vectors; this structure typically stores ordinary vectors, also known as dense vectors.

[0083] In some embodiments, the user profile information is converted into a corresponding dense vector, and the user's long-term interest features are extracted based on the dense vector corresponding to the user profile information.

[0084] Specifically, for users' long-term interest characteristics, a sub-module of user vectors is modeled through the UserProf module to represent these characteristics. The data, after being processed by the embedding lookup layer, is input into the UserProf module for Prof Embedding. This is then processed by the Transformer Layer to obtain the ProfUnit, which represents the user's long-term interest characteristics.

[0085] In some embodiments, extracting long-term interest features of a user based on the user profile information includes: obtaining user profile information, which includes user demographic attribute information and user preference information obtained statistically based on the user's historical behavior data; and extracting long-term interest features of the user based on the user demographic attribute information and the user preference information.

[0086] The user profile information includes user demographic attributes and user preference information obtained from statistical analysis of user historical behavior data. For example, user demographic attributes include gender, age, and demographic group, while user preference information includes processed features obtained from the processing and statistical analysis of user historical behavior data. These processed features include user main interests and tags that the user is interested in.

[0087] Step 102: Extract the user's short-term interest features based on the click sequence information in the user's historical behavior data.

[0088] For example, taking video projects as an example, click sequence information can include the video IDs that the user clicked and the corresponding video's attribute characteristics.

[0089] For short-term user interest features, the weighted-transformer module within the user behavior sequence (ItemSeq) module processes the user's click sequences, enabling parallel and optimized computation of the latent information within the click sequences. The weighted-transformer module improves upon the Transformer by employing a multi-head attention mechanism. It divides the model into multiple heads, forming multiple subspaces, and assigns different weights to different heads, allowing the model to focus on different aspects of information. The Transformer module is a self-attention-based translation module, using the attention mechanism as the core component of the encoder-decoder to perform translation operations. For example, the weighted-transformer module can utilize a unified online translation service.

[0090] Specifically, the data processed by the embedding lookup layer is input into the user behavior sequence module. The Sequence Embed Pooler in the user behavior sequence module is used to synthesize various embedded dense vectors. The Sequence Embed Pooler is a pooler layer in the transformer structure. Then, the synthesized dense vectors are processed sequentially through the first weighted translation self-attention layer (Weighted-transformer Layer #1) and the second weighted translation self-attention layer (Weighted-transformer Layer #2) to obtain sequence units. These sequence units are used to represent the user's short-term interest features.

[0091] In some embodiments, the click sequence information is converted into a corresponding dense vector, and the user's short-term interest features are extracted based on the dense vector corresponding to the click sequence information.

[0092] Step 103: Calculate the user's item attribute preference vector based on the item attribute information in the user's historical behavior data.

[0093] The project attribute information includes project tags, project category, project title word segmentation, and project theme. Taking a video project as an example, the project attribute information includes video tags, video category, video title word segmentation, and video theme.

[0094] In some embodiments, calculating the user's item attribute preference vector based on the item attribute information in the user's historical behavior data includes: fusing the user's long-term interest features and short-term interest features to obtain a first fusion vector; and processing the first fusion vector and the item attribute information to calculate the user's item attribute preference vector.

[0095] Specifically, for the user's item attribute preference vector, an item attribute preference module was designed, for example, named the Attr2Item module. This module fuses the user's long-term and short-term interest features to obtain a first fused vector (the user's long-term and short-term interest vector). This first fused vector is then input into the Attr2Item module, which uses a multi-head attention mechanism to process the first fused vector and item attribute information to calculate the user's item attribute preference vector. The Attr2Item module is used to model the relationship between the attributes of the user's historical interaction items and the recommendation structure. For example, which tags the user mainly focuses on are used in recommendations. The Attr2Item structure integrates how to obtain which tags the user focuses on and how to establish a relationship with the recommendation structure. Specifically, the long-term interest features of the user output from the user profiling module and the short-term interest features of the user output from the user click sequence module are fused (aggregated) to obtain a first fused vector (the user's long-term and short-term interest vector). This first fused vector is then input into the topN attention layer in the Attr2Item module, along with the item attribute information from the shared data embedding (item attribute embedding). The topN attention layer processes the first fused vector and the item attribute information to calculate the user's item attribute preference vector for the item. The topN attention layer focuses attention on the top N data points.

[0096] In some embodiments, the project attribute information is converted into a corresponding dense vector, and the user's project attribute preference vector is calculated based on the dense vector corresponding to the project attribute information.

[0097] Step 104: The user's long-term interest features, short-term interest features, and project attribute preference information are fused to obtain a user vector.

[0098] For example, by adding vectors of a user's long-term interest features, short-term interest features, and item attribute preference information, we can aggregate them into a final user representation vector, which is the user vector.

[0099] Step 105: Optimize multiple targets in the target model based on the user vector to obtain the optimized target model.

[0100] In some embodiments, the plurality of objectives includes a playback objective and a sharing objective, and optimizing the plurality of objectives in the objective model based on the user vector to obtain an optimized objective model includes:

[0101] The playback target and the sharing target in the target model are optimized based on the user vector to obtain the optimized target model.

[0102] In some embodiments, the target model includes a multi-task learning module, and the optimization of the playback target and the sharing target in the target model based on the user vector to obtain an optimized target model includes:

[0103] The user vector is processed by the multi-task learning module to split the user vector into common information and difference information. The common information is the common information corresponding to the playback target and the sharing target, and the difference information is the difference information corresponding to the playback target and the sharing target.

[0104] The playback target and the sharing target in the target model are optimized based on the shared information and the difference information to obtain an optimized target model.

[0105] For example, the SharedMMoE module optimizes multi-objective tasks based on user vectors. Specifically, the shared MMoE layer within the SharedMMoE module processes user vectors to optimize the sharing unit and playback unit. Then, different vectors are generated for each objective, and these generated vectors are combined (Concat). Figure 3 The structure for extracting and fusing long-term interest features, short-term interest features, and attribute preference vectors into user vectors is unified. Then, the SharedMMoE module splits the obtained user vectors into data required for two optimization objectives. During data splitting, multiple expert networks work together. The SharedMMoE module uses the user vectors that are uniformly input from the lower layer.

[0106] Specifically, based on online serving, the cosine similarity between the user vector of a sample user and the item vectors involved in the user's user behavior history is calculated to obtain the candidate item recall for sample user U, i.e., to calculate the top K items (top K) that the sample user is most interested in. Additionally, based on offline serving, the Play Target Embed, Share Target Embed, and the results processed based on the share and play targets in the SharedMMoE module are input into the overall loss function (sampled loss) of the target model for training. This trains the model to find the best match between the top K items that the sample user is most interested in and the historical recommended items, thus simultaneously optimizing multiple targets in the target model to obtain the optimized target model.

[0107] like Figure 3 As shown, after aggregating user vectors, the MMoE module optimizes the user's playback and sharing goals for the project. This embodiment further improves the MMoE module. Each different expert network in the MMoE module can extract different information, and different Gates control the input of each target Tower. In the target recommendation scenario, since the user's click-through rate and share rate are highly correlated, this embodiment believes that different expert networks may contain shared interests. Therefore, a GateShared structure is designed in the MMoE module to extract this shared information, supplementing the information that different Gates can only extract from different expert networks. The basic MMoE module is designed for multiple expert networks with the same input. This embodiment considers the differences between different goals and introduces a common structure (such as the GateShared structure) to model the shared information (information of the same or similar parts) in multiple goals. By splitting the original single input information into shared input information and target-specific differential input information, it explicitly differentiates the input.

[0108] This application embodiment selects the sequence within each user's playback history (session) from each user's historical behavior data as the input data for the ItemSeq structure. Because the time differences within a user's playback history are relatively small, the recommendation results obtained using this training data can receive user feedback within a shorter time interval. Conversely, if the time span of the training data is too large, the recommendation results are prone to low real-time performance and are less likely to capture the user's short-term interests. For example, if the historical length of the training data can be increased, it can take into account both the user's long-term interests and short-term data; for instance, to balance performance and online mode constraints, the historical length can be set to 64. The ItemSeq structure is used to obtain the user behavior sequence corresponding to the Item in the user's playback history. The storage module in the recommendation system stores each user's historical behavior (such as the user's playback history).

[0109] The training data includes user profile information, user historical behavior information, user historical behavior attribute information, target item information, and user behavior information related to the target item (such as sharing and playing).

[0110] After encoding the training data using a Transformer-based multi-objective optimization network, user vectors E need to be used. u To predict the results to be recommended. Assume the project candidate pool is [X1, X2, X3, ..., X...]. N ] Calculate the embedding vector of each candidate and the user vector E. u Cosine similarity: Given two vectors, the cosine similarity between them is calculated as follows (4), and then the vectors with the highest cosine similarity are recommended:

[0111]

[0112] Where u represents the user vector; v represents the item vector, i.e., the embedding vector of each candidate item.

[0113] Specifically, for user U, user profile information, user click sequence information, and project attribute information are fused to obtain user vector E. u Then calculate the user vector E uFor each candidate, a Cosine similarity score is calculated. Two values ​​need to be determined: one for the sharing goal (whether the item is shared) and the other for the playback goal (whether the item is played). In a target recommendation scenario, the estimated values ​​for both the sharing and clicking goals are assigned weights, and then weighted to obtain the final score. For example, different goals correspond to different Top N items and scores. The final score for each item is obtained by combining multiple goals, providing a recommendation candidate set. Assuming the candidate pool has only 10 items, the result after calculation is the similarity score with the user vector E. u The most similar candidates are X3, X5, and X7. Therefore, the most similar candidates (X3, X5, and X7) are used as recall candidates so that the recommendation system can complete the candidate recall for user U.

[0114] In some embodiments, the plurality of objectives further includes a retention objective, and the method further includes:

[0115] Obtain project feedback information that includes retention rates;

[0116] Based on the user vector and the project feedback information, the playback target, the sharing target, and the retention target in the target model are optimized to obtain the optimized target model.

[0117] In some embodiments, optimizing the playback target, sharing target, and retention target in the target model based on the user vector and the project feedback information to obtain an optimized target model includes: optimizing the playback target in the target model based on the user vector; optimizing the sharing target and retention target in the target model based on the user vector and the project feedback information; and obtaining the optimized target model based on the optimized playback target, sharing target, and retention target.

[0118] The retention rate in project feedback information affects the sharing goal, but not the playback goal. For example, when modeling the user sharing goal, considering that most videos shared by users in a target recommendation scenario are from new users, it is necessary to model the retention rate of new users after receiving the shared video. Therefore, the retention rate of users after receiving videos shared by others is used as an additional modeling goal, which can be called the retention rate goal, and this retention rate goal is a weight among multiple goals. Related to the target recommendation scenario, which emphasizes the number of daily active users (DAU), sharing an item will retain the item and bring in more users, thus increasing DAU. The additional goal is the number of users brought in after each share. The entire recommendation system tends to prioritize recommending candidates with a larger number of users brought in under the same conditions. Retention rate can represent the number of users brought in after sharing.

[0119] Therefore, this retention rate is added as a new target and optimized together with the previous sharing and playback targets: on the one hand, all targets share the underlying input; on the other hand, the retention rate (the number of users brought in after sharing) helps the recommendation system make biased choices to select candidates that can improve DAU.

[0120] From a real-world perspective, if this multi-objective optimization model doesn't model new user retention and only optimizes the sharing and playback objectives, then new users who share videos with others might not revisit the site within a few days. This means that optimizing the sharing objective doesn't bring any positive benefits to the recommendation system; instead, it loses some potential long-term users. Therefore, while optimizing playback and sharing objectives, it's necessary to incorporate retention objectives for optimization as well.

[0121] like Figure 4 As shown, in Figure 3 Building upon this foundation, item feedback information containing retention data (users brought in after sharing) is introduced. Then, user vectors are processed through the MMoE layer (Wgted Shared MMoELayer) in the WgtSharedMMoE module to optimize the sharing unit, play unit, and retention unit. Furthermore, based on online serving, the cosine similarity between the sample user's user vector and the item vectors involved in the sample user's user behavior history is calculated to obtain the candidate item recall for sample user U, i.e., calculating the top K items (topK) that the sample user is most interested in. In addition, based on offline serving, the Play Target Embed, Share Target Embed, and the results processed based on the share and play targets in the WgtSharedMMoE module are input into the overall loss function (Sampled Loss) of the target model for training. Additionally, the item feedback information and the results processed based on the share and play targets in the WgtSharedMMoE module are input into the retention target loss function (Wgt Loss) for training. This process trains the model to find the K most interesting items for the sample users that best match the historical recommended items, thus simultaneously optimizing multiple targets (play, share, and retention targets) in the target model to obtain the optimized target model.

[0122] Step 106: Process the project information in the project information database using the optimized target model to generate recall information corresponding to the target user, and make recommendations based on the recall information.

[0123] In some embodiments, the step of processing the project information in the project information database using the optimized target model to generate recall information corresponding to the target user, and making recommendations based on the recall information, includes: obtaining the user vector of the target user using the optimized target model; calculating the cosine similarity between the user vector of the target user and the project vectors in the project information database; generating recall information corresponding to the target user based on the cosine similarity; determining recommendation information from the recall information; and recommending the recommendation information to the target user.

[0124] For example, the optimized target model processes the user profile information, click sequence information, and item attribute information corresponding to the target user to obtain the target user's user vector. Then, the cosine similarity between the target user's user vector and the item vectors in the item information database is calculated. Based on this cosine similarity, recall information corresponding to the target user is generated. Recommendation information is then determined from the recall information and recommended to the target user. Here, the target user is either a user currently using the recommendation system or a currently logged-in user of an application that uses the recommendation system.

[0125] For example, when you need to use the user vectors of the target users to predict the recommended results, such as when the candidate pool in the project information base is [Y1, Y2, Y3, ..., Y...] N ], Calculate the cosine similarity between the item vector and the user vector of each candidate item in the item information database. Given two vectors, the cosine similarity between them can be calculated using formula (4), which will not be elaborated here. Then, recommend the vectors with the highest cosine similarity. For the target user, the user profile information, click sequence information, and item attribute information of the target user are fused to obtain the user vector of the target user. Then, calculate the cosine similarity score between the user vector and the item vector of each candidate item. After calculation, the candidates most similar to the target user vector are Y1, Y2, and Y5. The most similar candidates (Y1, Y2, and Y5) are used as recall candidates (recall information of the target user) so that the recommendation system can complete the recall of candidate items for the target user U.

[0126] In some embodiments, generating recall information corresponding to the target user based on the cosine similarity includes: determining a first candidate item set from the item information in the item information database based on the cosine similarity; performing a preliminary selection on the first candidate item set to obtain a second candidate item set, wherein the number of items in the second candidate item set is less than the number of items in the first candidate item set; sorting the candidate items in the second candidate item set, and selecting the top K candidate items from the sorted second candidate item set as the recall information corresponding to the target user.

[0127] For example, a recommendation system for a specific application scenario may include a recall logic unit, a preliminary selection logic unit, and a ranking logic unit. The recall logic unit retrieves data (documents) based on the user's profile information, according to various dimensions such as precise personalization, general personalization, and popularity. The preliminary selection logic unit performs initial screening on a large number of recalled results according to specific preliminary selection rules (e.g., specific preliminary selection rules based on parameters such as user document relevance, timeliness, geographic location, and diversity) to reduce the computational scale of the ranking logic unit. For instance, the recall logic unit might retrieve tens of thousands of items, which are then filtered by the preliminary selection logic unit to reduce the number of items from tens of thousands to thousands. The ranking logic unit ranks the final results output by the target model and recommends them to the user.

[0128] For example, based on the cosine similarity between the target user's user vector and the item vectors in the item information database, a first candidate item set is determined from the item information in the database. For instance, the recall logic unit, based on the MMoE multi-task learning model, can calculate multiple predictions, and then use these predictions to find the Top K candidate items that the user is most interested in, selecting these Top K candidate items as the first candidate item set. In the recall process, the candidate logic involves storing each item vector offline and generating a user vector online. Because the user's online context is dynamic, the recommendation system must ensure that user vectors are generated online to capture the user's interests and preferences at the current moment.

[0129] For example, a first candidate item set is initially selected based on a specific initial selection rule to obtain a second candidate item set, where the number of items in the second candidate item set is less than the number of items in the first candidate item set. This specific initial selection rule might be set based on parameters such as user document relevance, timeliness, geographic location, and diversity. For instance, in the initial selection logic unit, the MMoE multi-task learning model can be used not only to assist in candidate item selection but also in the ranking stage. Each user's user vector has already been calculated in the recall stage. To calculate the cosine similarity between the candidate item's embedding vector and the target user's user vector in the initial selection stage, items with higher scores are selected and input into the ranking logic unit for processing.

[0130] For example, the candidates in the second candidate item set are sorted, and the top K candidate items are selected from the sorted second candidate item set as the recall information corresponding to the target user. For example, the final results output by the target model are sorted by the sorting logic unit, and the top K items are recommended to the target user.

[0131] The recall information can be used to recommend all items to the target user, or it can recommend items ranked higher. In other words, all items in the recall information can be selected as recommended items, or items ranked higher than a preset position can be selected as recommended items and recommended to the target user.

[0132] All of the above technical solutions can be combined in any way to form optional embodiments of this application, and will not be described in detail here.

[0133] This application embodiment extracts long-term interest features of users based on user profile information; extracts short-term interest features of users based on click sequence information in user historical behavior data; calculates the user's item attribute preference vector based on item attribute information in user historical behavior data; fuses the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector; optimizes multiple objectives in the target model based on the user vector to obtain an optimized target model; processes item information in the item information database using the optimized target model to generate recall information corresponding to the target user, and makes recommendations based on the recall information. This application embodiment aggregates user profile information, click sequence information, and item attribute information into a user vector representing user interests and preferences, and simultaneously optimizes multiple objectives in the target model based on the user vector, making the model more reflective of user interests and preferences, thereby improving the model's recommendation accuracy.

[0134] To facilitate better implementation of the information recommendation method of this application, this application also provides an information recommendation device. Please refer to...Figure 5 , Figure 5 A schematic diagram of the structure of the information recommendation device provided in this application embodiment. The information recommendation device 500 may include:

[0135] The first extraction unit 501 is used to extract the user's long-term interest features based on the user profile information;

[0136] The second extraction unit 502 is used to extract the user's short-term interest features based on the click sequence information in the user's historical behavior data;

[0137] The calculation unit 503 is used to calculate the user's item attribute preference vector based on the item attribute information in the user's historical behavior data;

[0138] The fusion unit 504 is used to fuse the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector.

[0139] The optimization unit 505 is used to optimize multiple targets in the target model according to the user vector to obtain the optimized target model;

[0140] The recommendation unit 506 is used to process the project information in the project information database through the optimized target model to generate recall information corresponding to the target user, and to make recommendations based on the recall information.

[0141] In some embodiments, the first extraction unit 501 is configured to: acquire user profile information, the user profile information including user demographic attribute information and user preference information obtained statistically based on the user's historical behavior data; and extract the user's long-term interest features based on the user demographic attribute information and the user preference information.

[0142] In some embodiments, the calculation unit 503 is configured to: fuse the user's long-term interest features and short-term interest features to obtain a first fusion vector; and process the first fusion vector and the project attribute information to calculate the user's project attribute preference vector.

[0143] In some embodiments, the plurality of targets includes a playback target and a sharing target, and the optimization unit 505 is configured to optimize the playback target and the sharing target in the target model according to the user vector to obtain an optimized target model.

[0144] In some embodiments, the target model includes a multi-task learning module, and the optimization unit 505 is configured to: process the user vector through the multi-task learning module to split the user vector into common information and difference information, wherein the common information is the common information corresponding to the playback target and the sharing target, and the difference information is the difference information corresponding to the playback target and the sharing target; optimize the playback target and the sharing target in the target model according to the common information and the difference information to obtain an optimized target model.

[0145] In some embodiments, the plurality of objectives further includes a retention objective, and the optimization unit 505 is configured to: obtain project feedback information containing retention data; and optimize the playback objective, the sharing objective, and the retention objective in the target model based on the user vector and the project feedback information to obtain an optimized target model.

[0146] In some embodiments, the optimization unit 505 is configured to: optimize the playback target in the target model based on the user vector; optimize the sharing target and the retention target in the target model based on the user vector and the project feedback information; and obtain an optimized target model based on the optimized playback target, the sharing target, and the retention target.

[0147] In some embodiments, the recommendation unit 506 is configured to: obtain a user vector of a target user through the optimized target model; calculate the cosine similarity between the user vector of the target user and the project vector in the project information database; generate recall information corresponding to the target user based on the cosine similarity; determine recommendation information from the recall information; and recommend the recommendation information to the target user.

[0148] In some embodiments, the recommendation unit 506 is configured to generate recall information corresponding to the target user based on the cosine similarity, specifically including: determining a first candidate item set from the item information in the item information database based on the cosine similarity; performing a preliminary selection on the first candidate item set to obtain a second candidate item set, wherein the number of items in the second candidate item set is less than the number of items in the first candidate item set; sorting the candidate items in the second candidate item set, and selecting the top K ranked candidate items from the sorted second candidate item set as the recall information corresponding to the target user.

[0149] In some embodiments, the first extraction unit 501 is used to convert the user profile information into a corresponding dense vector, and extract the user's long-term interest features based on the dense vector corresponding to the user profile information.

[0150] The second extraction unit 502 is used to convert the click sequence information into a corresponding dense vector, and extract the user's short-term interest features based on the dense vector corresponding to the click sequence information;

[0151] The calculation unit 503 is used to convert the project attribute information into a corresponding dense vector, and calculate the user's project attribute preference vector based on the dense vector corresponding to the project attribute information.

[0152] All of the above technical solutions can be combined in any way to form optional embodiments of this application, and will not be described in detail here.

[0153] The information recommendation device 500 provided in this application embodiment includes: a first extraction unit 501 extracting long-term interest features of users based on user profile information; a second extraction unit 502 extracting short-term interest features of users based on click sequence information in user historical behavior data; a calculation unit 503 calculating a user's item attribute preference vector based on item attribute information in user historical behavior data; a fusion unit 504 fusing the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector; an optimization unit 505 optimizing multiple objectives in a target model based on the user vector to obtain an optimized target model; and a recommendation unit 506 processing item information in an item information database using the optimized target model to generate recall information corresponding to the target user, and making recommendations based on the recall information. This application embodiment aggregates user profile information, click sequence information, and item attribute information into a user vector representing user interests and preferences, and simultaneously optimizes multiple objectives in the target model based on the user vector, making the model better reflect user interests and preferences, thereby improving the model's recommendation accuracy.

[0154] Accordingly, embodiments of this application also provide a computer device, which can be a terminal or a server. For example... Figure 6 As shown, the computer device may include a radio frequency (RF) circuit 601, a memory 602 including one or more computer-readable storage media, an input unit 603, a display unit 604, a sensor 605, an audio circuit 606, a wireless Fidelity (WiFi) module 607, a processor 608 including one or more processing cores, and a power supply 609, among other components. Those skilled in the art will understand that... Figure 6 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0155] The RF circuit 601 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink information from the base station and hands it over to one or more processors 608 for processing; additionally, it transmits uplink data to the base station. Furthermore, the RF circuit 601 can also communicate wirelessly with networks and other devices.

[0156] The memory 602 can be used to store software programs and modules. The processor 608 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function, etc.; the data storage area may store data created according to the use of the computer device, etc.

[0157] The input unit 603 can be used to receive input digital or character information, and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0158] Display unit 604 can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of computer devices. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. Display unit 604 may include a display panel.

[0159] The computer device may also include at least one sensor 605, such as a light sensor, a motion sensor, and other sensors.

[0160] Audio circuitry 606, a speaker, and a microphone provide an audio interface between the user and the computer device. Audio circuitry 606 converts received audio data into electrical signals, transmits them to the speaker, and the speaker converts them into sound signals for output. Conversely, the microphone converts collected sound signals into electrical signals, which are then received by audio circuitry 606, converted back into audio data, and processed by processor 608. The processed data is then transmitted via RF circuitry 601 to, for example, another computer device, or output to memory 602 for further processing. Audio circuitry 606 may also include an earphone jack to facilitate communication between peripheral headphones and the computer device.

[0161] WiFi is a short-range wireless transmission technology. Computer devices using a WiFi module 607 can help users send and receive emails, browse web pages, and access streaming media, providing users with wireless broadband internet access. Although Figure 6 WiFi module 607 is shown, but it is understood that it is not an essential component of computer equipment and can be omitted as needed without changing the nature of the invention.

[0162] The processor 608 is the control center of the computer device. It connects various parts of the mobile phone through various interfaces and lines. By running or executing software programs and / or modules stored in the memory 602, and calling data stored in the memory 602, it performs various functions of the computer device and processes data, thereby monitoring the computer device as a whole.

[0163] The computer device also includes a power supply 609 (such as a battery) that supplies power to various components. Preferably, the power supply can be logically connected to the processor 608 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system.

[0164] Although not shown, the computer device may also include a camera, Bluetooth module, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 608 in the computer device loads the executable files corresponding to the processes of one or more computer programs into the memory 602 according to the following instructions, and the processor 608 runs the computer programs stored in the memory 602 to realize various functions:

[0165] Based on user profile information, long-term interest features of users are extracted; based on click sequence information in user historical behavior data, short-term interest features of users are extracted; based on item attribute information in the user historical behavior data, the user's item attribute preference vector is calculated; the user's long-term interest features, short-term interest features, and item attribute preference information are fused to obtain a user vector; multiple objectives in the target model are optimized based on the user vector to obtain an optimized target model; the optimized target model is used to process item information in the item information database to generate recall information corresponding to the target user, and recommendations are made based on the recall information.

[0166] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0167] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0168] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the information recommendation method of this application embodiment; for brevity, further details are omitted here.

[0169] This application also provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding process in the information recommendation method of this application embodiment. For simplicity, further details are omitted here.

[0170] This application also provides a computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding process in the information recommendation method of this application embodiment. For brevity, further details are omitted here.

[0171] It should be understood that the processor in the embodiments of this application may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0172] It is understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0173] It should be understood that the above-described memory is exemplary and not a limiting description. For example, the memory in the embodiments of this application may also be static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM), etc. That is to say, the memory in the embodiments of this application is intended to include, but is not limited to, these and any other suitable types of memory.

[0174] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0175] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0176] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0177] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0178] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0179] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer or a server) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0180] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An information recommendation method, characterized in that, The method includes: Based on user profile information, long-term interest features of users are extracted. The long-term interest features are obtained by converting the user profile information into a corresponding dense vector and extracting it based on the dense vector corresponding to the user profile information. Specifically, this includes: inputting the user profile information after the embedding lookup layer processing into the user profile module for user profile embedding in the user profile module; and processing it through a translation layer to obtain profile units, which are used to represent the user's long-term interest features. Based on click sequence information in user historical behavior data, short-term interest features of users are extracted. These short-term interest features are obtained by converting the click sequence information into corresponding dense vectors and extracting them from these dense vectors. Specifically, this includes: inputting user historical behavior data processed by an embedding lookup layer into a user behavior sequence module; performing comprehensive processing on the dense vectors corresponding to the click sequence information through a sequence embedding pooling layer in the user behavior sequence module; and then processing the comprehensive dense vectors corresponding to the click sequence information through a first weighted translation self-attention layer and a second weighted translation self-attention layer to obtain sequence units, which are used to represent the user's short-term interest features. Based on the item attribute information in the user's historical behavior data, the user's item attribute preference vector is calculated, including: fusing the user's long-term interest features and short-term interest features to obtain a first fusion vector; processing the first fusion vector and the item attribute information using a multi-head attention mechanism of a top-N self-attention layer through the item attribute preference module to calculate the user's item attribute preference vector, wherein the item attribute preference module is used to model the relationship between the attributes of the user's historical interaction items and the recommendation structure, wherein the top-N self-attention layer is used to focus attention on the top N data; the item attribute preference vector is obtained by converting the item attribute information into a corresponding dense vector and calculating it based on the dense vector corresponding to the item attribute information. The user's long-term interest characteristics, short-term interest characteristics, and item attribute preference information are fused together to obtain a user vector; The multi-task learning module optimizes multiple objectives in the target model based on the user vector to obtain an optimized target model. This multi-task learning module includes multiple different expert networks and multiple different gates. It also includes a common structure for constructing common information among the multiple objectives, allowing the user vector to be split into common and differential information for differentiated processing. The multiple objectives include a playback objective, a sharing objective, and a retention objective. Specifically, this includes: obtaining project feedback information containing retention data, where retention represents the number of users brought in after sharing; and optimizing the playback objective, the sharing objective, and the retention objective based on the user vector and the project feedback information. Optimization is performed to obtain an optimized target model. Specifically, based on offline services, the playback target embedding, sharing target embedding, and the results processed by the multi-task learning module based on the sharing target and the playback target are input into the overall loss function of the target model for training. Additionally, the item feedback information and the results processed by the multi-task learning module based on the sharing target and the playback target are input into the retention target loss function for training. When the top K items most interested in by the sample users best match the historical recommended items, the synchronous optimization of the playback target, the sharing target, and the retention target in the target model is completed, resulting in the optimized target model. The optimized target model is used to process the project information in the project information database to generate recall information for the target user, and recommendations are made based on the recall information.

2. The information recommendation method as described in claim 1, characterized in that, The step of extracting long-term interest features of users based on user profile information includes: Obtain user profile information, which includes user demographic attribute information and user preference information obtained based on the user's historical behavior data; Based on the user demographic information and the user preference information, extract the user's long-term interest features.

3. The information recommendation method as described in claim 1, characterized in that, The step of optimizing multiple objectives in the target model based on the user vector to obtain an optimized target model includes: The playback target and the sharing target in the target model are optimized based on the user vector to obtain the optimized target model.

4. The information recommendation method as described in claim 3, characterized in that, The target model includes a multi-task learning module. The optimization of the playback target and the sharing target in the target model based on the user vector to obtain an optimized target model includes: The user vector is processed by the multi-task learning module to split the user vector into common information and difference information. The common information is the common information corresponding to the playback target and the sharing target, and the difference information is the difference information corresponding to the playback target and the sharing target. The playback target and the sharing target in the target model are optimized based on the shared information and the difference information to obtain an optimized target model.

5. The information recommendation method as described in claim 1, characterized in that, Based on the user vector and the project feedback information, the playback target, the sharing target, and the retention target in the target model are optimized to obtain an optimized target model, including: Based on the user vector, optimize the playback target in the target model; Based on the user vector and the project feedback information, the sharing target and the retention target in the target model are optimized; Based on the optimized playback target, sharing target, and retention target, the optimized target model is obtained.

6. The information recommendation method as described in claim 1, characterized in that, The process of processing project information in the project information database using the optimized target model to generate recall information corresponding to the target user, and making recommendations based on the recall information, includes: The user vector of the target user is obtained through the optimized target model; Calculate the cosine similarity between the user vector of the target user and the project vectors in the project information of the project information database; Based on the cosine similarity, recall information corresponding to the target user is generated; Recommendation information is determined from the recall information and recommended to the target user.

7. The information recommendation method as described in claim 6, characterized in that, The step of generating recall information corresponding to the target user based on the cosine similarity includes: A first candidate project set is determined from the project information in the project information database based on the cosine similarity. The first candidate item set is initially selected to obtain a second candidate item set, wherein the number of items in the second candidate item set is less than the number of items in the first candidate item set; The candidates in the second candidate item set are sorted, and the top K candidate items in the sorted second candidate item set are selected as the recall information corresponding to the target user.

8. An information recommendation device, characterized in that, The device includes: The first extraction unit is used to extract long-term interest features of users based on user profile information. The long-term interest features are obtained by converting the user profile information into a corresponding dense vector and extracting it based on the dense vector corresponding to the user profile information. Specifically, it includes: inputting the user profile information after the embedding lookup layer processing into the user profile module for user profile embedding in the user profile module; and processing it through the translation layer to obtain a profile unit, which is used to represent the user's long-term interest features. The second extraction unit is used to extract the user's short-term interest features based on the click sequence information in the user's historical behavior data. The short-term interest features are obtained by converting the click sequence information into a corresponding dense vector and extracting it based on the dense vector corresponding to the click sequence information. Specifically, it includes: inputting the user's historical behavior data after the embedding lookup layer processing into the user behavior sequence module; performing comprehensive processing on the dense vector corresponding to the click sequence information through the sequence embedding pooling layer in the user behavior sequence module; and processing the comprehensive processing of the dense vector corresponding to the click sequence information through a first weighted translation self-attention layer and a second weighted translation self-attention layer to obtain a sequence unit, which is used to represent the user's short-term interest features. The calculation unit is used to calculate the user's item attribute preference vector based on the item attribute information in the user's historical behavior data, including: fusing the user's long-term interest features and short-term interest features to obtain a first fusion vector; processing the first fusion vector and the item attribute information using a multi-head attention mechanism of a topN self-attention layer through an item attribute preference module to calculate the user's item attribute preference vector, wherein the item attribute preference module is used to model the relationship between the attributes of the user's historical interaction items and the recommendation structure, wherein the topN self-attention layer is used to focus attention on the top N data; the item attribute preference vector is obtained by converting the item attribute information into a corresponding dense vector and calculating it based on the dense vector corresponding to the item attribute information. The fusion unit is used to fuse the user's long-term interest features, short-term interest features, and item attribute preference information to obtain a user vector. An optimization unit is used to optimize multiple objectives in the target model based on the user vector using a multi-task learning module to obtain an optimized target model. The multi-task learning module includes multiple different expert networks and multiple different gates. The multi-task learning module also includes a common structure for constructing common information among the multiple objectives, so as to decompose the user vector into common information and differential information for differentiated processing. The multiple objectives include a playback objective, a sharing objective, and a retention objective. Specifically, this includes: obtaining project feedback information containing retention data, where retention data represents the number of users brought in after sharing; and optimizing the playback objective, the sharing objective, and the retention objective based on the user vector and the project feedback information. The target objectives are optimized to obtain an optimized target model. Specifically, based on offline services, the playback target embedding, sharing target embedding, and the results processed by the multi-task learning module based on the sharing and playback targets are input into the overall loss function of the target model for training. Additionally, the item feedback information and the results processed by the multi-task learning module based on the sharing and playback targets are input into the retention target loss function for training. When the top K items most interested in by the sample users best match the historical recommended items, the playback target, sharing target, and retention target in the target model are simultaneously optimized to obtain the optimized target model. The recommendation unit is used to process the project information in the project information database through the optimized target model to generate recall information corresponding to the target user, and to make recommendations based on the recall information.

9. The information recommendation device as described in claim 8, characterized in that, The optimization unit is further configured to optimize the playback target and the sharing target in the target model based on the user vector to obtain an optimized target model.

10. The information recommendation device as described in claim 8, characterized in that, The target model includes a multi-task learning module. The optimization unit is further configured to: process the user vector through the multi-task learning module to split the user vector into common information and difference information, wherein the common information is the common information corresponding to the playback target and the sharing target, and the difference information is the difference information corresponding to the playback target and the sharing target; optimize the playback target and the sharing target in the target model based on the common information and the difference information to obtain an optimized target model.

11. The information recommendation device as described in claim 8, characterized in that, The optimization unit is further configured to: optimize the playback target in the target model based on the user vector; optimize the sharing target and the retention target in the target model based on the user vector and the project feedback information; and obtain the optimized target model based on the optimized playback target, the sharing target, and the retention target.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted for loading by a processor to perform the steps of the information recommendation method as described in any one of claims 1-7.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, and the processor executing the steps of the information recommendation method according to any one of claims 1-7 by calling the computer program stored in the memory.

Images