Multi-interest news recommendation method and system based on popularity perception timeliness difference

Abstract

The invention discloses a multi-interest news recommendation method and system based on popularity perception timeliness difference. The method comprises the following steps: respectively extracting global and local semantic features of a news title; aggregating expressions of user historical news in each semantic dimension through a self-attention network to form a multi-interest feature; fusing the popularity information of the words and the entities, and generating popularity scores of the news through a multi-layer self-attention mechanism and a cross attention mechanism; converting the news existence duration into an embedded vector, and mapping the embedded vector into a time score through an MLP (Markup Language Protocol); mapping the popularity score and the time score of the news to a K-dimensional semantic channel through nonlinear transformation, and fusing the popularity score and the time score for adaptive weighting to realize dynamic modulation of popularity and time decay; correlation between user interest and news semantics is calculated through cross-channel attention, and matching scores of all channels are aggregated to obtain click prediction scores; and carrying out model training. According to the invention, the perception ability and recommendation accuracy of the recommendation system to the hot content are improved.

Description

Technical Field

[0001] This invention belongs to the field of news recommendation technology, and in particular relates to a multi-interest news recommendation method and system based on the difference in popularity perception and timeliness. Background Technology

[0002] With the rapid development of online news platforms and the explosive growth of information, news recommendation systems have become a key technology for improving the efficiency of users' information acquisition. However, early neural network news recommendation methods typically modeled user interests and news content as separate single feature vectors and made recommendations based on the similarity between vectors. This approach faces significant limitations in practical applications. Specifically, on the one hand, user interests are inherently multifaceted; for example, the same user may simultaneously follow multiple fields such as sports, technology, and politics, and a single vector cannot fully capture their diverse interest structure. On the other hand, news content itself encompasses multi-faceted semantic information, such as event background, social impact, and industry dynamics; using only a single vector representation inevitably loses these rich semantic layers. Therefore, such methods are clearly insufficient in capturing users' diverse interests and preferences and the multi-dimensional semantic connotations of news, ultimately limiting the accuracy and timeliness of recommendation results.

[0003] To overcome the limitations of single-vector representation methods, recent work has actively explored multi-interest news recommendation methods based on decoupled representations. The core idea of ​​these innovative methods is to decompose the representations of users and news into multiple relatively independent dimensions to better capture their multifaceted characteristics. For example, the MINS model designs a multi-channel recurrent network based on GRU, where each channel models the dependencies in the news sequence for each interest; the MCCM model focuses on learning multiple representations of words or entire news articles, thereby improving the ability to characterize news semantics; and the DRUIC model attempts to decouple representations simultaneously on the user and news sides, thus compensating to some extent for the shortcomings of single-end decoupling. These methods have advanced the development of multi-interest modeling techniques from different perspectives, providing multiple feasible technical paths to overcome the deficiencies of single-vector representations.

[0004] However, the multi-interest or multi-semantic representations obtained by the above methods are still static embeddings, failing to fully consider the timeliness differences of different semantic dimensions of news and the dynamic evolution of user interests over time. Therefore, when faced with the rapid decline in the value of trending news or the phased shift of user interests, the recommendation effect of such methods is significantly limited. Specifically, for example, the click potential of a sports event news item on the day of the match is much higher than several days later; the spread of a social hot news item will rise sharply in the early stages of the event and then quickly subside. When certain semantic dimensions of the news highly match the current social focus or the user's immediate interest, related keywords are very likely to stimulate the user's click behavior; however, as the event's popularity fades or the focus of interest shifts, even if the news content remains unchanged, its attractiveness will decrease significantly. Existing static representations cannot adapt to such dynamic fluctuations in semantic value and real-time changes in interest weights, making it difficult to achieve accurate and timely recommendations in practical application scenarios. Summary of the Invention

[0005] The purpose of this invention is to address a key deficiency in existing news recommendation technologies: the static nature of user interest representation and news semantic embedding. Existing methods fail to effectively capture the temporal decay patterns of news value and real-time popularity changes, and are also ill-suited to the dynamic evolution of user interests. This invention proposes a multi-interest news recommendation method and system based on popularity-aware timeliness differences. By designing a dynamic semantic modulation mechanism, this invention deeply fuses the temporal decay characteristics of news with real-time popularity signals, generating a dynamic news representation with time awareness. Simultaneously, a multi-channel interactive network is constructed to achieve fine-grained matching between multiple user interests and multiple news semantics. This technical solution significantly improves the adaptability of the recommendation system to dynamic environments, greatly enhancing the timeliness of recommendation results while maintaining accuracy. Experimental results on public datasets demonstrate that this method outperforms existing mainstream recommendation models in multiple evaluation metrics.

[0006] To achieve the objectives of this invention, this invention provides a multi-interest news recommendation method based on popularity-aware timeliness differences, the method comprising:

[0007] Step 1: Extract global and local semantic features of news headlines using multi-head self-attention and convolutional neural networks respectively, and then concatenate and fuse them into a multi-semantic representation in each semantic channel;

[0008] Word sequence of candidate news headlines Performing multi-dimensional semantic parsing involves the following processing steps: First, a global semantic model of the word sequence is performed using a multi-head self-attention mechanism to generate a global representation containing K semantic channels. Each attention head focuses on capturing different aspects of global semantic features; simultaneously, a convolutional neural network extracts local semantic features from the word sequence, using a convolutional kernel with a window size of w to scan the word sequence and generate corresponding local semantic representations. Finally, the global features of each semantic channel k are... With local features By splicing and merging the data, a complete multi-semantic representation of the news can be formed. Each of the channels The corresponding news is a complete representation on a specific latent semantic dimension; where K is set to 5 and w is set to 3.

[0009] Step 2: Aggregate the representations of users' historical news across various semantic dimensions using a self-attention network to form multi-interest features;

[0010] Historical click sequences of the target user First, using the news decoupling method in step 1, each news item in the sequence is decoupled. Generate decoupled representations corresponding to K latent semantic dimensions. Based on this, for each specific semantic dimension k, k=1,2,...,K, the decoupled representation of all historical news under the corresponding dimension is obtained. As input, a self-attention network is used to model the intrinsic relationships between different news items in their corresponding semantic dimensions. The self-attention mechanism calculates the relevance weights between news representations within a sequence to highlight important news items, and then obtains the user's interest representation in dimension k through weighted aggregation. Finally, the interest representations of the K semantic dimensions are integrated to form a complete multi-interest vector for the user. Each dimension This corresponds to capturing users' interests and preferences in specific semantic aspects.

[0011] Step 3: Integrate the popularity information of words and entities, and generate a popularity score for the news through a multi-layered self-attention mechanism and a cross-attention mechanism;

[0012] Based on user behavior data, we construct a word popularity set W(t) and an entity popularity set E(t) for news articles. First, we map W(t) and E(t) to a vector space through an embedding layer to obtain the word popularity embedding matrix. and entity popularity embedding matrix Subsequently, a self-attention mechanism is used to model the popularity associations within the word set and the entity set, respectively, to generate internal relation representations. Simultaneously, cross-modal popularity information between words and entities is fused through a cross-attention mechanism to obtain interactive representations. ;Will and After addition, an attention mechanism is used again for fusion to generate a word-side popularity representation. Perform the same operation on the entity side to generate an entity popularity representation. Finally, and The concatenation is used to form a unified representation, which is further fused through an attention mechanism, and then mapped to the final popularity score P using a fully connected layer, specifically represented as follows: ,in, Indicates a fully connected network. This is an attention fusion operation.

[0013] Step 4: Convert the duration of the news story into an embedding vector and map it to a time fraction using an MLP;

[0014] Temporal modeling of the existence duration t of the news: First, the scalar duration t is mapped into a dense vector representation through a temporal embedding layer. Subsequently, a small multilayer perceptron (MLP) is used to convert the temporal embedding vector into a temporal fraction T, the process of which is represented as follows: ,in and There are two weight matrices. and There are two bias terms. It is an activation function, and the final time feature vector T, i.e., the time score, will be used as the input of the subsequent dynamic semantic modulation module.

[0015] Step 5: Map the popularity score and time score of the news to K-dimensional semantic channels through a nonlinear transformation. The time feature is modeled for timeliness through a decay function in each channel. Finally, the channel weights generated by fusing the two are used to adaptively weight the multi-semantic representation of the news channel by channel to achieve dynamic modulation of popularity and time decay.

[0016] The computer device performs dynamic semantic modulation on the news multi-semantic representation generated in step 1 to effectively integrate the real-time popularity characteristics and time decay characteristics of the news. By establishing a dynamic weight allocation mechanism for multi-signal perception, adaptive adjustment of different semantic channels is achieved.

[0017] Specifically, the popularity score P is first converted into a multi-dimensional weight vector that matches the number of semantic channels; the conversion process is implemented through a parameterized nonlinear transformation layer: ,in The weight matrix, For bias terms, The Sigmoid activation function is used; the transformation decouples the single global popularity signal into K independent channel weights, enabling the model to distinguish the differences in the response of different semantic dimensions to the popularity signal.

[0018] For time features, a channel decoupling mechanism is used to process the time feature vector T to obtain an intermediate representation. Subsequently, regarding Time-modulated weights are generated using a decay function. This ensures that the weights of each semantic channel exhibit a reasonable decay trend over time, and that different channels maintain independent decay rates; in obtaining popularity weights With time decay weight Then, signal fusion is achieved through element-wise multiplication: Finally, the obtained fusion modulation vector is used Multi-semantic representation of original news A new multi-semantic representation of the news is obtained by performing channel-by-channel weighting: .

[0019] Step 6: Calculate the association between user interests and news semantics through cross-channel attention, and aggregate the matching scores of each channel to obtain the click prediction score;

[0020] Based on the user multi-interest representation obtained in step 2 Compared with the dynamic news semantic representation generated in step 5 To establish fine-grained cross-channel interaction relationships, the process is as follows: First, an interest-semantic association graph is established through a cross-channel attention mechanism: ,in , Let d be two trainable projection matrices, where d is the vector dimension, and the attention matrix is... Each element This represents the association strength between the i-th user interest channel and the j-th news semantic channel; based on the attention matrix, the user-news interaction representation is further calculated: ,in This is a trainable projection matrix; the transformation achieves many-to-many semantic mapping, enabling each user's interest channel to aggregate all relevant news semantic information. This is achieved by obtaining the interactive representation. Then, fine-grained matching scores are calculated on each independent channel: Finally, by aggregating the scores from all channels, we obtain the overall click prediction score for the candidate news items. .

[0021] Step 7: Train the model using the Adam optimizer with negative sampling and negative log-likelihood loss function.

[0022] Training samples are constructed using negative sampling techniques. Each positive news item clicked by a user is randomly paired with four unclicked negative news items to form batch training data. The model training uses the negative log-likelihood loss function, defined as follows: Where S represents the set of positive samples, The predicted score represents the positive sample. The predicted score represents the negative sample; this loss function formalizes the click prediction task as a pseudo-classification problem, optimizing the model parameters by maximizing the score difference between positive and negative samples. The model updates the network parameters using the Adam optimizer with a learning rate of 0.0001, a batch size of 32, and a Dropout ratio of 0.2 to prevent overfitting.

[0023] This application also provides a multi-interest news recommendation system based on popularity-aware timeliness differences. This system is used to implement the method provided in this application, and the system includes:

[0024] The extraction module is used to extract global and local semantic features of news titles through multi-head self-attention and convolutional neural networks, respectively, and then splice and fuse them into a multi-semantic representation in each semantic channel;

[0025] The aggregation module is used to aggregate the representations of users' historical news across various semantic dimensions through a self-attention network, forming multi-interest features;

[0026] The popularity score module is used to integrate the popularity information of words and entities, and generate the popularity score of news through multi-layer self-attention mechanism and cross-attention mechanism;

[0027] The time score module is used to convert the duration of news events into an embedding vector and then map it to a time score using an MLP.

[0028] The fusion module maps the popularity score and time score of the news to K-dimensional semantic channels through nonlinear transformation. The time feature is modeled for timeliness by a decay function in each channel. Finally, the channel weights generated by fusing the two are used to adaptively weight the multi-semantic representation of the news channel by channel to achieve dynamic modulation of popularity and time decay.

[0029] The association module is used to calculate the association between user interests and news semantics through cross-channel attention, and to aggregate the matching scores of each channel to obtain the click prediction score;

[0030] The training module is used to train the model using the Adam optimizer with negative sampling and negative log-likelihood loss functions.

[0031] It should be noted that the technical details of the system and the method in this application are the same, and will not be repeated here.

[0032] The significant advancement of this invention compared to existing technologies lies in:

[0033] By introducing a popularity perception mechanism into news semantic representation, and by modeling real-time popularity signals and decoupling multiple channels, the model can dynamically capture the degree of social attention to news content, thereby significantly improving the recommendation system's ability to perceive trending content and the accuracy of recommendations.

[0034] During the dynamic modulation process, an adaptive time decay weight mechanism is adopted to automatically adjust the decay rate of each semantic channel according to the duration of the news, so that the model can accurately reflect the timeliness change pattern of news value, thereby effectively enhancing the freshness and timeliness of the recommendation results.

[0035] In the matching calculation stage, a multi-channel interactive attention network is used to establish a fine-grained many-to-many mapping relationship between user interests and news semantics, thereby realizing deep semantic interaction across channels and improving the matching accuracy and result stability of the recommendation system.

[0036] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method provided in the embodiments of this application. Detailed Implementation

[0038] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0039] Figure 1 This document illustrates a flowchart illustrating how an embodiment of the present invention combines the dual-end decoupled representation technique in existing news recommendation systems with the dynamic semantic modulation and multi-channel interactive matching of the present invention to guide the model in news recommendation. Figure 1 The paper describes a model that learns from user behavior data and news data, and improves the effectiveness of news recommendation by combining the dynamic semantic adjustment of this invention to extract more optimized news features.

[0040] The implementation process consists of three parts:

[0041] The first part is the two-ended decoupled representation module, which contains two data streams:

[0042] User behavior data --> User encoder --> "Multi-interest user representation": This means that in existing technologies, the user encoder receives the user's historical news sequence and generates the user's multi-interest representation through a multi-head attention mechanism.

[0043] Candidate news data --> News encoder - "Multi-semantic news representation": This represents the process in existing technologies where the news encoder receives candidate news content and generates a multi-semantic representation of the news through CNN and attention mechanisms.

[0044] The second part is the dynamic semantic modulation module, which is the core innovative module proposed in this invention. It receives multi-semantic news representations from the news encoder and processes two types of dynamic signals simultaneously. It includes the following data streams:

[0045] Popularity data --> Popularity encoder --> "Popularity features" --> Popularity score calculation: The popularity of representative words and the popularity of entities are processed by the popularity encoder, concatenated to generate popularity features, and then the popularity score is obtained through a fully connected network.

[0046] Popularity score calculation --> nonlinear transformation --> "Popularity attention vector": represents mapping the popularity score through a parameterized nonlinear transformation layer into a weight vector that matches the number of semantic channels.

[0047] News duration --> Time encoder --> "Time feature" --> Time score calculation: This represents the conversion of news duration into a time feature representation through a time embedding layer and MLP.

[0048] Time fraction calculation --> Nonlinear transformation --> "Multi-channel time fraction" --> Time decay function --> "Time attention vector": This represents processing the time fraction through a nonlinear transformation and decay function to generate a time decay weight vector.

[0049] "Popularity Attention Vector" and "Time Attention Vector" --> Fusion --> "Multi-Signal Modulation Vector": This represents the fusion of the weights of the two dynamic signals through element-wise multiplication.

[0050] "Multi-signal modulation vector" and "multi-semantic news representation" --> element-wise multiplication - "dynamic multi-semantic news representation": This represents the generation of a dynamic news representation by weighting the original news semantic representation obtained in the first part through the modulation vector channel by channel.

[0051] The third part is the multi-channel interactive network module, another core innovative module of this invention. It receives multi-interest user representations from the first part and dynamic multi-semantic news representations from the second part. Through a cross-channel attention mechanism, it establishes an interest-semantic association graph, achieving fine-grained matching between user multi-interests and news multi-semantics. It mainly includes the following data streams:

[0052] "Multi-interest user representation" and "Dynamic multi-semantic news representation" --> Cross-attention: This represents the calculation of the correlation strength between user interests and news semantics through a cross-channel attention matrix, and the generation of user representations after interaction.

[0053] Cross-attention --> Match score calculation and "Dynamic multi-semantic news representation" --> Match score calculation: This represents calculating fine-grained matching scores on each semantic channel separately, and then aggregating them to obtain the final click prediction score.

[0054] This embodiment provides a multi-interest news recommendation method based on popularity perception and timeliness differences. By integrating the real-time popularity signal and time-series decay characteristics of news, it achieves dynamic evaluation of the semantic value of news, thereby improving the accuracy and timeliness of recommendation results.

[0055] Referring to Table 1, the significant differences in recommendation effectiveness between the embodiments of the present invention and existing methods are clearly demonstrated. Table 1 uses a two-column comparison format. The left two columns present the recommendation results and scores of the present invention, while the right two columns show the recommendation results of existing methods. The bolded news IDs represent the news actually clicked by the user, intuitively reflecting the gap in recommendation accuracy between the two methods.

[0056] Table 1: Comparison of Results

[0057] ID News ID ID News ID 1 N15347 1 N38951 2 N21681 2 N55036 3 N29862 3 N21681 4 N56969 4 N56969 5 N1952 5 N7618 6 N46749 6 N15347 7 N7618 7 N46749 8 N55036 8 N1952 9 N38951 9 N29862

[0058] The following example, using user U2115's recommendation scenario on November 15, 2019, illustrates the specific implementation process of this invention. The system constructs a user interest profile based on the user's historical click sequence and simultaneously processes a recommendation task involving a candidate set of 9 news articles {N21681, N56969, N1952, N46749, N38951, N29862, N15347, N7618, N55036}.

[0059] The system first performs deep modeling of user historical behavior through a dual-end decoupled representation module. The user encoder employs a 5-head self-attention mechanism to extract 5-dimensional interest representation vectors from 50 historical news articles. During the news encoding stage, the system performs multi-dimensional semantic parsing on each candidate news article. Taking news article N15347 as an example, the news encoder extracts local semantic patterns from the title using a convolutional neural network, and then combines this with the self-attention mechanism to generate a news representation containing 5 semantic channels, each corresponding to a different semantic dimension. The dynamic semantic modulation module plays a crucial role in this process. The system uses pre-collected popularity data of words in candidate news titles and entities such as "National Rugby League" and "Team X" at the current moment to obtain a popularity score. Simultaneously, it combines the current duration of the news article with nonlinear transformation and weight fusion to generate a modulation vector, thus obtaining a dynamically adjusted news representation for the 5 semantic channels. A multi-channel interactive network then establishes fine-grained interest-semantic associations. Based on the learned attention weights, the system generates a deeply fused user representation and calculates matching scores on each semantic channel. Ultimately, news article N15347 achieves a comprehensive recommendation score of 0.73. The system sorted all candidate news articles by score from highest to lowest to generate a recommendation list, successfully placing news article N15347, which the user actually clicked, at the top of the recommendation list. In contrast, the existing method resulted in news article N15347 having a final recommendation score of 0.62, ranking it sixth on the recommendation list.

[0060] This invention, by introducing a dynamic semantic modulation mechanism, can more accurately capture the timeliness changes of content value in news recommendations. Even if some news items have low static semantic matching scores, if their popularity characteristics are prominent and their publication time is appropriate, the method of this invention can still identify their potential value. This helps the recommendation system better balance content quality and real-time popularity, thereby improving recommendation performance. For example, in the example above, although existing methods may rank the actually clicked news item N15347 sixth due to over-reliance on long-term interests, this invention, by integrating real-time popularity and time decay signals, correctly identifies its current value and places it first. This demonstrates that this invention can guide the recommendation system to discover more timely and high-quality content, improving user experience and engagement.

[0061] 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-interest news recommendation method based on popularity perception and timeliness differences, characterized in that, The method includes: Step 1: Extract global and local semantic features of news headlines using multi-head self-attention and convolutional neural networks respectively, and then concatenate and fuse them into a multi-semantic representation in each semantic channel; Step 2: Aggregate the representations of users' historical news across various semantic dimensions using a self-attention network to form multi-interest features; Step 3: Integrate the popularity information of words and entities, and generate a popularity score for the news through a multi-layered self-attention mechanism and a cross-attention mechanism; Step 4: Convert the duration of the news story into an embedding vector and map it to a time fraction using an MLP; Step 5: Map the popularity score and time score of the news to K-dimensional semantic channels through nonlinear transformation. The time feature is modeled for timeliness through a decay function in each channel. Finally, the channel weights generated by the fusion of the two are used to adaptively weight the news multi-semantic representation for each channel to achieve dynamic modulation of popularity and time decay. Step 6: Calculate the association between user interests and news semantics through cross-channel attention, and aggregate the matching scores of each channel to obtain the click prediction score; Step 7: Train the model using the Adam optimizer with negative sampling and negative log-likelihood loss function.

2. The method according to claim 1, characterized in that, Step 1: Extract global and local semantic features of news headlines using multi-head self-attention and convolutional neural networks respectively, and then concatenate and fuse them from each semantic channel into a multi-semantic representation, including: Word sequence of candidate news headlines Performing multi-dimensional semantic parsing involves the following processing steps: First, a global semantic model of the word sequence is performed using a multi-head self-attention mechanism to generate a global representation containing K semantic channels. Each attention head focuses on capturing different aspects of global semantic features; simultaneously, a convolutional neural network extracts local semantic features from the word sequence, using a convolutional kernel with a window size of w to scan the word sequence and generate corresponding local semantic representations. Finally, the global features of each semantic channel k are... With local features By splicing and merging the data, a complete multi-semantic representation of the news can be formed. Each of the channels The corresponding news is a complete representation on a specific latent semantic dimension; where K is set to 5 and w is set to 3.

3. The method according to claim 2, characterized in that, Step 2: Aggregate the representations of users' historical news across various semantic dimensions using a self-attention network to form multi-interest features; including: Historical click sequences of the target user First, a news decoupling method is used to decouple each news item in the sequence. Generate decoupled representations corresponding to K latent semantic dimensions. Based on this, for each specific semantic dimension k, k=1,2,...,K, the decoupled representation of all historical news under the corresponding dimension is obtained. As input, a self-attention network is used to model the intrinsic relationships between different news items in their corresponding semantic dimensions. The self-attention mechanism calculates the relevance weights between news representations within a sequence to highlight important news items, and then obtains the user's interest representation in dimension k through weighted aggregation. Finally, the interest representations of the K semantic dimensions are integrated to form a complete multi-interest vector for the user. Each dimension This corresponds to capturing users' interests and preferences in specific semantic aspects.

4. The method according to claim 3, characterized in that, Step 3: Integrate the popularity information of words and entities, and generate a popularity score for the news through multi-layered self-attention and cross-attention mechanisms; including: Based on user behavior data, we construct a word popularity set W(t) and an entity popularity set E(t) for news articles. First, we map W(t) and E(t) to a vector space through an embedding layer to obtain the word popularity embedding matrix. and entity popularity embedding matrix Subsequently, a self-attention mechanism is used to model the popularity associations within the word set and the entity set, respectively, to generate internal relation representations. Simultaneously, cross-modal popularity information between words and entities is fused through a cross-attention mechanism to obtain interactive representations. ;Will and After addition, an attention mechanism is used again for fusion to generate a word-side popularity representation. Perform the same operation on the entity side to generate an entity popularity representation. Finally, and The concatenation is used to form a unified representation, which is further fused through an attention mechanism, and then mapped to the final popularity score P using a fully connected layer, specifically represented as follows: ,in, Indicates a fully connected network. This is an attention fusion operation.

5. The method according to claim 4, characterized in that, Step 4: Convert the duration of the news item into an embedding vector and map it to a time fraction using an MLP; including: Temporal modeling of the existence duration t of the news: First, the scalar duration t is mapped into a dense vector representation through a temporal embedding layer. Subsequently, the temporal embedding vector is converted into a temporal fraction T using a multilayer perceptron (MLP), the process of which is represented as follows: ,in and There are two weight matrices. and There are two bias terms. It is an activation function, which ultimately yields the time feature vector T, i.e., the time score.

6. The method according to claim 5, characterized in that, Step 5: Map the news's popularity score and time score to K-dimensional semantic channels via a nonlinear transformation. The time feature is modeled for timeliness using a decay function in each channel. Finally, the channel weights generated by fusing the two scores are adaptively weighted channel-by-channel to dynamically modulate the news's multi-semantic representation, achieving dynamic modulation of popularity and time decay. This includes: Dynamic semantic modulation is performed on the news multi-semantic representation generated in step 1. By establishing a dynamic weight allocation mechanism for multi-signal perception, adaptive adjustment of different semantic channels is achieved.

7. Specifically, the popularity score P is first converted into a multi-dimensional weight vector that matches the number of semantic channels; the conversion process is implemented through a parameterized nonlinear transformation layer: ,in The weight matrix, For bias terms, Use the Sigmoid activation function; For time features, a channel decoupling mechanism is used to process the time feature vector T to obtain an intermediate representation. Subsequently, regarding Time-modulated weights are generated using a decay function. This ensures that the weights of each semantic channel exhibit a reasonable decay trend over time, and that different channels maintain independent decay rates; in obtaining popularity weights With time decay weight Then, signal fusion is achieved through element-wise multiplication: Finally, the obtained fusion modulation vector is used Multi-semantic representation of original news A new multi-semantic representation of the news is obtained by performing channel-by-channel weighting: .

8. The method according to claim 6, characterized in that, Step 6: Calculate the correlation between user interests and news semantics through cross-channel attention, and aggregate the matching scores of each channel to obtain a click prediction score; including: Based on user multi-interest representation Semantic representation of dynamic news To establish fine-grained cross-channel interaction relationships, the process is as follows: First, an interest-semantic association graph is established through a cross-channel attention mechanism: ,in , Let d be two trainable projection matrices, where d is the vector dimension, and the attention matrix is... Each element This represents the association strength between the i-th user interest channel and the j-th news semantic channel; based on the attention matrix, the user-news interaction representation is further calculated: ,in For trainable projection matrices; in obtaining interactive representations Then, fine-grained matching scores are calculated on each independent channel: Finally, by aggregating the scores from all channels, we obtain the overall click prediction score for the candidate news items. .

9. The method according to claim 7, characterized in that, Step 7: Train the model using the Adam optimizer with negative sampling and negative log-likelihood loss functions; including: Training samples are constructed using negative sampling techniques. Each positive news item clicked by a user is randomly paired with four unclicked negative news items to form batch training data. The model training uses the negative log-likelihood loss function, defined as follows: Where S represents the set of positive samples, The predicted score represents the positive sample. The predicted score represents the negative sample; the model updates the network parameters using the Adam optimizer with a learning rate of 0.0001, the batch size is set to 32, and the Dropout ratio is set to 0.2 to prevent overfitting.

10. A multi-interest news recommendation system based on popularity-aware timeliness differences, the system being used to implement any one of the methods of claims 1 to 8, characterized in that, The system includes: The extraction module is used to extract global and local semantic features of news titles through multi-head self-attention and convolutional neural networks, respectively, and then splice and fuse them into a multi-semantic representation in each semantic channel; The aggregation module is used to aggregate the representations of users' historical news across various semantic dimensions through a self-attention network, forming multi-interest features; The popularity score module is used to integrate the popularity information of words and entities, and generate the popularity score of news through multi-layer self-attention mechanism and cross-attention mechanism; The time score module is used to convert the duration of news events into an embedding vector and then map it to a time score using an MLP. The fusion module maps the popularity score and time score of the news to K-dimensional semantic channels through nonlinear transformation. The time feature is modeled for timeliness by a decay function in each channel. Finally, the channel weights generated by fusing the two are used to adaptively weight the multi-semantic representation of the news channel by channel to achieve dynamic modulation of popularity and time decay. The association module is used to calculate the association between user interests and news semantics through cross-channel attention, and to aggregate the matching scores of each channel to obtain the click prediction score; The training module is used to train the model using the Adam optimizer with negative sampling and negative log-likelihood loss functions.

Images