A multi-domain multi-behavior adaptive news recommendation system and method

Abstract

The application relates to the field of news recommendation technology and natural language query, and discloses a multi-domain and multi-behavior adaptive news recommendation system and method. Through the double-layer mechanism of the BIPN network, invalid behavior noise such as accurate filtering of false clicks and quick removal can be filtered, the effective browsing prediction accuracy can be improved, the recommendation deviation caused by noise can be avoided, the behavior noise suppression effect is remarkable, through the multi-domain mixed expert mechanism, the preference of a certain domain can be seamlessly transferred to other domains, the bad experience caused by the fragmentation of multi-end recommendation can be avoided, the cross-domain preference transfer capability is strong, through the GCN enhancement layer, high-order neighbor correlation is realized, preference information is supplemented for new users and small theme news, the situation that a hot news monopolizes a recommendation list is broken, the individualization coverage rate is improved, the AutoML double-layer optimization does not need manual parameter adjustment, can automatically adapt to the domain distribution and behavior distribution of different news platforms, compared with manual parameter adjustment, the model generalization error is reduced, and the training efficiency is improved.

Description

Technical Field

[0001] This invention relates to news recommendation technology and natural language query, specifically to a multi-domain, multi-behavior adaptive news recommendation system and method. Background Technology

[0002] With the widespread adoption of mobile internet, users' news consumption scenarios are becoming increasingly fragmented. Users browse news content through multiple terminals (including mobile news apps, WeChat mini-programs, and in-depth report pages on PCs) and multiple interactive behaviors (including clicking, browsing, saving, and sharing). This multi-terminal switching and multi-behavioral interaction mode presents unique challenges to news recommendation systems. While existing technologies have made progress in single-domain recommendation or single-behavior modeling, they have significant shortcomings in multi-domain, multi-behavior collaborative scenarios. (1) Lack of cross-domain adaptability of single-domain multi-behavior model: Single-domain model can improve the accuracy of single-domain recommendation by filtering noise such as accidental clicks and fast swipes, but it cannot achieve cross-domain preference transfer; at the same time, the model weights rely on manual tuning for single domains, and when a new domain is added, it needs to be retrained, resulting in extremely poor generalization. (2) Behavioral noise and lack of text perception preference in multi-domain models: Although multi-domain recommendation methods can capture cross-domain commonalities through the shared layer, they are not optimized for the behavioral specificity of news texts. On the one hand, "misclicks" and "effective browsing" are regarded as equally effective signals, which leads the model to mistakenly regard noise as real preferences and recommend a large amount of content that users are not interested in. (3) Performance drop in sparse data scenarios: Cold start and topic sparsity are common problems in news recommendation. For example, the new user registration model cannot infer their preferences, and niche news topics are difficult to learn text features due to low interaction volume, resulting in the recommendation list being monopolized by hot news for a long time, and users' personalized needs cannot be met. Existing technologies cannot simultaneously solve the three core problems of cross-domain preference transfer, behavioral noise filtering, and data sparsity mitigation. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a multi-domain, multi-behavior adaptive news recommendation system and method, which has advantages such as improving the accuracy and relevance of news recommendations and solving the aforementioned technical problems.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a multi-domain, multi-behavior adaptive news recommendation system, comprising a feature acquisition module, a joint representation generation module, a multi-view expert module, a graph enhancement module, an adaptive prediction module, and a recommendation list generation module; The joint representation generation module generates a domain-behavior joint feature representation based on the target user features, candidate news features, target domain features, and target behavior features obtained by the feature acquisition module. The multi-view expert module generates a comprehensive preference representation based on the domain-behavior joint feature representation. The graph enhancement module performs high-order information enhancement on the comprehensive preference representation based on the trained graph product network to obtain the final preference representation; The adaptive prediction module generates predicted interaction probabilities based on the final preference representation; The recommendation list generation module generates a news recommendation list based on the predicted interaction probability.

[0005] As a preferred technical solution of the present invention, the joint representation generation module generates a domain-behavior joint feature representation based on the features of the target user, the features of the candidate news, the features of the target domain, and the features of the target behavior obtained by the feature acquisition module, including the following steps: Step A1: Through tensor product Integration of cross-domain shared weights Domain-specific weights Then through element-wise product Incorporating behavioral specific weights The weight fusion is completed, and the specific expression is: in, Indicates the fusion weights; Step A2: Original feature encoding and bias correction, through weight fusion The specific steps for encoding the concatenation of target user features and candidate news features are as follows: in, Indicates the characteristics of the target user, Indicate the characteristics of candidate news, This represents the concatenation of user characteristics and candidate news characteristics. Indicates the field bias, Indicates shared bias. Indicates behavioral bias. This represents the encoded domain-behavioral characteristics; Step A3: Construct a domain-independent mapping network From the original full scene features Extracting domain-behavior adaptive information, which is then added to the basic features as residual terms, and outputting the joint domain-behavior feature representation, the specific expression is as follows: in, Represents the uniform mapping weights, used to... Mapping the 2D dimension to the d dimension, b c For mapping bias, Representation of domain-behavior joint feature representation, This represents the original full-scene features, derived from the features of the target user. Characteristics of candidate news Target domain features and target behavioral characteristics It is composed of splicing parts.

[0006] As a preferred embodiment of the present invention, the domain-independent mapping network in step A3 It consists of two fully connected layers with ReLU activation function and 128 hidden layer dimensions.

[0007] As a preferred technical solution of the present invention, the multi-view expert module includes a shared expert module, a domain expert module, and a behavior expert module; The shared expert module outputs a public preference representation through a scene-adaptive gating fusion mechanism. The domain expert module extracts domain-specific preferences through the BIPN network, while simultaneously integrating information from other domains to supplement sparse data; The behavior expert module outputs text-aware preferences through a BIPN network that is consistent with but has independent parameters in the domain expert module; The multi-view expert module further includes weighted fusion of the outputs from the shared expert module, domain expert module, and behavior expert module to output a comprehensive preference representation, the specific expression of which is as follows: in, This is a comprehensive preference representation. For module fusion weights, and , For public preference representation, This represents a domain-specific preference representation. This represents a specific text preference.

[0008] As a preferred technical solution of the present invention, the shared expert module outputs a public preference representation through a scene-adaptive gating fusion mechanism, the specific expression of which is as follows: in, Indicates the weight of shared experts. This indicates the bias of shared experts. Represents the ReLU activation function. Representation layer normalization, This indicates shared expert embedding features. Indicates shared expert embedding weights. and For gating parameters, Indicates the number of shared experts. This represents the Softmax function. For public preference representation, It is the element-wise product.

[0009] As a preferred embodiment of the present invention, the domain expert module extracts domain-specific preferences through a BIPN network and simultaneously integrates information from other domains to supplement sparse data. The specific steps are as follows: Step B1: Establish the BIPN network, including a pre-filtering layer, a text-aware layer, and a post-filtering layer; The formula for the pre-filter layer is as follows: in, For the Sigmoid function, and For filtering parameters, Represents the behavioral embedding dimension. This represents the filtered user response. For element-wise product, Represents the user embedding vector. This represents the weight vector learned based on domain-behavior joint features and behavior embeddings; The formula for the text perception layer is as follows: in, This represents the activation function. and For sensing parameters, For text perception preferences, Indicates splicing, Represents the news embedding vector. The behavior represents a one-hot embedding vector; The formula for the post-filter layer is as follows: in, and These are post-filtering parameters. Let d represent the behavior-perceived preferences. Indicates splicing, The behavior represents a one-hot embedding vector. This represents the quadratic weight vector learned based on domain-behavior joint features and behavior embeddings; Step B2: Cross-domain information fusion, the formula is as follows: in, Indicates the current domain. Indicates a non-current domain. Indicates the current domain weight. The average weight of other domains, This represents the behavioral perception preference features in the current domain. This represents the behavioral perception preference features outside the current domain. Indicates the total number of fields.

[0010] As a preferred embodiment of the present invention, the behavior expert module outputs text-aware preferences through a BIPN network that is consistent with but has independent parameters in the domain expert module. Specifically, the behavior expert network is established, which is a BIPN network that is consistent with but has independent parameters in the domain expert module. The specific expression is as follows: in, This represents the activation function. Indicates splicing, For element-wise product, Represents the user embedding vector. Represents the news embedding vector. The behavior represents a one-hot embedding vector. These are the BIPN parameters specific to behavior b in the behavioral expert network. User preferences under the current behavior; The formula for cross-behavioral information fusion is as follows: Where b represents the current action, Indicates a non-current action. Indicates the current behavior weight. This represents user preferences under the current behavior. This represents user preferences not related to the current action. This indicates the total number of news items.

[0011] As a preferred embodiment of the present invention, the image enhancement module includes the following steps: Step C1: Construct a bipartite graph G=(V, E) with a node set V=U∪T, containing all user nodes and news nodes; Where U represents the user set, T represents the news set, E represents the edge set, and ∪ represents the union set, satisfying that if user u interacts with news a under any domain d and any action b, then there exists an edge (u, a) with an edge weight of 1. Step C2: Perform graph convolution using LightGCN, specifically by constructing an adjacency matrix. Then, normalized graph convolution is performed, followed by uniformly weighted fusion of two layers of GCN embedding, as shown in the following expression: in, For the first Layer node embedding, Indicates the first Layer node embedding, , These represent the number of users and the number of news items, respectively. This represents a diagonal matrix recording the degree of all nodes. To express summation, This indicates the node embedding after GCN fusion, including the user's GCN enhanced embedding. Enhanced GCN embedding with news ; Step C3: Enhance representation fusion, the specific expression is as follows: in, This represents the final preference representation. Indicates a linear layer. This indicates that GCN enhances weights. Indicates splicing, This represents a comprehensive preference.

[0012] As a preferred embodiment of the present invention, the adaptive prediction module generates predicted interaction probabilities based on the final preference representation, and the specific steps are as follows: Step D1: Define each hyperparameter as a 1-dimensional trainable tensor The formula is as follows: in, This represents the Sigmoid function. This represents the optimized hyperparameters. ; Step D2: A two-level optimization approach is adopted, first optimizing the model parameters and then optimizing the hyperparameters. The formula is as follows: in, Represents all trainable parameters. Indicates a given hyperparameter The optimal model parameters at that time, Indicates that given hyperparameters Under the premise of finding the optimal model parameters, such that the total loss function To obtain the minimum value, The total loss function is expressed as follows: in, Represents cross-entropy loss, This represents the Bayesian personalized ranking loss. This is represented as L2 regularization, with a regularization coefficient of 1e-3; The recommendation list generation module generates a news recommendation list based on the predicted interaction probability, as shown in the following expression: in, Indicates the predicted interaction probability. For domain-behavior specific prediction parameters, according to Sort in descending order and use the Top-K results to generate a recommendation column.

[0013] This invention also provides a multi-domain, multi-behavior adaptive news recommendation method, based on the above-mentioned multi-domain, multi-behavior adaptive news recommendation system, comprising the following steps: S1: Obtain the characteristics of the target user, the characteristics of the candidate news, the characteristics of the target domain, and the characteristics of the target behavior; S2: Generate a domain-behavior joint feature representation based on the characteristics of the target user, the characteristics of the candidate news, the characteristics of the target domain, and the characteristics of the target behavior; S3: Generate a comprehensive preference representation by processing the domain-behavior joint feature representation; S4: Based on the trained graph product network, perform high-order information augmentation on the comprehensive preference representation to obtain the final preference representation; S5: Generate predicted interaction probabilities based on the final preference representation, and generate a news recommendation list based on the predicted interaction probabilities.

[0014] Compared with existing technologies, this invention provides a multi-domain, multi-behavior adaptive news recommendation system and method, which has the following beneficial effects: 1. This invention utilizes a two-layer mechanism of the BIPN network to accurately filter out invalid behavioral noise such as accidental clicks and rapid scrolling. It improves the accuracy of effective browsing prediction, avoids recommendation bias caused by noise, and demonstrates significant behavioral noise suppression. Furthermore, through a multi-domain hybrid expert mechanism, preferences from one domain can be seamlessly transferred to other domains, avoiding the unpleasant experience of fragmented recommendations across multiple platforms, and exhibiting strong cross-domain preference transfer capabilities.

[0015] 2. This invention achieves high-order neighbor association through the GCN enhancement layer, supplementing new users and niche news topics with preference information, breaking the monopoly of hot news in the recommendation list, improving personalized coverage, and the AutoML two-layer optimization eliminates the need for manual parameter tuning. It can automatically adapt to the domain distribution and behavior distribution of different news platforms. Compared with manual parameter tuning, it reduces the model generalization error and improves training efficiency. Attached Figure Description

[0016] Figure 1 This is the architecture diagram of the MDMB-ANR model; Figure 2 This is a screenshot of the data cleaning results. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1-2 A multi-domain, multi-behavior adaptive news recommendation system includes a feature acquisition module, a joint representation generation module, a multi-view expert module, a graph enhancement module, an adaptive prediction module, and a recommendation list generation module. The joint representation generation module generates a domain-behavior joint feature representation based on the target user features, candidate news features, target domain features, and target behavior features obtained by the feature acquisition module. The joint representation generation module generates a domain-behavior joint feature representation based on the target user features, candidate news features, target domain features, and target behavior features obtained by the feature acquisition module. This includes the following steps: Step A1: Through tensor product Integration of cross-domain shared weights Domain-specific weights Then through element-wise product Incorporating behavioral specific weights The weight fusion is completed, and the specific expression is: in, Indicates the fusion weights; Step A2: Original feature encoding and bias correction, through weight fusion The specific steps for encoding the concatenation of target user features and candidate news features are as follows: in, Indicates the characteristics of the target user, Indicate the characteristics of candidate news, This represents the concatenation of the characteristics of the target user and the characteristics of the candidate news. Indicates the field bias, Indicates shared bias. Indicates behavioral bias. This represents the encoded domain-behavioral characteristics; Step A3: Construct a domain-independent mapping network (Constituted by a 2-layer fully connected network with ReLU activation function and 128 hidden layer dimensions), it maps features from different domains and behaviors to the same scale space, avoiding feature imbalance caused by differences in domain distribution (for example, the frequency of user interaction on the APP is 3 times that on the mini-program, so the feature range needs to be unified through mapping), and obtains features from the original full set of scene features. Extracting domain-behavior adaptive information, which is then added to the basic features as residual terms, and outputting the joint domain-behavior feature representation, the specific expression is as follows: in, Represents the uniform mapping weights, used to... Mapping the 2D dimension to the d dimension, b c For mapping bias, Representation of domain-behavior joint feature representation, This represents the original full-scene features, derived from the features of the target user. Characteristics of candidate news Target domain features and target behavioral characteristics The final output is composed of multiple parts. The joint basic representation of "user-news-domain-behavior" has initially captured the correlation characteristics between domain and behavior; The multi-view expert module processes domain-behavior joint feature representations to generate comprehensive preference representations; The multi-view expert module includes a shared expert module, a domain expert module, and a behavioral expert module. In news recommendation, user preferences can be broken down into three categories: 1) cross-domain and cross-behavior common preferences, such as users' long-term preferences for current affairs news, which exist across all terminals and behaviors; 2) domain-specific preferences, such as PC users preferring long-form in-depth reports, while APP users prefer short-form trending articles; 3) behavior-specific preferences, such as the "content depth" of news being considered in the "collection" behavior, and the "headline attractiveness" being considered in effective browsing. Meanwhile, there is a large amount of noise in different behaviors, such as accidental clicks, which needs to be filtered specifically. Therefore, this layer designs a multi-view structure of shared experts, domain experts, and behavioral experts, and each expert integrates a Behavioral Context Text Preference Network (BIPN) to achieve the dual functions of preference extraction and noise filtering. The shared expert module outputs a public preference representation through a scenario-adaptive gating fusion mechanism; Shared Expert Module (S): This module extracts common preferences across domains and behaviors, capturing user preferences shared by all terminals and behaviors, such as users' general interest in "local news" and "breaking political news." These preferences do not change with terminals or behaviors and need to be extracted uniformly by shared experts to avoid redundant learning across multiple domains.

[0019] 1) Establish an expert network: Each shared expert is a single-layer fully connected network. LayerNorm is added to avoid gradient explosion caused by differences in the distribution of news text features. ReLU activation function is added to introduce non-linearity and adapt to the complex relationship between topic and user. The formula is as follows: in, To determine the weights and biases of shared experts, the number of shared experts is adaptively determined based on the dataset size (e.g., the MIND dataset is set to 2, focusing on two common preferences: "hot topics" and "regional associations" respectively).

[0020] 2) Incorporate a scenario-adaptive gating fusion mechanism: This mechanism assigns expert weights to different combinations of "user-news-domain-behavior" by introducing gating. For example, when a user browses "local breaking news," the "regional association" expert is activated; when browsing "national hot news," the "hot topic" expert is activated, avoiding interference from ineffective experts. The formula is as follows: in, Represents the ReLU activation function. Representation layer normalization, This indicates shared expert embedding features. Indicates shared expert embedding weights. and For gating parameters, Indicates the number of shared experts. This represents the Softmax function. This indicates public preference. The final output is the element-wise product. This represents public preferences, such as a general preference vector for users regarding "local political news in Beijing"; The domain expert module extracts domain-specific preferences through the BIPN network, while also incorporating information from other domains to supplement sparse data; Domain Expert Module (D): Domain-specific preferences are extracted, while information from other domains is integrated to supplement sparse data. For example, user interaction data for "embedded information pages within WeChat official accounts" is scarce; browsing preferences from the app can be used to assist in modeling, avoiding insufficient preference learning caused by data sparsity within the domain. Each domain expert integrates a BIPN network to filter behavioral noise within that domain.

[0021] 1) Establishing the BIPN network: The BIPN network uses a three-layer structure of "pre-filtering - text awareness - post-filtering" to accurately extract text preferences under domain-behavior, while filtering out invalid noise. The pre-filtering layer filters out information embedded in the user profile that is irrelevant to the current user action, based on the "user-news-behavior" feature. For example, accidental clicks within the app. To filter user preferences for "long-form in-depth reports" and retain only those related to "short-form headlines," the filtering formula is as follows: in, For the Sigmoid function, (|B| represents the behavioral embedding dimension) For filtering parameters, This is the filtered user representation.

[0022] The text perception layer captures users' specific text preferences for the current news. For example, user u's "headline attractiveness rating" and "topic relevance" for news a directly determine whether the user will take effective action, and the formula is as follows: Where tanh is the activation function. , For sensing parameters, Text perception preference.

[0023] The post-filtering layer further filters out noise in text perception preferences. For example, outliers with "high topic relevance" caused by accidental clicks (a user accidentally clicks on sports news but actually prefers technology) have their weight reduced by the post-filtering layer, as shown in the following formula: in, and These are post-filtering parameters. For text-aware preference representation under domain d Indicates splicing, The behavior represents a one-hot embedding vector. This represents the quadratic weight vector learned based on domain-behavior joint features and behavior embeddings. This represents the text-aware preference representation (i.e., domain expert output) under domain d.

[0024] 2) Cross-domain information fusion: To avoid domain experts relying too heavily on single-domain data (such as sparse data in embedded pages of WeChat official accounts), domain fusion weights are introduced. This integrates supplementary information from expert outputs in other domains while ensuring dominance in the current domain. The formula is as follows: in, This indicates a domain other than the current domain (e.g., the current domain is an embedded page within a WeChat Official Account). (For APP, mini-program, PC) The average weight of other domains, Indicates the current domain weight. This represents the total number of fields, ensuring that the total weight sums to 1. For example, (Mini Program) indicates that BIPN output accounts for 60% on the Mini Program side, while the output of the APP and PC sides together account for 40%. It retains the short text preference of the Mini Program side and also supplements the topic preferences of other domains. The behavioral expert module outputs text-aware preferences through a BIPN network that is consistent with but parameter-independent in the domain expert module; Behavioral Expert Module (B) Capture the specific text preferences of individual behaviors. For example, effective browsing behavior focuses on the "headline appeal" and "lead completeness" of news articles, while saving behavior focuses on "content depth" and "long-term value," and sharing behavior focuses on "social dissemination." At the same time, integrate information from other behaviors to improve generalization (e.g., use the topic preferences of "effective browsing" to assist in predicting "saving" behavior).

[0025] 1) Establish a behavioral expert network: The structure of the behavioral expert network is completely consistent with that of the domain expert BIPN, but the parameters are independent. It outputs the text-aware preference under behavior b, and its formula is as follows: in, BIPN parameters are specific to behavior b. For example, the BIPN for the "favorite" behavior will increase the weight of the "content depth" feature, while the BIPN for the "effective browsing" behavior will increase the weight of the "title attractiveness" feature. Represents the behavioral embedding dimension. .

[0026] 3) Cross-behavioral information fusion: Introducing behavioral fusion weights It integrates supplementary information from expert outputs on other behaviors to balance the correlation between the current behavior and other behaviors. The formula is as follows: For example, BIPN outputs indicating valid views account for 70%, while outputs from accidental clicks and bookmarks account for 30% combined. Indicates the current behavior weight. This indicates the total number of news items. This represents user preferences under the current behavior. This represents user preferences under conditions other than the current action. The multi-view expert module also includes a weighted fusion of the outputs from the shared expert module, domain expert module, and behavior expert module to output a comprehensive preference representation, as shown in the following expression: in, This represents a comprehensive preference expression. For module fusion weights, and , This indicates public preference. This represents a domain-specific preference representation. This represents a specific text preference. The graph augmentation module enhances the comprehensive preference representation with higher-order information based on the trained graph product network, resulting in the final preference representation. Data sparsity is a core pain point in news recommendation, and the preference representation output by the expert layer may be biased (e.g., inferring user preferences based on only one accidental click). This layer constructs a unified user-news interaction graph that encompasses the entire domain and all behaviors, and uses GCN to extract higher-order neighbor relationships (e.g., "User A → News X → User B": User A accidentally clicks on News X, and User B saves News X, suggesting that User A may also be interested in the topic of News X). This supplements the preference information under sparse data and enhances the robustness of the final representation. Step C1: Construct a unified interaction graph; Integrate user-news interaction data from all domains and all valid behaviors (valid browsing, bookmarking) to construct a bipartite graph G=(V, E). The node set V=U∪T contains all user nodes and news nodes; the edge set E exists if user u interacts with news a in any domain d and any behavior b, with an edge weight of 1.

[0027] Step C2: LightGCN higher-order information aggregation; LightGCN is used for graph convolution because the association of features in news text is mainly based on topic matching, which does not require complex nonlinear transformations. LightGCN's linear aggregation is more efficient and less prone to overfitting. The specific implementation steps are as follows: 1) Construct the adjacency matrix: ; in, For user-news interaction matrix ( R u，a =1 indicates a valid interaction; otherwise, it is 0. Let be the adjacency matrix of the bipartite graph. express The transpose of .

[0028] 2) Normalization and Higher-Order Aggregation: To avoid embedding bias caused by differences in node degree (e.g., the interaction frequency of trending news is tens of times higher than that of niche news), normalization is needed to balance the impact. A matrix D is introduced. in, v This represents the target node whose degree is to be calculated; Indicates the neighboring nodes being traversed; express and If there is an edge between them, the value is 1; otherwise, it is 0. express The degree, statistics How many edges are there, and how many normalized graph convolutions are performed? in, For the first Layer node embedding ( l =0 indicates the initial user embedding. e u News Embedding e a ), , These represent the number of users and the number of news items, respectively. Let represent a diagonal matrix recording the degree of all nodes, and .

[0029] 3) Multi-layer embedding fusion: A uniformly weighted fusion 2-layer GCN embedding is adopted (2 layers can capture the second-order association of "user-news-user", which is sufficient to cover the topic association requirements of news recommendation): in, For the first Layer node embedding, , These represent the number of users and the number of news items, respectively. To express summation, This indicates the embedding of nodes after GCN fusion. This includes the user's GCN enhanced embedding. Enhanced GCN embedding with news , The weights are layered (shallow embeddings emphasize direct interaction, while deeper embeddings emphasize global topic associations), ultimately... E gcn Includes user-enhanced GCN embedding (Incorporating the preferences of similar users) and enhanced GCN embedding with news (It incorporates themes from similar news stories).

[0030] Step C3: Enhanced Representation Fusion The GCN is introduced to enhance the weight λ, balancing the "personalized preferences" of the expert layer with the global topic association of the GCN. That is, when the data is sparse, it relies on global association, and when the data is abundant, it relies on personalized preferences. The formula is as follows: Among them, Linear( The first layer is a linear layer, which mainly maps the stitched 2D dimension to the d dimension, consistent with the output dimension of the expert layer. This represents the final preference representation based on "personalization + global collaboration." For example, if a new user u has only accidentally clicked on one tech news article, GCN can supplement user u's tech topic preferences through the association of "user u → tech news X → user B (who has saved news X and prefers tech)" to avoid recommendation bias. The adaptive prediction module generates predicted interaction probabilities based on the final preference representation; The adaptive prediction module generates predicted interaction probabilities based on the final preference representation. The specific steps are as follows: Step D1: Define each hyperparameter as a 1-dimensional trainable tensor The formula is as follows: in, This represents the Sigmoid function. Indicates the optimized hyperparameters, subscript ; Step D2: A two-level optimization approach is adopted, first optimizing the model parameters and then optimizing the hyperparameters. The formula is as follows: in, Represents all trainable parameters. Indicates a given hyperparameter w The optimal model parameters at that time, Indicates that given hyperparameters Under the premise of finding the optimal model parameters, such that the total loss function To obtain the minimum value, The total loss function is expressed as follows: in, Represents cross-entropy loss, This represents the Bayesian personalized ranking loss. This is represented as L2 regularization, with a regularization coefficient of 1e-3; The cross-entropy loss primarily optimizes the accuracy of behavior prediction, determining whether a user will have a valid interaction with news item a under "domain d - behavior b", such as valid browsing or saving. in, y For behavior labels (y=1 indicates a valid interaction, y=0 indicates an invalid interaction such as a mistaken click), The inner product of expert-level output and news embedding reflects the user's degree of news preference. This is the sample set for training data.

[0031] r The Bayesian personalized ranking loss primarily optimizes the ranking quality of the recommendation list, ensuring that news preferred by users appears at the top of the information feed. in, ,in, O The training triplet observation set consists of "user u + positive sample news a (news with effective user interaction) + negative sample news c (samples with no effective user interaction)". R + For users u News collection R For news collection, The predicted score of user u’s preference for positive sample a under domain d and behavior b; The predicted preference score of user u for negative sample c under domain d and behavior b.

[0032] The recommendation list generation module generates a news recommendation list based on the predicted interaction probability. For different "domain-behavior" combinations, the probability of user-news interaction is output to generate a personalized recommendation list, and the formula is as follows: in, These are prediction parameters specific to "domain d-behavior b", for example, prediction parameters for effective browsing on the APP. W d1，b2 It will emphasize the weight of "title features" and predict parameters for PC-based bookmarking. W d3，b3 The weighting will be emphasized for "content depth". The final ranking will be based on... Sort in descending order, then select the Top-K to generate a recommendation list. Indicates the predicted interaction probability; This invention also provides a multi-domain, multi-behavior adaptive news recommendation method, based on the above-mentioned multi-domain, multi-behavior adaptive news recommendation system, comprising the following steps: S1: Obtain the characteristics of the target user, the characteristics of the candidate news, the characteristics of the target domain, and the characteristics of the target behavior; S2: Generate a domain-behavior joint feature representation based on the characteristics of the target user, the characteristics of the candidate news, the characteristics of the target domain, and the characteristics of the target behavior; S3: Generate a comprehensive preference representation by processing the domain-behavior joint feature representation; S4: Based on the trained graph product network, perform high-order information augmentation on the comprehensive preference representation to obtain the final preference representation; S5: Generate predicted interaction probabilities based on the final preference representation, and generate a news recommendation list based on the predicted interaction probabilities.

[0033] Example: (a) Data preprocessing (taking the MIND dataset as an example): The MIND dataset is a publicly released news recommendation dataset from Microsoft, containing 6040 users, 3706 news articles, and 1 million user-news interaction records. It covers two domains: "mobile apps" and "WeChat mini-programs," and three behaviors: "accidental clicks," "valid browsing," and "favorites." It is a standard dataset for validating multi-domain, multi-behavior news recommendation models. The data preprocessing workflow is as follows: (1) Data cleaning: The first step is to remove duplicate interaction records between users, news, domains, and behaviors to avoid duplicate data interfering with preference learning.

[0034] The second step is outlier filtering: 1) Delete users with less than 2 valid interactions, filtering a total of 1203 users, leaving 4837 valid users; 2) Delete news articles with less than 5 valid interactions, filtering a total of 892 news articles, leaving 2814 valid news articles; 3) Delete records where users clicked more than 15 times in 1 minute, filtering a total of 52,000 records, leaving 948,000 valid interactions.

[0035] The third step, missing value imputation, involves: 1) imputing with the average age of users in the same region (if a user's age is missing, use the average age of all users); 2) using the NaiveBayes text classification model to infer the topic based on the news title and summary; and 3) inferring imputation based on interaction time and text length. A screenshot of the data cleaning results is shown below: (2) Feature construction: A fusion feature vector with a total dimension of 80 is constructed, including: 32-dimensional user features, 32-dimensional news features, 8-dimensional domain features, and 8-dimensional behavioral features. The specific feature allocation is shown in Tables 2-5 below. Table 2 User Feature Allocation Table Table 3 News Feature Allocation Table Table 4 Domain Feature Allocation Table Table 5. Behavioral Characteristic Allocation Table (3) Data partitioning: A combination of time-series partitioning and leave-one-out method is used to ensure that the validation and test sets meet the real-time requirements of news recommendation. The data partitioning method is as follows: Training set (80%): User interaction data from January to October 2023, covering non-hot topics and regular topics, to learn long-term preferences; Validation set (10%): User interaction data in November 2023, including recent hot topics, used to fine-tune the model's timeliness adaptation; Test set (10%): User interaction data in December 2023, including the latest trending topics, to verify the model's real-time recommendation capabilities; Leave-one-out validation: For each user in the test set, retain the news corresponding to their last valid browsing as the news to be recommended, and the remaining news as candidate news, and evaluate whether the model can rank the news to be recommended in the Top-K list.

[0036] (II) Model parameter settings: Based on the characteristics of the MIND dataset, set the following basic parameters, and AutoML will automatically optimize the hyperparameters. The parameter list is shown in Table 6 below: Table 6 Model Parameter Settings (III) Model Training Process: The model training employs a three-stage strategy of pre-training, joint training, and fine-tuning to ensure stable convergence and sufficient learning of the domain-behavior association features of the news scene. The three-stage strategy is as follows: (1) Phase 1 pre-training for more than 20 rounds: The main goal is to allow the domain-behavior joint representation extraction layer to learn the basic association features of "domain-behavior" first, so as to avoid oscillations in parameters during subsequent joint training due to insufficient learning of domain-behavior characteristics.

[0037] Parameter freezing: Freeze the parameters of the behavior-aware multi-view expert layer (shared / domain / behavior expert) and the AutoML layer, and retain only the parameters of the domain-behavior joint representation extraction layer for training. Optimization objective: Optimize only the cross-entropy loss. L bce It focuses on matching the basic features of "domain-behavior-news"; Training details: The learning rate is set to 1.5e-3 to avoid excessively rapid parameter updates during the pre-training phase; evaluation is performed on the validation set every 5 rounds. L bce Save the model with the minimum loss as the pre-training weights to ensure that the basic features are learned sufficiently.

[0038] (2) More than 60 rounds of joint training in Phase 2: The parameters of all layers are jointly optimized to learn the collaborative dependencies between shared preferences, domain-specific preferences, and behavior-specific preferences.

[0039] Parameter Unfreezing: Unfreeze all parameters of the behavior-aware multi-view expert layer and the AutoML layer; Optimization objective: Optimize total loss ; Two-layer optimization process: 1) Fixed hyperparameter tensors e w Training model parameters W 1 epoch: Using the Adam optimizer with a learning rate of 3e-3, calculate the total loss. L And backpropagate updates W ; 2) Fixed model parameters W Take one mini-batch of data, with a size of 1024, and calculate the hyperparameters. w Gradient: Calculated using the chain rule L right e w The derivative is updated using the learning rate 3e-4.e w ; 3) Repeat steps 1-2 above, evaluating HR@10 and NDCG@20 on the validation set every 2 rounds, and saving the model weight with the highest NDCG@20.

[0040] (3) Stage 3 fine-tuning for more than 20 rounds: To address the timeliness characteristics of the test set, we fine-tuned the hyperparameters and prediction layer parameters to ensure the model can adapt to the latest hot topics and improve real-time recommendation performance.

[0041] Parameter freezing: Freeze the parameters of the domain-behavior joint representation extraction layer, GCN enhancement layer, and expert layer to prevent the destruction of general features; Optimization goal: Optimize only the hyperparameters of the AutoML layer. w With the prediction layer W d，b , b d，b The optimization objective remains the total loss. L ; Training details: The learning rate is set to 3e-4 to avoid drastic parameter changes; HR@10 is evaluated on the validation set every 2 rounds, and the model with the highest HR@10 is saved as the final model.

[0042] (iv) Performance verification: To verify the superiority of MDMB-ANR in news recommendation scenarios, four representative baseline models were selected and compared on the MIND dataset. The names and characteristics of the four models are shown in Table 7 below: Table 7 List of Baseline Models HR@K (hit rate, reflecting effective reach) and NDCG@K (normalized depreciation cumulative gain, reflecting ranking quality) were used as the core evaluation indicators. All results are the average of 5 replicate experiments. The validation results are shown in Table 8 below: Table 8. Model Validation Results Experimental results show that the MDMB-ANR model of this invention successfully solves the shortcomings of previous models in cross-domain preference transfer and behavioral noise filtering, thus achieving better recommendation results than other models in multi-domain and multi-behavioral news recommendation scenarios.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-domain, multi-behavior adaptive news recommendation system, characterized in that: It includes a feature acquisition module, a joint representation generation module, a multi-view expert module, a graph enhancement module, an adaptive prediction module, and a recommendation list generation module; The joint representation generation module generates a domain-behavior joint feature representation based on the target user features, candidate news features, target domain features, and target behavior features obtained by the feature acquisition module. The multi-view expert module generates a comprehensive preference representation based on the domain-behavior joint feature representation. The graph enhancement module performs high-order information enhancement on the comprehensive preference representation based on the trained graph product network to obtain the final preference representation; The adaptive prediction module generates predicted interaction probabilities based on the final preference representation; The recommendation list generation module generates a news recommendation list based on the predicted interaction probability.

2. The multi-domain, multi-behavior adaptive news recommendation system according to claim 1, characterized in that: The joint representation generation module generates a domain-behavior joint feature representation based on the target user features, candidate news features, target domain features, and target behavior features obtained by the feature acquisition module, including the following steps: Step A1: Through tensor product Integration of cross-domain shared weights Domain-specific weights Then through element-wise product Incorporating behavioral specific weights The weight fusion is completed, and the specific expression is: in, Indicates the fusion weights; Step A2: Original feature encoding and bias correction, through weight fusion The specific steps for encoding the concatenation of target user features and candidate news features are as follows: in, Indicates the characteristics of the target user, Indicate the characteristics of candidate news, This represents the concatenation of user characteristics and candidate news characteristics. Indicates the field bias, Indicates shared bias. Indicates behavioral bias. This represents the encoded domain-behavioral characteristics; Step A3: Construct a domain-independent mapping network From the original full scene features Extracting domain-behavior adaptive information, which is then added to the basic features as residual terms, and outputting the joint domain-behavior feature representation, the specific expression is as follows: in, Represents the uniform mapping weights, used to... Mapping the 2D dimension to the d dimension, b c For mapping bias, Representation of domain-behavior joint feature representation, This represents the original full-scene features, derived from the features of the target user. Characteristics of candidate news Target domain features and target behavioral characteristics It is composed of splicing parts.

3. The multi-domain, multi-behavior adaptive news recommendation system according to claim 2, characterized in that: In step A3, the domain-independent mapping network It consists of two fully connected layers with ReLU activation function and 128 hidden layer dimensions.

4. The multi-domain, multi-behavior adaptive news recommendation system according to claim 2, characterized in that: The multi-view expert module includes a shared expert module, a domain expert module, and a behavior expert module; The shared expert module outputs a public preference representation through a scene-adaptive gating fusion mechanism. The domain expert module extracts domain-specific preferences through the BIPN network, while simultaneously integrating information from other domains to supplement sparse data; The behavior expert module outputs text-aware preferences through a BIPN network that is consistent with but has independent parameters in the domain expert module; The multi-view expert module further includes weighted fusion of the outputs from the shared expert module, domain expert module, and behavior expert module to output a comprehensive preference representation, the specific expression of which is as follows: in, This is a comprehensive preference representation. For module fusion weights, and , For public preference representation, This represents a domain-specific preference representation. This represents a specific text preference.

5. The multi-domain, multi-behavior adaptive news recommendation system according to claim 4, characterized in that: The shared expert module outputs a common preference representation through a scene-adaptive gating fusion mechanism, the specific expression of which is as follows: in, Indicates the weight of shared experts. This indicates the bias of shared experts. Represents the ReLU activation function. Representation layer normalization, This indicates shared expert embedding features. Indicates shared expert embedding weights. and For gating parameters, Indicates the number of shared experts. This represents the Softmax function. For public preference representation, It is the element-wise product.

6. The multi-domain, multi-behavior adaptive news recommendation system according to claim 5, characterized in that: The domain expert module extracts domain-specific preferences through the BIPN network, and simultaneously integrates information from other domains to supplement the sparse data. The specific steps are as follows: Step B1: Establish the BIPN network, including a pre-filtering layer, a text-aware layer, and a post-filtering layer; The formula for the pre-filter layer is as follows: in, For the Sigmoid function, and For filtering parameters, Represents the behavioral embedding dimension. This represents the filtered user response. For element-wise product, Represents the user embedding vector. This represents the weight vector learned based on domain-behavior joint features and behavior embeddings; The formula for the text perception layer is as follows: in, This represents the activation function. and For sensing parameters, For text perception preferences, Indicates splicing, Represents the news embedding vector. The behavior represents a one-hot embedding vector; The formula for the post-filter layer is as follows: in, and These are post-filtering parameters. Let d represent the behavior-perceived preferences. Indicates splicing, The behavior represents a one-hot embedding vector. This represents the quadratic weight vector learned based on domain-behavior joint features and behavior embeddings; Step B2: Cross-domain information fusion, the formula is as follows: in, Indicates the current domain. Indicates a non-current domain. Indicates the current domain weight. The average weight of other domains, This represents the behavioral perception preference features in the current domain. This represents the behavioral perception preference features outside the current domain. Indicates the total number of fields.

7. The multi-domain, multi-behavior adaptive news recommendation system according to claim 6, characterized in that: The behavior expert module outputs text-aware preferences through a BIPN network that is consistent with but parameter-independent in the domain expert module. Specifically, this involves establishing a behavior expert network, which is a BIPN network consistent with but parameter-independent in the domain expert module. The specific expression is as follows: in, This represents the activation function. Indicates splicing, For element-wise product, Represents the user embedding vector. Represents the news embedding vector. The behavior represents a one-hot embedding vector. These are the BIPN parameters specific to behavior b in the behavioral expert network. User preferences under the current behavior; The formula for cross-behavioral information fusion is as follows: Where b represents the current action, Indicates a non-current action. Indicates the current behavior weight. This represents user preferences under the current behavior. This represents user preferences not related to the current action. This indicates the total number of news items.

8. The multi-domain, multi-behavior adaptive news recommendation system according to claim 4, characterized in that: The graph enhancement module includes the following steps: Step C1: Construct a bipartite graph G=(V, E) with a node set V=U∪T, containing all user nodes and news nodes; Where U represents the user set, T represents the news set, E represents the edge set, and ∪ represents the union set, satisfying that if user u interacts with news a under any domain d and any action b, then there exists an edge (u, a) with an edge weight of 1. Step C2: Perform graph convolution using LightGCN, specifically by constructing an adjacency matrix. Then, normalized graph convolution is performed, followed by uniformly weighted fusion of two layers of GCN embedding, as shown in the following expression: in, For the first Layer node embedding, Indicates the first Layer node embedding, , These represent the number of users and the number of news items, respectively. This represents a diagonal matrix recording the degree of all nodes. To express summation, This indicates the node embedding after GCN fusion, including the user's GCN enhanced embedding. Enhanced GCN embedding with news ; Step C3: Enhance representation fusion, the specific expression is as follows: in, This represents the final preference representation. Indicates a linear layer. This indicates that GCN enhances weights. Indicates splicing, This represents a comprehensive preference.

9. The multi-domain, multi-behavior adaptive news recommendation system according to claim 8, characterized in that: The adaptive prediction module generates predicted interaction probabilities based on the final preference representation, and the specific steps are as follows: Step D1: Define each hyperparameter as a 1-dimensional trainable tensor The formula is as follows: in, This represents the Sigmoid function. This represents the optimized hyperparameters. ; Step D2: A two-level optimization approach is adopted, first optimizing the model parameters and then optimizing the hyperparameters. The formula is as follows: in, Represents all trainable parameters. Indicates a given hyperparameter The optimal model parameters at that time, Indicates that given hyperparameters Under the premise of finding the optimal model parameters, such that the total loss function To obtain the minimum value, The total loss function is expressed as follows: in, Represents cross-entropy loss, This represents the Bayesian personalized ranking loss. This is represented as L2 regularization, with a regularization coefficient of 1e-3; The recommendation list generation module generates a news recommendation list based on the predicted interaction probability, as shown in the following expression: in, Indicates the predicted interaction probability. For domain-behavior specific prediction parameters, according to Sort in descending order and use the Top-K results to generate a recommendation column.

10. A multi-domain, multi-behavior adaptive news recommendation method, based on the multi-domain, multi-behavior adaptive news recommendation system described in any one of claims 1-9, characterized in that: The process includes the following steps: S1: Obtain the characteristics of the target user, the characteristics of the candidate news, the characteristics of the target domain, and the characteristics of the target behavior; S2: Generate a domain-behavior joint feature representation based on the characteristics of the target user, the characteristics of the candidate news, the characteristics of the target domain, and the characteristics of the target behavior; S3: Generate a comprehensive preference representation by processing the domain-behavior joint feature representation; S4: Based on the trained graph product network, perform high-order information augmentation on the comprehensive preference representation to obtain the final preference representation; S5: Generate predicted interaction probabilities based on the final preference representation, and generate a news recommendation list based on the predicted interaction probabilities.

Images