Content recommendation method, apparatus, device, and readable storage medium
By introducing a hierarchical sharing mechanism of expert network layers into the multi-task model and using the attention mechanism for feature focusing, the problem of incomplete feature representation in existing technologies is solved, thereby improving the accuracy and precision of content recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN YAYUE TECH CO LTD
- Filing Date
- 2021-02-20
- Publication Date
- 2026-05-01
AI Technical Summary
In existing technologies, multi-task models only focus on feature sharing between different layers when recommending content, failing to achieve comprehensive feature representation and resulting in low learning accuracy.
By using attention between the representation vector of the recommendation task and the representation vectors of all layers of the expert network, instead of the simple gated flat-layer expert feature sharing in MMoE, we can achieve comprehensive focus on feature representation and improve the prediction accuracy of multi-task models.
It improves the prediction accuracy of multi-task models and the recommendation accuracy of candidate content, and achieves comprehensive focus of feature expression by focusing attention on important information.
Smart Images
Figure CN113569130B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence, and in particular to a content recommendation method, apparatus, device, and readable storage medium. Background Technology
[0002] In the field of artificial intelligence, content recommendation is a process of recommending candidate content to a user account by combining the account characteristics of the user account and the content characteristics of the candidate content. Optionally, content that users are interested in can be recommended to users based on their interests and browsing behavior. The account characteristics of the user account are obtained by feature extraction based on the settings of interests and browsing behavior.
[0003] In related technologies, when content recommendation involves at least two recommendation tasks, it is necessary to focus not only on features from different perspectives related to the recommendation tasks, but also on feature extraction at different levels related to the recommendation tasks themselves. Therefore, a multi-gate Mixture-of-Experts (MMoE) model was proposed, which shares features of flat expert networks through gating.
[0004] However, related technologies only focus on feature sharing between layers, failing to achieve comprehensive focus on feature representation, resulting in low learning accuracy for multi-task models. Summary of the Invention
[0005] This application provides a content recommendation method, apparatus, device, and readable storage medium, which can improve the accuracy of content recommendation. The technical solution is as follows:
[0006] On the one hand, a content recommendation method is provided, the method comprising:
[0007] Obtain content features, which are features obtained by feature extraction from candidate content;
[0008] The content features are processed by a task network to obtain task features, wherein the task network is a network set up for a recommendation task based on the candidate content.
[0009] The content features are processed by at least two expert network layers to obtain at least two expert features, where the i-th expert network layer corresponds to the i-th expert feature, and i is a positive integer. The at least two expert network layers are used to process the content features based on the corresponding extraction and summarization requirements.
[0010] Based on the task features and the expert features, determine the attention focus information corresponding to the recommendation task and the at least two expert network layers;
[0011] Recommendation of candidate content is made based on the attention focus information.
[0012] On the other hand, a content recommendation device is provided, the device comprising:
[0013] The acquisition module is used to acquire content features, which are features obtained by feature extraction of candidate content.
[0014] The processing module is used to perform first feature processing on the content features through a task network to obtain task features, wherein the task network is a network set up based on the recommendation task of the candidate content;
[0015] The processing module is further configured to perform a second feature processing on the content features through at least two expert network layers to obtain at least two expert features, wherein the i-th expert network layer corresponds to the i-th expert feature, and i is a positive integer, wherein the at least two expert network layers are used to perform feature processing on the content features based on the corresponding extraction and summarization requirements respectively;
[0016] The determination module is used to determine the attention focus information corresponding to the recommendation task and the at least two expert network layers based on the task features and the expert features;
[0017] The recommendation module is used to recommend candidate content based on the attention focus information.
[0018] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one program, the at least one program being loaded and executed by the processor to implement the content recommendation method as described in any of the embodiments of this application above.
[0019] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the content recommendation method as described in any of the embodiments of this application above.
[0020] On the other hand, a computer program product or computer program is provided, which 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 any of the content recommendation methods described in the above embodiments.
[0021] The beneficial effects of the technical solutions provided in this application include at least the following:
[0022] By performing attention between the representation vector of the recommendation task and the representation vectors of all layers of the expert network, the attention focus of the recommendation task on each layer of the expert network is obtained. The focus result of the attention is used as the input of each subsequent prediction process to replace the simple gating method of sharing features of experts in the same layer in MMoE, so as to achieve comprehensive focus on feature representation, thereby improving the prediction accuracy of the multi-task model and the recommendation accuracy of candidate content. Attached Figure Description
[0023] 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.
[0024] Figure 1 This is a schematic diagram of the structure of the MMoE model provided in related technologies;
[0025] Figure 2 This is a schematic diagram of the structure of a multi-task model provided in an exemplary embodiment of this application;
[0026] Figure 3 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of this application;
[0027] Figure 4 This is a schematic diagram of the structure of a news recommendation system provided in an exemplary embodiment of this application;
[0028] Figure 5 This is a flowchart of a content recommendation method provided in an exemplary embodiment of this application;
[0029] Figure 6 This is a flowchart of a content recommendation method provided in another exemplary embodiment of this application;
[0030] Figure 7 This is a flowchart of a content recommendation method provided in another exemplary embodiment of this application;
[0031] Figure 8 Based on Figure 7 The illustrated embodiment provides a flowchart of the process for acquiring attention-focusing information;
[0032] Figure 9 This is a structural block diagram of a content recommendation apparatus provided in an exemplary embodiment of this application;
[0033] Figure 10 This is a structural block diagram of a content recommendation apparatus provided in another exemplary embodiment of this application;
[0034] Figure 11 This is a schematic diagram of the structure of a computer device provided in an exemplary embodiment of this application. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.
[0036] First, a brief introduction to the terms used in the embodiments of this application:
[0037] 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.
[0038] 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.
[0039] In this embodiment, the main focus is on machine learning models. These models predict the click-through rate of candidate content, thereby ranking the candidate content in the candidate content set and recommending the top-ranked candidate content.
[0040] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instruction-based learning.
[0041] Recommendation systems are tools that automatically connect users with recommended content, assisting users in discovering potentially interesting content in information-overloaded environments. In some embodiments, after candidate content to be recommended is input into the recommendation system for a target account, the system determines the probability that the target account will select the candidate content based on the content characteristics of the candidate content and the account characteristics of the target account. For example, it determines the click-through rate of the target account selecting candidate content; or, it determines the conversion rate of the target account browsing candidate content.
[0042] In some embodiments, when recommending content, recommendations are typically made to a target account by inputting a set of candidate content. This set includes multiple candidate content items. The candidate content items are ranked based on the probability that each item will be selected by the user corresponding to the target account, and then recommended based on this ranking. The recommendation system described in this application can be applied to various programs such as news programs, video programs, and product purchase programs to achieve content recommendation. For example, a news program is used. A target account is logged into the news program. During the news recommendation process, n candidate news items (n being a positive integer) are first obtained from the content pool. The recommendation system then predicts the click-through rate (CTR) of these n candidate news items, obtaining the CTR prediction result for each candidate news item. Based on the CTR prediction results, the n candidate news items are ranked and then recommended.
[0043] Attention mechanism: refers to a focusing method derived from human vision. In the process of visually perceiving objects, we usually do not browse a scene completely, but pay attention to a specific part according to the needs. The attention mechanism learns to focus attention on the part that the user will pay attention to when a similar scene appears. The essence of the attention mechanism is weighting.
[0044] Typically, multi-task learning in deep neural networks is widely used in business applications such as recommendation systems. For example, when recommending information to users in a recommendation system, it's not only necessary to recommend information that users are interested in, but also to promote the conversion rate of information completion. Therefore, it's necessary to model both click-through rate and conversion rate simultaneously. For multi-task models, it's necessary to focus not only on features from different angles related to the task itself (i.e., flat features), but also on feature extraction at different levels related to the task itself (i.e., hierarchical attention), thereby improving the accuracy of the multi-task model. In related technologies, multi-task models (Multi-gate Mixture-of-Experts, MMoE) are used for click-through rate prediction and ranking of candidate content by sharing feature vectors or sharing flat expert network features.
[0045] This is illustrative; please refer to it. Figure 1 The diagram illustrates the structure of the MMoE model in related technologies. The MMoE model learns the relationships between multiple tasks by sharing a lower layer and using gating. Each task uses a separate gating network, and the features of the expert layer are selectively utilized through the weights output by the gating network. Figure 1 As shown, task 110 corresponds to gating network 111, and task 120 corresponds to gating network 121. The features of expert layers 131, 132, and 133 are selectively utilized through the weights output by gating network 111 for the prediction process of subsequent task 110; and the features of expert layers 131, 132, and 133 are selectively utilized through the weights output by gating network 121 for the prediction process of subsequent task 120.
[0046] However, the MMoE multi-task model only focuses on feature sharing between horizontal layers (horizontal focus) at the feature level, while ignoring feature sharing between different levels after feature extraction (hierarchical focus).
[0047] Drawing inspiration from Convolutional Neural Networks (CNNs) for extracting and representing image features, the level of generalization of the extracted representations gradually increases from different layers. For example, pixels, textures, and shapes are extracted sequentially from the bottom to the top. The layer-by-layer extraction of features by neural networks also exhibits a similar effect of progressively increasing abstraction.
[0048] Given the understanding of progressively abstracting feature representations, this application provides a multi-task association learning model based on layer-sharing of hierarchical expert networks. By learning the original input feature information, feature representations for different tasks are obtained. The attention results between the task feature representations and the hierarchical expert network are used as the input of the subsequent prediction network to replace the simple gating method of feature sharing in flat-layer expert networks in MMoE.
[0049] This is illustrative; please refer to it. Figure 2 It illustrates a structural schematic diagram of a multi-task model provided in an exemplary embodiment of this application, such as... Figure 2As shown, task features are extracted for tasks 210 and 220 respectively, resulting in task feature 211 for task 210 and task feature 221 for task 220. Based on the progressively higher expert network layers 231, 232 and 233, attention analysis is performed on task features 211 and 221 to obtain attention distribution information, and subsequent task predictions are made, such as click-through rate prediction or completion rate prediction.
[0050] Figure 3 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of this application, such as... Figure 3 As shown, the implementation environment includes a terminal 310 and a server 320, wherein the terminal 310 and the server 320 are connected through a communication network 330.
[0051] Terminal 310 is used to send a content retrieval request to server 320 for content recommendation. The content recommendation includes at least one of the following forms: news content recommendation, video content recommendation, product content recommendation, etc. Terminal 310 sends the content retrieval request to server 320 via communication network 330, and instructs server 320 to sort and recommend candidate content based on the content retrieval request and the user profile of the target account logged into terminal 310.
[0052] The above description uses the example of terminal 310 sending a content acquisition request to server 320 to obtain content recommendations. In some embodiments, server 320 actively retrieves candidate content from the content pool based on user profiles and recommends it to terminal 310 every preset time interval.
[0053] Server 320 includes a multi-task model 321, which is used to recommend candidate content to terminal 310 based on at least two recommendation tasks.
[0054] After receiving the content retrieval request, the server 320 combines the user profile of the target account to extract features from the candidate content in the content pool, and uses a multi-task model 321 corresponding to at least two recommendation tasks to predict the click-through rate and / or conversion rate of each candidate content, thereby feeding back recommended content to the terminal 310 based on the prediction results.
[0055] It is worth noting that the content recommendation method provided in this application embodiment can be implemented by a terminal, a server, or a combination of both. That is, in the above embodiments, the scene recognition model 321 is located in the server 320 as an example. In some embodiments, the scene recognition model 321 can also be configured in the terminal 310, or part of the network in the scene recognition model 321 can be configured in the terminal 310, while other parts can be configured in the server 320.
[0056] The terminals include at least one of smartphones, tablets, laptops, desktop computers, smart speakers, and smart wearable devices. The servers can be physical servers or cloud servers providing cloud computing services. A server can be a single server or a server cluster or distributed system composed of multiple servers. When the terminals and servers collaborate to implement the solutions provided in this application embodiment, the terminals and servers can be directly or indirectly connected via wired or wireless communication methods; this application embodiment does not limit this connection.
[0057] Let's take the recommendation of news content as an example to illustrate this. Figure 4 This is a schematic diagram of the structure of a news recommendation system provided in an exemplary embodiment of this application, such as... Figure 4 As shown, the news recommendation system 400 includes four modules: user profile module 410, recall module 420, click-through rate (CTR) prediction module 430, and reordering module 440.
[0058] After the recall module 420 filters and roughly sorts the news articles from the content pool 401, it outputs them to the CTR prediction module 430. The CTR prediction module 430 performs fine sorting on these recalled articles, and then the re-sorting module 440 finally recommends them to users. The user profiling module 410 assists the recall, CTR prediction, and re-sorting processes with user profiles. The user profiling module 410 includes user interest tags and historical interaction data. Historical interaction data refers to the user's interaction data with historical news content within a specified historical time period, such as click data, browsing data, collection data, and forwarding data.
[0059] Based on the above explanation, Figure 5 A flowchart of a content recommendation method provided as an exemplary embodiment of this application is illustrated using an example of the method being applied to a server. Figure 5 As shown, the method includes:
[0060] Step 501: Obtain content features, which are the features obtained by feature extraction of candidate content.
[0061] In some embodiments, content features are obtained through a specified feature extraction network. In some embodiments, a content pool is obtained, which includes at least two candidate contents to be ranked. Feature extraction is performed on the candidate contents in the content pool to obtain the content features corresponding to each candidate content.
[0062] The methods for obtaining content features include at least one of the following:
[0063] First, the content in the content pool is updated regularly. When the content in the content pool is updated, the features of the candidate content in the content pool are extracted in advance to obtain content features. Then, when the user needs content recommendations, the extracted content features are directly obtained based on the content recommendation request.
[0064] Second, the content in the content pool is updated periodically. When a user requests content recommendations, candidate content from the content pool is retrieved based on the recommendation request, and feature extraction is performed to obtain content features.
[0065] Step 502: Perform first feature processing on the content features through the task network to obtain task features.
[0066] The task network is a network set up for recommendation tasks based on candidate content.
[0067] In some embodiments, the model includes task networks corresponding to at least two recommendation tasks. The content features are processed by the fully connected layers of the task networks to obtain vector representations corresponding to different recommendation tasks.
[0068] Step 503: Perform second feature processing on the content features through at least two expert network layers to obtain at least two expert features.
[0069] The i-th expert network layer corresponds to the i-th expert feature, where i is a positive integer. At least two expert network layers are used to perform feature processing on the content features based on their respective extraction and generalization requirements. Each expert network layer is a neural network formed by mapping the original input features to different feature spaces; that is, each expert network layer corresponds to a feature space mapping method.
[0070] In some embodiments, at least two expert network layers are sequentially arranged. These at least two expert network layers are arranged in order of progressively increasing generalization level when processing content features. For example, for news content, if the generalization process progressively increases in the order of text, semantics, and summary, then the at least two expert network layers extract features sequentially from bottom to top in the order of text, semantics, and summary. Specifically, after the k-th expert network layer performs second feature processing on the content features to obtain the k-th expert feature, if the number of expert network layers is greater than k, the (k+1)-th expert network layer performs second feature processing on the k-th expert feature to obtain the (k+1)-th expert feature, and so on, until at least two expert features are obtained.
[0071] Indicatively, the model includes an expert network layer A for text feature extraction, an expert network layer B for semantic feature extraction, and an expert network layer C for summary feature extraction. When expert network layer A processes the content features using text features, it obtains expert feature a. Expert feature a is then input into expert network layer B, which processes it using semantic features to obtain expert feature b. Expert feature b is then input into expert network layer C, which processes it using summary features to obtain expert feature c. Thus, expert features a, b, and c are at least two expert features extracted through at least two expert network layers.
[0072] Between expert network layers A, B, and C, the level of generalization of the feature representations extracted at different levels increases progressively according to the order of the three expert network layers. The input of an upper expert network layer is the output of the lower adjacent expert network layer. This simultaneously focuses on feature sharing between layers and feature sharing between different layers.
[0073] It is worth noting that steps 502 and 503 are two parallel steps. Step 502 can be executed first, or step 503 can be executed first, or both steps 502 and 503 can be executed simultaneously. The multiple expert network layers in step 503 are arranged in a hierarchical order, with the order of the expert network layers from the bottom to the top according to the progressively increasing level of generalization.
[0074] Additionally, it is worth noting that in this embodiment, at least two expert network layers in an expert network are used as an example for illustration. In some embodiments, it can also be implemented as at least two expert network layers in multiple expert networks. This embodiment does not limit this.
[0075] Step 504: Based on task features and expert features, determine the attention focus information corresponding to the recommendation task and at least two expert network layers.
[0076] Attention analysis is performed on task features and expert features using the attention mechanism to obtain the focus result of attention, that is, the attention focus of at least two expert network layers corresponding to the recommendation task.
[0077] Step 505: Recommend candidate content based on attention focus information.
[0078] In some embodiments, candidate content is ranked based on attention-focusing information according to a recommendation task. For example, if the recommendation task includes a click-through rate prediction task, then the click-through rate of candidate content is first predicted based on attention-focusing information to obtain the click-through rate of candidate content in the content pool. Then, the candidate content in the content pool is ranked based on the click-through rate, and the candidate content is recommended based on the ranking result.
[0079] In another embodiment, the recommendation task includes a click-through rate prediction task and a conversion rate prediction task, wherein the conversion rate prediction refers to the prediction of the degree to which a user completes browsing the recommended content.
[0080] The task involves two parts: Click-through rate (CTR) prediction (Task 1) and conversion rate prediction (Task 2). Task 1 first predicts the CTR of candidate content based on attention focus information, obtaining the CTR of candidate content in the content pool. Then, for Task 1, candidate content in the content pool is ranked based on CTR. Similarly, Task 2 first predicts the conversion rate of candidate content based on attention focus information, obtaining the conversion rate of candidate content in the content pool. Then, for Task 2, candidate content in the content pool is ranked based on conversion rate. Finally, the ranking results from Task 1 and Task 2 are combined to obtain a comprehensive ranking result for the candidate content. Recommendations are then made based on this comprehensive ranking result.
[0081] In summary, the content recommendation method provided in this embodiment performs attention between the expression vector of the recommendation task and the expression vectors of all layers of the expert network layer, respectively, to obtain the attention focus of the recommendation task on each layer of the expert network layer. The focus result of the attention is used as the input of each subsequent prediction process to replace the simple gating method of sharing expert features in MMoE, so as to achieve comprehensive focus on feature expression, thereby improving the prediction accuracy of the multi-task model and the recommendation accuracy of candidate content.
[0082] In some embodiments, feature mapping of content features is performed using at least two task networks. Figure 6 This is a flowchart of a content recommendation method provided in another exemplary embodiment of this application. The method is illustrated using an example of its application in a server. Figure 6 As shown, the method includes:
[0083] Step 601: Obtain content features, which are the features obtained by feature extraction of candidate content.
[0084] In some embodiments, content features are obtained through a specified feature extraction network. In some embodiments, a content pool is obtained, which includes at least two candidate contents to be ranked. Feature extraction is performed on the candidate contents in the content pool to obtain the content features corresponding to each candidate content.
[0085] Step 602: Perform first feature processing on the content features through at least two task networks to obtain at least two task features.
[0086] The m-th task network corresponds to the m-th task feature, where m is a positive integer. At least two task networks are networks set up based on at least two recommendation tasks for candidate content, and the m-th task network corresponds to the m-th recommendation task.
[0087] The task network is designed based on recommendation tasks for candidate content; that is, each candidate content corresponds to at least two recommendation tasks, and at least one task network is set up for feature mapping for each recommendation task. For example, in a news recommendation scenario, candidate news content corresponds to click-through rate (CTR) and conversion rate (CTR) tasks. The task network includes a first network based on the CTR task and a second network based on the CTR task. The content features are processed using the first and second networks respectively to obtain the first task features corresponding to the first network and the second task features corresponding to the second task. Both the first and second networks include fully connected layers of neural networks trained using sample data. The sample data can be historical content interaction data of users.
[0088] Step 603: Perform second feature processing on the content features through at least two expert network layers to obtain at least two expert features.
[0089] The i-th expert network layer corresponds to the i-th expert feature, where i is a positive integer. At least two expert network layers are used to perform feature processing on the content features based on their respective extraction and generalization requirements. Each expert network layer is a neural network formed by mapping the original input features to different feature spaces; that is, each expert network layer corresponds to a feature space mapping method.
[0090] It is worth noting that steps 602 and 603 are two parallel steps. Step 602 can be executed first, or step 603 can be executed first, or steps 602 and 603 can be executed simultaneously. Alternatively, they can be executed alternately for at least two task networks and at least two expert networks; this embodiment does not limit this.
[0091] Step 604: Based on the features of the m-th task and at least two expert features, determine the attention focus information corresponding to the m-th recommendation task and at least two expert network layers.
[0092] Determine the attention analysis results between each task feature and at least two expert features in at least two task features, that is, determine the attention focus information corresponding to each recommendation task and at least two expert network layers in at least two recommendation tasks.
[0093] To illustrate, taking a news recommendation scenario as an example, candidate content corresponds to click-through rate (CTR) tasks and conversion rate tasks. The first task feature corresponding to the CTR task and the second task feature corresponding to the conversion rate task are extracted. Then, the first task feature is used to perform attention with at least two expert features to determine the attention focus information corresponding to the CTR task and at least two expert network layers; similarly, the second task feature is used to perform attention with at least two expert features to determine the attention focus information corresponding to the conversion rate task and at least two expert network layers.
[0094] Step 605: Recommend candidate content based on attention focus information.
[0095] Optionally, for the m-th recommendation task, the attention focusing results from all levels of at least two expert network layers are concatenated and then input into the prediction network model corresponding to the m-th recommendation task to obtain the output result corresponding to the m-th recommendation task.
[0096] For illustrative purposes, please refer to Formula 1 below, which expresses the recommendation method for a recommendation task:
[0097] Formula 1:
[0098] in, This represents the output of the recommendation task. denoted as the hidden layer, i.e., the hidden layer in the prediction network model structure; concat represents the concatenation of attention-focusing information; and E represents the expression vector corresponding to the expert features output by each layer of the expert network. This represents the k-th task feature. This represents the attention focus information for the k-th recommendation task.
[0099] In some embodiments, the task network includes a first network set up based on the click-through rate task. Then, based on the attention focus information of at least two expert network layers corresponding to the click-through rate task, the click-through rate of candidate content in the click-through rate task is predicted to obtain the click-through rate prediction result. Based on the click-through rate prediction result, the candidate content is recommended. Schematic, the candidate content in the content pool is sorted based on the click-through rate prediction result, and the candidate content is recommended according to the sorting result.
[0100] In other embodiments, since the recommendation model is a multi-task-based recommendation model, the task network also includes a second network set up based on the conversion rate task. Therefore, when recommending candidate content, the model further includes predicting the conversion rate of the candidate content in the conversion rate task based on the attention focus information of at least two expert network layers corresponding to the conversion rate task, obtaining a conversion rate prediction result, and recommending candidate content based on the click-through rate prediction result and the conversion rate prediction result. Illustratively, candidate content in the content pool is first ranked based on the click-through rate prediction result, and second ranked based on the conversion rate prediction result. The results of the first and second rankings are then combined to obtain a comprehensive ranking result, and candidate content is recommended based on this comprehensive ranking result. For example, the weighted average of the order of candidate content in the first ranking and the order in the second ranking is used to obtain the comprehensive ranking result of the candidate content.
[0101] For example, if the target candidate content is ranked 20th in the first ranking with a weight of 0.8, and 5th in the second ranking with a weight of 0.2, then the overall ranking result of the target candidate content is 16+1, which is the 17th position. If the value is not an integer, it is rounded down by taking the integer part. It is worth noting that the above method of obtaining the overall ranking result is only an illustrative example. In this application embodiment, the overall ranking result can also be obtained in other ways, such as analyzing the ranking result through an additional neural network layer, etc. This application embodiment does not limit this method.
[0102] In summary, the content recommendation method provided in this embodiment performs attention between the expression vector of the recommendation task and the expression vectors of all layers of the expert network layer, respectively, to obtain the attention focus of the recommendation task on each layer of the expert network layer. The focus result of the attention is used as the input of each subsequent prediction process to replace the simple gating method of sharing expert features in MMoE, so as to achieve comprehensive focus on feature expression, thereby improving the prediction accuracy of the multi-task model and the recommendation accuracy of candidate content.
[0103] The method provided in this embodiment enables each recommendation task to selectively focus on the information of the expert network layer from different angles and at different levels of granularity, thereby placing attention on important information that needs attention and improving attention accuracy.
[0104] Compared with the simple gating method in MMoE for sharing features of horizontal experts, the multi-task association learning model based on hierarchical expert network layer sharing proposed in this application focuses on features from different angles related to the task itself (horizontal attention) and features from different levels related to the task itself (hierarchical attention), achieving comprehensive focus on feature representation and thus improving the learning accuracy of the model.
[0105] In an optional embodiment, attention analysis is performed on task features and expert features using an attention mechanism. Figure 7 This is a flowchart of a content recommendation method provided in another exemplary embodiment of this application. The method is illustrated using an example of its application in a server. Figure 7 As shown, the method includes:
[0106] Step 701: Obtain content features, which are the features obtained by feature extraction of candidate content.
[0107] In some embodiments, content features are obtained through a specified feature extraction network. In some embodiments, a content pool is obtained, which includes at least two candidate contents to be ranked. Feature extraction is performed on the candidate contents in the content pool to obtain the content features corresponding to each candidate content.
[0108] Step 702: Perform first feature processing on the content features through the task network to obtain task features.
[0109] The task network is a network set up for recommendation tasks based on candidate content.
[0110] In some embodiments, the model includes task networks corresponding to at least two recommendation tasks. The content features are processed by the fully connected layers of the task networks to obtain vector representations corresponding to different recommendation tasks.
[0111] Step 703: Perform second feature processing on the content features through at least two expert network layers to obtain at least two expert features.
[0112] The i-th expert network layer corresponds to the i-th expert feature, where i is a positive integer. At least two expert network layers are used to perform feature processing on the content features based on their respective extraction and generalization requirements. Each expert network layer is a neural network formed by mapping the original input features to different feature spaces; that is, each expert network layer corresponds to a feature space mapping method.
[0113] It is worth noting that steps 702 and 703 are two parallel steps. Step 702 can be executed first, or step 703 can be executed first, or steps 702 and 703 can be executed simultaneously.
[0114] Step 704: Perform attention analysis on task features and expert features based on the attention mechanism to obtain attention focus information corresponding to the recommendation task and at least two expert network layers.
[0115] In some embodiments, when performing attention analysis on task features and expert features, the matrix product of the task features and the i-th expert feature is determined, the matrix product is subjected to embedding scaling to obtain a scaling matrix, the scaling matrix is then classified, and based on the classification results corresponding to at least two expert network layers, attention focus information corresponding to the recommendation task and at least two expert network layers is determined.
[0116] As an illustration, please refer to Formula Two below for the method of acquiring information with this focused attention:
[0117] Formula 2:
[0118] Where E represents the expression vector corresponding to the expert features output by each layer of the expert network. Let d represent the k-th task feature, d represent the feature mapping dimension, and T represent the number of expert network layers.
[0119] According to Formula 1 above, by performing attention between the expression vectors (i.e., task features) of different task networks and the expression vectors (expert features) of all levels of the expert network layer, we can obtain the attention focus of different recommendation tasks on all levels of the expert network layer. The focus result of the attention is then used as the input for the subsequent prediction process.
[0120] like Figure 8As shown, in the process of acquiring attention-focusing information, the task feature 810 and the expert feature 820 are first multiplied by a matrix to obtain a matrix product 830. The matrix product 830 is then scaled to obtain a scaling matrix 840. After classification analysis based on the scaling matrix 840, a classification result 850 is obtained. Based on the classification results of at least two expert network layers, the attention-focusing information 860 corresponding to the recommendation task and at least two expert network layers is determined. Schematic, the classification results of at least two expert network layers are multiplied by a matrix to obtain the attention-focusing information 860.
[0121] Step 705: Recommend candidate content based on attention focus information.
[0122] In some embodiments, candidate content is ranked based on attention-focusing information according to a recommendation task. For example, if the recommendation task includes a click-through rate prediction task, then the click-through rate of candidate content is first predicted based on attention-focusing information to obtain the click-through rate of candidate content in the content pool. Then, the candidate content in the content pool is ranked based on the click-through rate, and the candidate content is recommended based on the ranking result.
[0123] In another embodiment, the recommendation task includes a click-through rate prediction task and a conversion rate prediction task, wherein the conversion rate prediction refers to the prediction of the degree to which a user completes browsing the recommended content.
[0124] In summary, the content recommendation method provided in this embodiment performs attention between the expression vector of the recommendation task and the expression vectors of all layers of the expert network layer, respectively, to obtain the attention focus of the recommendation task on each layer of the expert network layer. The focus result of the attention is used as the input of each subsequent prediction process to replace the simple gating method of sharing expert features in MMoE, so as to achieve comprehensive focus on feature expression, thereby improving the prediction accuracy of the multi-task model and the recommendation accuracy of candidate content.
[0125] The method provided in this embodiment enables each recommendation task to selectively focus on the information of the expert network layer from different angles and at different levels of granularity, thereby placing attention on important information that needs attention and improving attention accuracy.
[0126] Figure 9 This is a structural block diagram of a content recommendation apparatus provided in an exemplary embodiment of this application, such as... Figure 9 As shown, the device includes:
[0127] The acquisition module 910 is used to acquire content features, wherein the content features are features obtained by feature extraction of candidate content;
[0128] The processing module 920 is used to perform first feature processing on the content features through a task network to obtain task features, wherein the task network is a network set up based on the recommendation task of the candidate content;
[0129] The processing module 920 is further configured to perform a second feature processing on the content features through at least two expert network layers to obtain at least two expert features, wherein the i-th expert network layer corresponds to the i-th expert feature, and i is a positive integer, wherein the at least two expert network layers are used to perform feature processing on the content features based on the corresponding extraction and summarization requirements respectively;
[0130] The determination module 930 is used to determine the attention focus information corresponding to the recommendation task and the at least two expert network layers based on the task features and the expert features;
[0131] The recommendation module 940 is used to recommend candidate content based on the attention focus information.
[0132] In some embodiments, the at least two expert network layers are arranged in a sequential order;
[0133] The processing module 920 is further configured to perform a second feature processing on the content feature through the kth expert network layer to obtain the kth expert feature, where k is a positive integer;
[0134] The processing module 920 is further configured to, in response to the number of expert network layers being greater than k, perform a second feature processing on the k-th expert feature through the (k+1)-th expert network layer to obtain the (k+1)-th expert feature, until at least two expert features are obtained.
[0135] In some embodiments, the at least two expert network layers are arranged sequentially based on an increasing degree of generalization when processing the content features.
[0136] In some embodiments, the processing module 920 is further configured to perform a first feature processing on the content features through at least two task networks to obtain at least two task features, wherein the m-th task network corresponds to the m-th task feature, m is a positive integer, the at least two task networks are networks set up based on at least two recommendation tasks of the candidate content, and the m-th task network corresponds to the m-th recommendation task.
[0137] In some embodiments, the determining module 930 is further configured to determine the attention focus information corresponding to the m-th recommendation task and the at least two expert network layers based on the m-th task feature and the at least two expert features.
[0138] In some embodiments, the task network includes a first network based on click-through rate task settings;
[0139] like Figure 10 As shown, the recommendation module 940 includes:
[0140] Prediction unit 941 is used to predict the click-through rate of the candidate content in the click-through rate task based on the attention focus information of the at least two expert network layers corresponding to the click-through rate task, and obtain the click-through rate prediction result;
[0141] Recommendation unit 942 is used to recommend candidate content based on the click-through rate prediction results.
[0142] In some embodiments, the task network includes a second network based on conversion rate task settings;
[0143] The prediction unit 941 is further configured to predict the conversion rate of the candidate content in the conversion rate task based on the attention focus information of the at least two expert network layers corresponding to the conversion rate task, and obtain the conversion rate prediction result.
[0144] The recommendation unit 942 is also used to recommend the candidate content based on the click-through rate prediction result and the conversion rate prediction result.
[0145] In some embodiments, the determining module 930 is further configured to perform attention analysis on the task features and the expert features based on an attention mechanism to obtain attention focus information corresponding to the recommendation task and the at least two expert network layers.
[0146] In some embodiments, the determining module 930 is further configured to determine the matrix product of the task feature and the i-th expert feature;
[0147] The processing module 920 is further configured to perform embedding scaling processing on the matrix product to obtain a scaling matrix;
[0148] The determining module 930 is further configured to classify the scaling matrix and, based on the classification results corresponding to the at least two expert network layers, determine the attention focus information corresponding to the recommendation task and the at least two expert network layers.
[0149] In summary, the content recommendation device provided in this embodiment performs attention between the expression vector of the recommendation task and the expression vectors of all layers of the expert network layer, respectively, to obtain the attention focus of the recommendation task on each layer of the expert network layer. The focus result of the attention is used as the input of each subsequent prediction process to replace the simple gating method of sharing features among experts in MMoE, thereby achieving comprehensive focus on feature expression, improving the prediction accuracy of the multi-task model, and improving the recommendation accuracy of candidate content.
[0150] It should be noted that the content recommendation device provided in the above embodiments is only an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the content recommendation device provided in the above embodiments belongs to the same concept as the content recommendation method embodiments, and its specific implementation process can be found in the method embodiments, which will not be repeated here.
[0151] Figure 11 This application shows a schematic diagram of the structure of a computer device provided in an exemplary embodiment, which can be implemented as described above. Figure 1 The server shown is 120. Specifically:
[0152] Computer device 1100 includes a central processing unit (CPU) 1101, a system memory 1104 including random access memory (RAM) 1102 and read-only memory (ROM) 1103, and a system bus 1105 connecting the system memory 1104 and the CPU 1101. Computer device 1100 also includes a mass storage device 1106 for storing an operating system 1113, application programs 1114, and other program modules 1115.
[0153] Mass storage device 1106 is connected to central processing unit 1101 via a mass storage controller (not shown) connected to system bus 1105. Mass storage device 1106 and its associated computer-readable media provide non-volatile storage for computer device 1100. That is, mass storage device 1106 may include computer-readable media (not shown) such as hard disk or compact disc read-only memory (CD-ROM) drive.
[0154] Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 1104 and mass storage device 1106 described above can be collectively referred to as memory.
[0155] According to various embodiments of this application, the computer device 1100 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1100 can be connected to the network 1112 via the network interface unit 1111 connected to the system bus 1105, or the network interface unit 1111 can be used to connect to other types of networks or remote computer systems (not shown).
[0156] The aforementioned memory also includes one or more programs, which are stored in the memory and configured to be executed by the CPU.
[0157] Embodiments of this application also provide a computer device, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the content recommendation method provided in the above-described method embodiments.
[0158] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the content recommendation method provided in the above-described method embodiments.
[0159] Embodiments of this application also provide a computer program product or 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 any of the content recommendation methods described in the above embodiments.
[0160] Optionally, the computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. The random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM). The sequence numbers of the embodiments in this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.
[0161] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.
[0162] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A content recommendation method, characterized in that, The method includes: Obtain content features, which are features obtained by feature extraction from candidate content; The content features are processed by a task network to obtain task features, wherein the task network is a network set up for a recommendation task based on the candidate content. The content features are processed by at least two expert network layers that are sequentially arranged and connected in an adjacent manner to obtain at least two expert features. The i-th expert network layer corresponds to the i-th expert feature, where i is a positive integer. The at least two expert network layers are used to process the content features based on their respective extraction and summarization requirements. During the acquisition of the at least two expert features, the input of the upper expert network layer is the output of the lower adjacent expert network layer. Based on the task features and the expert features, determine the attention focus information corresponding to the recommendation task and the at least two expert network layers; The candidate content is recommended based on the attention focus information. The attention focus information is obtained by performing attention calculations on the task features and the expert features of all levels of the expert network layer. The attention focus information of all levels of the expert network layer is concatenated and used as the input of the prediction network model of the recommendation task to obtain the output result of the corresponding recommendation task.
2. The method according to claim 1, characterized in that, The second feature processing of the content features is performed through at least two expert network layers that are sequentially arranged and connected in an adjacent manner, resulting in at least two expert features, including: The content features are processed by the kth expert network layer to obtain the kth expert feature, where k is a positive integer. In response to the fact that the number of expert network layers is greater than k, the kth expert feature is processed by the (k+1)th expert network layer to obtain the (k+1)th expert feature, until at least two expert features are obtained.
3. The method according to claim 2, characterized in that, The at least two expert network layers are arranged sequentially based on the increasing degree of generalization required for extracting and summarizing the content features.
4. The method according to any one of claims 1 to 3, characterized in that, The first feature processing of the content features through the task network to obtain task features includes: The content features are processed by at least two task networks to obtain at least two task features, wherein the m-th task network corresponds to the m-th task feature, and m is a positive integer. The at least two task networks are networks set up based on at least two recommendation tasks of the candidate content, and the m-th task network corresponds to the m-th recommendation task.
5. The method according to claim 4, characterized in that, The step of determining the attention focus information corresponding to the recommendation task and the at least two expert network layers based on the task features and the expert features includes: Based on the m-th task feature and the at least two expert features, determine the attention focus information corresponding to the m-th recommendation task and the at least two expert network layers.
6. The method according to any one of claims 1 to 3, characterized in that, The task network includes a first network based on click-through rate task settings; The recommendation of candidate content based on the attention focus information includes: Based on the attention focus information of the at least two expert network layers corresponding to the click-through rate task, the click-through rate of the candidate content in the click-through rate task is predicted to obtain the click-through rate prediction result. The candidate content is recommended based on the click-through rate prediction results.
7. The method according to claim 6, characterized in that, The task network includes a second network based on conversion rate tasks; The recommendation of candidate content based on the attention focus information includes: Based on the attention focus information of the at least two expert network layers corresponding to the conversion rate task, the conversion rate of the candidate content in the conversion rate task is predicted to obtain the conversion rate prediction result. The candidate content is recommended based on the click-through rate prediction results and the conversion rate prediction results.
8. The method according to any one of claims 1 to 3, characterized in that, The step of determining the attention focus information corresponding to the recommendation task and the at least two expert network layers based on the task features and the expert features includes: Attention analysis is performed on the task features and expert features based on the attention mechanism to obtain the attention focus information corresponding to the recommendation task and the at least two expert network layers.
9. The method according to claim 8, characterized in that, The attention analysis of the task features and expert features based on the attention mechanism to obtain the attention focus information corresponding to the recommendation task and the at least two expert network layers includes: Determine the matrix product of the task feature and the i-th expert feature; The matrix product is then subjected to embedding scaling to obtain a scaling matrix; The scaling matrix is classified, and based on the classification results corresponding to the at least two expert network layers, the attention focus information corresponding to the recommendation task and the at least two expert network layers is determined.
10. A content recommendation device, characterized in that, The device includes: The acquisition module is used to acquire content features, which are features obtained by feature extraction of candidate content. The processing module is used to perform first feature processing on the content features through a task network to obtain task features, wherein the task network is a network set up based on the recommendation task of the candidate content; The processing module is further configured to perform a second feature processing on the content features through at least two expert network layers that are sequentially arranged and connected in a contiguous manner, to obtain at least two expert features, wherein the i-th expert network layer corresponds to the i-th expert feature, and i is a positive integer. The at least two expert network layers are used to perform feature processing on the content features based on corresponding extraction and summarization requirements. During the acquisition of the at least two expert features, the input of the upper expert network layer is the output of the lower adjacent expert network layer. The determination module is used to determine the attention focus information corresponding to the recommendation task and the at least two expert network layers based on the task features and the expert features; The recommendation module is used to recommend candidate content based on the attention focus information. The attention focus information is obtained by performing attention calculations on the task features and the expert features of all levels of the expert network layer. The attention focus information of all levels of the expert network layer is concatenated and used as the input of the prediction network model of the recommendation task to obtain the output result of the corresponding recommendation task.
11. The apparatus according to claim 10, characterized in that, The processing module is further configured to perform a second feature processing on the content feature through the kth expert network layer to obtain the kth expert feature, where k is a positive integer; The processing module is further configured to, in response to the number of expert network layers being greater than k, perform a second feature processing on the k-th expert feature through the (k+1)-th expert network layer to obtain the (k+1)-th expert feature, until at least two expert features are obtained.
12. The apparatus according to claim 11, characterized in that, The at least two expert network layers are arranged sequentially based on the increasing level of generalization when processing the content features.
13. The apparatus according to any one of claims 10 to 12, characterized in that, The processing module is also used for: The content features are processed by at least two task networks to obtain at least two task features, wherein the m-th task network corresponds to the m-th task feature, and m is a positive integer. The at least two task networks are networks set up based on at least two recommendation tasks of the candidate content, and the m-th task network corresponds to the m-th recommendation task.
14. The apparatus according to claim 13, characterized in that, The determining module is further configured to: Based on the m-th task feature and the at least two expert features, determine the attention focus information corresponding to the m-th recommendation task and the at least two expert network layers.
15. The apparatus according to any one of claims 10 to 12, characterized in that, The task network includes a first network based on click-through rate task settings; The recommendation module includes: The prediction unit is used to predict the click-through rate of the candidate content in the click-through rate task based on the attention focus information of the at least two expert network layers corresponding to the click-through rate task, and obtain the click-through rate prediction result. The recommendation unit is used to recommend candidate content based on the click-through rate prediction results.
16. The apparatus according to claim 15, characterized in that, The task network includes a second network based on conversion rate tasks; The prediction unit is further configured to predict the conversion rate of the candidate content in the conversion rate task based on the attention focus information of the at least two expert network layers corresponding to the conversion rate task, and obtain the conversion rate prediction result. The recommendation unit is also used to recommend the candidate content based on the click-through rate prediction result and the conversion rate prediction result.
17. The apparatus according to any one of claims 10 to 12, characterized in that, The determining module is further configured to: Attention analysis is performed on the task features and expert features based on the attention mechanism to obtain the attention focus information corresponding to the recommendation task and the at least two expert network layers.
18. The apparatus according to claim 17, characterized in that, The determining module is further configured to determine the matrix product of the task feature and the i-th expert feature; The processing module is also used to perform embedding scaling processing on the matrix product to obtain a scaling matrix; The determining module is further configured to classify the scaling matrix and, based on the classification results corresponding to the at least two expert network layers, determine the attention focus information corresponding to the recommendation task and the at least two expert network layers.
19. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one program, which is loaded and executed by the processor to implement the content recommendation method as described in any one of claims 1 to 9.
20. A computer-readable storage medium, characterized in that, The storage medium stores at least one program segment, which is loaded and executed by a processor to implement the content recommendation method as described in any one of claims 1 to 9.
21. A computer program product, characterized in that, The computer program product 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 content recommendation method as described in any one of claims 1 to 9.