News recommendation method fusing activation spreading theory and ebbinghaus forgetting theory

Abstract

This invention relates to a news recommendation method integrating activation diffusion theory and Ebbinghaus forgetting theory. The method includes the following steps: acquiring a knowledge graph dataset and a news dataset, and preprocessing them respectively; using BERT to obtain vector representations of news words; aggregating word and entity vectors in a convolutional neural network (CNN) using activation diffusion theory to obtain feature vectors for the news; improving the attention mechanism using Ebbinghaus forgetting theory to obtain the user's feature vector based on the user's click history; and predicting click-through rates and recommending news based on the feature vectors of candidate news and the user's feature vector. Compared with existing technologies, this invention can accurately mine the relationship between users and candidate news, thereby improving the accuracy of recommendations.

Description

Technical Field

[0001] This invention relates to the fields of data analysis and mining and deep learning technology, and in particular to a knowledge graph-based news recommendation method that utilizes activation diffusion theory and Ebbinghaus forgetting theory. Background Technology

[0002] Recommender systems involve interdisciplinary research and are one of the most important applications in data analysis and mining, and information retrieval. A typical recommender system consists of three parts: a user model, an item model, and a recommendation algorithm, with the recommendation algorithm undoubtedly being the core issue. Based on different recommendation algorithms, they can be broadly categorized into traditional machine learning-based recommender systems and deep learning-based recommender systems.

[0003] Traditional recommendation systems can be categorized into content-based recommendation, collaborative filtering-based recommendation, and hybrid recommendation. Content-based recommendation uses user profiles and item features to derive user preferences from items the user has already interacted with. It then matches these preferences with item content features, ultimately providing recommendations similar to the user's preferences. Collaborative filtering-based recommendation uses algorithms to mine historical data of the target user to predict their interest in uninterrupted items or calculates similarity between target items to find neighboring items for candidate items and recommends items with high predicted ratings and no prior interaction. Hybrid recommendation combines these algorithms across different domains.

[0004] Depending on the deep learning model used, deep learning-based recommendation systems can be categorized in several ways. Among them, recommendation methods based on knowledge graphs combined with neural networks leverage the structural characteristics of knowledge graphs to connect user history records with the recommendation process. Furthermore, the semantic relationships contained in various types of items can be reflected in the knowledge graph, making this type of method widely used in recommendation systems across various general and vertical fields. Based on different underlying principles, they can be mainly divided into three types: First, connection-based recommendation, which utilizes the heterogeneous network characteristics of knowledge graphs to calculate the similarity between nodes and mine user interests through different relationships on the edges in the network, thereby completing the recommendation. Second, embedding-based recommendation, which maps entities and relationships in the graph to low-dimensional vectors, representing entities and relationships through graph embedding methods, thus expanding the semantic information of the original item and user representations. Third, hybrid recommendation, which obtains user interests across the entire knowledge graph and learns features from these interests through a graph embedding module.

[0005] In summary, current knowledge graph-based recommendation methods primarily address the problem of inaccurate recommendations caused by insufficient semantic information in the recommendation process, which fails to capture the evolution of user interests. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a news recommendation method based on the fusion activation diffusion theory and Ebbinghaus forgetting theory of knowledge graphs.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A news recommendation method integrating activation diffusion theory and Ebbinghaus forgetting theory, the news recommendation method comprising the following steps:

[0009] Obtain the knowledge graph dataset and the news dataset, and preprocess the data in the knowledge graph dataset and the news dataset;

[0010] Obtain vector representations of news words;

[0011] By using activation diffusion theory, word and entity vectors are aggregated in a convolutional neural network (CNN) to obtain the feature vector of the news.

[0012] We improve the attention mechanism by utilizing Ebbinghaus's forgetting theory and obtain the user's feature vector based on the user's click history in news articles.

[0013] Based on the feature vectors of candidate news articles and user feature vectors, click-through rates are predicted and news recommendations are made.

[0014] Furthermore, the process of obtaining the knowledge graph and news recommendation datasets, and preprocessing the data in the knowledge graph and news datasets, is as follows:

[0015] Knowledge graph preprocessing: Obtain the knowledge graph dataset, entity E i After training using the TransD knowledge graph representation learning algorithm, entity feature vectors are obtained.

[0016] News dataset preprocessing: Each news record clicked by a user is processed into the following information: user ID, sampling time, specific news title, and actual click details.

[0017] Furthermore, the process of obtaining the vector representation of news words is as follows:

[0018] Regarding the word "w" appearing in a news article i Using BERT to perform vectorization, we obtain the feature vectors of the words. Then concatenate the feature vectors of all words into a feature vector matrix for each word.

[0019] Where m is the word vector dimension, This represents the word vector of w1. Let n represent a matrix, where n is the total number of words.

[0020] Furthermore, the process of obtaining the entity vector is as follows:

[0021] Calculate semantic distance: Locate the entity corresponding to the news word in the knowledge graph, calculate the semantic association distance between the entity and its nth-order neighbor nodes according to the activation diffusion theory, and construct the context neighbor node set of the entity.

[0022] The context vector represents the entity E obtained from activation diffusion. i The feature vector of the context neighbor node set is represented as Concatenate the context vectors of the entities corresponding to all words in the news article into a context feature vector matrix.

[0023] Among them, neighbor(E i ) represents entity E i The set of all neighbor nodes after activation and diffusion, where L is the total number of neighbors of the current entity. For entity E in the knowledge graph l Entity feature vectors trained using TransD graph embedding. Let n represent the total number of words and m be the vector dimension. If a word does not have a corresponding entity in the knowledge graph, then m represents the vector dimension. Take 0.

[0024] Furthermore, the process of calculating the semantic distance is as follows:

[0025] The entity's first-order neighbors are used as the initial context set. Then, second-order neighbors are expanded outwards from these first-order neighbors. During this expansion, new nodes with smaller semantic association distances replace the original first-order neighbors, updating the entity's context set. This process is iterated until no new nodes with smaller semantic association distances are found, at which point the expansion stops. The formula for calculating the semantic association distance is:

[0026]

[0027] Among them, ti k and tj k They are entity E i and E j The values ​​of the vector in the k-th dimension, and It is entity E i and E jThe average value of the vector, where m is the vector dimension.

[0028] Furthermore, the process of aggregating word and entity vectors in a convolutional neural network (CNN) to obtain the feature vector of the news article is as follows:

[0029] The feature vector matrix H of the obtained n×m dimensional words w and the context feature vector matrix H of n×m dimensional entities C Stack them into an n×m×2 dimensional vector matrix, which serves as the input R:

[0030]

[0031] in, This represents the word vector of w1. The feature vector representing the set of context neighbor nodes of entity E1;

[0032] The convolutional layer uses multiple sliding windows of length L with a stride of 1. The initial weight matrix of the convolutional kernel is... The convolution kernel weight matrix in submatrix R i：i+L-1 The sliding window moves with a step size, the number of convolution kernels is k, and the local corresponding eigenvectors of the submatrix with respect to the convolution kernels are c. i The calculation formula is:

[0033] c i =[h*R i：i+L-1 +b]

[0034] Where b is the bias matrix, and all c i The features are combined into a feature matrix obtained by convolution with a single convolution kernel, and the ReLU function is used as the activation function.

[0035] The matrix is ​​dimensionality reduced in the pooling layer to obtain a vector of fixed size, and then max pooling is used to filter the feature vector c. t ;

[0036] The feature vector c obtained after max pooling is processed in the output layer. t By concatenating these elements, we can obtain the final representation of the news feature vector T(c) of a fixed length.

[0037] Furthermore, the improvement of the attention mechanism using Ebbinghaus forgetting theory, and the obtaining of the user's feature vector representation based on historical news records, is as follows:

[0038] The forgetting matrix is ​​constructed using the Ebbinghaus forgetting curve: The formula based on the Ebbinghaus forgetting curve is as follows:

[0039] f(x) = e (x-1)[0.0123×1n(x-1)]-0.0639

[0040] Where x represents the most recent x-th click, and f(x) represents the memory retention rate;

[0041] Based on the actual number of times users clicked on news articles, the memory retention rate for each click was obtained and normalized into a forgetting matrix:

[0042]

[0043] in, f(x) is the memory retention ratio after softmax normalization, which is used as the forgetting weight, and M is the number of times the user actually clicked the news.

[0044] Improve attention mechanism: Include candidate news items... k And user u's historical news records The attention weight coefficients are obtained by inner product and normalization, and the specific formula is as follows:

[0045]

[0046] Here, method ρ represents the operation of taking two feature vectors as input, performing an inner product along a specified dimension, and summing the results. and T(c k ) represents user records and candidate news c k Vector representation;

[0047] The historical records of user u are obtained by summing the corresponding weights and performing softmax normalization. The final weight of the corresponding vector

[0048] Calculate the user vector: Calculate the final weights. News vectors clicked by users Multiply and sum to obtain the user's vector representation T(u).

[0049] Furthermore, the click-through rate prediction is based on the obtained user vector T(u) and candidate news vector T(c). k To predict click-through rate, the calculation formula is:

[0050]

[0051] The method p combines the reduce_sum function from the TensorFlow library with the matrix inner product operation. It takes two feature vectors as input, performs an inner product along a specified dimension, and sums them.

[0052] Furthermore, the training process of the news recommendation method is as follows:

[0053] Set an iteration limit η, and use the input dataset to train the news recommendation method;

[0054] Compare the news recommendation prediction results with the actual results and calculate the loss function;

[0055] If the iteration limit is reached and convergence is achieved, the training is complete. The accuracy of all predicted click rates and actual click rates is evaluated and the results are output. Otherwise, training continues based on the hyperparameters.

[0056] Furthermore, the calculation of the loss function is as follows:

[0057] Based on the actual results and the predicted results p, the loss is calculated using the cross-entropy function and L2 regularization from the TensorFlow library, and then optimized using the Adam Optimizer from the TensorFlow library.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] Compared with other news recommendation methods, the method described in this invention obtains the feature vector of news through activation diffusion theory, which can make full use of the contextual information in the knowledge graph. It also obtains the feature vector of users through Ebbinghaus forgetting theory, which can accurately mine the relationship between users and candidate news, thereby improving the accuracy of recommendations. Attached Figure Description

[0060] Figure 1 This invention presents a news recommendation method based on the fusion activation diffusion theory and Ebbinghaus forgetting theory of knowledge graphs, and a flowchart of its training process. Detailed Implementation

[0061] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0062] Example 1

[0063] This invention proposes a news recommendation method based on knowledge graphs that integrates activation diffusion theory and Ebbinghaus forgetting theory. It uses activation diffusion theory to aggregate word and entity vectors within a convolutional neural network to represent news vectors, and utilizes Ebbinghaus forgetting theory to improve the attention mechanism, obtaining user vector representations for training the recommendation model. Finally, it can predict news recommendations based on the given data. The main features of this invention are as follows: first, in the process of representing news vectors using convolutional neural networks, rich entity contextual information is added to the knowledge graph using activation diffusion theory; second, the attention mechanism is improved using Ebbinghaus forgetting theory.

[0064] like Figure 1 The diagram shows the flowchart of the news recommendation method and its training provided by the present invention. The specific implementation steps of this process include the following 6 steps:

[0065] Step 1: Obtain the knowledge graph dataset and news dataset, and preprocess them separately. The specific process is as follows:

[0066] (1-1) Knowledge Graph Processing: Download the Wikidata knowledge graph dataset published by Wikipedia in OpenKE as the knowledge graph to be used. Download address: http: / / openke.thunlp.org / . Utilize the TransD knowledge graph representation learning algorithm in OpenKE tools, entity E... i Vector representations in the knowledge graph are obtained through training. m is the vector dimension. The Wikidata knowledge graph dataset contains a total of 20,982,733 entities, 594 different relations, and 68,904,773 triples.

[0067] (1-2) News Dataset Preprocessing: Download the Microsoft News Dataset (MIND) dataset released by Microsoft Research Asia in conjunction with Microsoft News Products Department as the data set for news recommendation. Download address: https: / / msnews.github.io / . After obtaining the required news recommendation dataset, process each news record clicked by the user into the following format: User ID, sampling time, specific news title (including words and entities), and actual click status (0 indicates no click, 1 indicates click). The MIND news dataset includes 156,965 different anonymous users, 1,816,594 interaction messages, 51,282 different news articles, 52,930 different words, and 26,904 different entities.

[0068] Step 2: Use BERT to obtain the vector representations of news words. The specific process is as follows:

[0069] For the word "wi" appearing in a news article, BERT was used for pre-training. The training model chosen was Google's "BERT-Base, Multilingual Cased (New, Recommended)", which can be downloaded from: https: / / github.com / google-research / bert. Tencent's bert-serving-server and bert-serving-client services were used. The BERT service was started by specifying the model, and the client was called to perform vectorization, resulting in word vectors. The vector dimension is 100.

[0070] By concatenating the feature vectors of all words together, we can obtain the feature vector matrix of the words. There are a total of 52,930 words.

[0071] Where m is the word vector dimension, This represents the word vector of w1. Let n represent a matrix, where n is the total number of entities.

[0072] Step 3: Using activation diffusion theory, aggregate word and entity vectors in a convolutional neural network (CNN) to obtain the feature vector representation of the news article. The specific process is as follows:

[0073] (3-1) Calculating Semantic Distance: Locate the entity corresponding to the news word in the knowledge graph. Based on activation diffusion theory, calculate the semantic association distance between the entity and its nth-order neighbor nodes, and construct the context neighbor set of the entity. The specific process is as follows: take the entity's first-order neighbors as the initial context set, and based on these neighbor nodes, expand outward to second-order neighbors. Replace the original first-order neighbors with new nodes with smaller semantic association distances during the diffusion process, update the entity's context set, and iterate this process continuously. If there are no new nodes with smaller semantic association distances, then... The average value of a vector, where m is the vector dimension.

[0074] (3-2) represents the context vector: the entity E obtained from activation diffusion in the previous step. i The feature vector of the context neighbor node set is represented as The context vectors of the entities corresponding to the words in the news are concatenated into a context feature vector matrix. There are a total of 26,904 context vectors for each entity, and each entity has an average of 6.58 context neighbor nodes.

[0075] Among them, neighbor(E i ) represents entity E i The set of all neighbor nodes after "activation diffusion", where L is the total number of neighbors of the current entity. For entity E in the knowledge graph l Entity feature vectors trained using TransD graph embedding. Let n represent the total number of words and m be the vector dimension. If a word does not have a corresponding entity in the knowledge graph, then m represents the vector dimension. Take 0.

[0076] (3-3) Obtain the feature vectors of the news using a convolutional neural network (CNN): The resulting n×m dimensional feature vector matrix H of the words... Wand the context feature vector matrix H of n×m dimensional entities C The input is a vector matrix formed by stacking elements using the `concat` function from the TensorFlow library, which is then used as the input.

[0077]

[0078] Multiple sliding windows of length L are used in the convolutional layer, specifically set to [1, 2, 3, 4], with a stride of 1. The initial weight matrix of the convolutional kernel is... The convolution kernel weight matrix in submatrix R i：i+L-1 The sliding window moves with a step size, the number of convolutional kernels is k (specifically set to 100), the learning rate is set to 0.001, and each Bach is set to 100. The formula for calculating the local eigenvectors of the submatrix with respect to the convolutional kernel is as follows:

[0079] c i =[h*R i：i+L-1 +b]

[0080] Where b is the bias matrix, and finally all c i The features are combined into a feature matrix obtained by convolution with a single convolution kernel, and the ReLU function is used as the activation function.

[0081] The matrix is ​​dimensionality reduced in the pooling layer to obtain a vector of fixed size. Max pooling is then used to select the feature vectors, and the calculation formula is as follows:

[0082] c t =max{c1, c2, ..., c n-L+1}

[0083] Where n is the total number of entities and L is the length of the sliding window.

[0084] In the output layer, the features obtained after max pooling are concatenated to obtain the final representation of the news feature vector T(c) of fixed length, as shown in the formula:

[0085]

[0086] Where k is the number of convolution kernels, and its size is the same as the word vector dimension m, set to 100, and c represents news.

[0087] Step 4: Improve the attention mechanism using Ebbinghaus forgetting theory. Obtain the user's feature vector representation based on historical news records. The process is as follows:

[0088] (4-1) Constructing the forgetting matrix using the Ebbinghaus forgetting curve: The formula is derived based on the Ebbinghaus forgetting curve:

[0089] y = e(x-1)[0.0123×ln(x-1)]-0.0639

[0090] Where x represents the most recent x-th click, and f(x) represents the memory retention rate.

[0091] Based on the actual number of times users clicked on news articles, the memory retention rate for each click was obtained and normalized into a forgetting matrix:

[0092]

[0093] in, The memory retention ratio of f(x) after softmax normalization, with a value in the range of [0, 1], is used as the forgetting weight. M is the number of times the user actually clicked on the news, and its maximum value is set to 30.

[0094] (4-2) Improve attention mechanism: select candidate news c k And user u's history The attention weight coefficients are obtained by inner product and normalization, and the specific formula is as follows:

[0095]

[0096] Here, method ρ combines the `reduce_sum` function from the TensorFlow library with the matrix inner product operation. It takes two feature vectors as input, performs an inner product along a specified dimension, and sums them. and T(c k ) represents user records and candidate news c k Vector representation.

[0097] The historical records of user u are obtained by summing the corresponding weights and performing softmax normalization. The final weight of the corresponding vector is calculated using the following formula:

[0098]

[0099] in, For user u's historical news records With News c k Attention weight coefficient, For News The number of points to forget.

[0100] (4-3) Calculate the user vector: calculate the final weight points. News vectors clicked by users

[0101] Step 5: Predict click patterns based on the feature vectors of candidate news articles and user feature vectors, compare the predicted results with the actual results, and calculate the loss function. The specific process is as follows:

[0102] (5-1) Click-through rate prediction: Based on the user vector T(u) and candidate news vector T(c) obtained in the previous steps... k To predict click-through rate, the calculation formula is:

[0103]

[0104] The method ρ combines the reduce_sum function from the TensorFlow library with the matrix inner product operation. It takes two feature vectors as input, performs an inner product along a specified dimension, and sums them.

[0105] (5-2) Calculate the loss function: Based on the actual result and the predicted result p, calculate the loss using the cross-entropy function and L2 regularization in the tensorflow library, and optimize it using the Adam Optimizer in the tensorflow library.

[0106] 6. If the iteration limit is reached and convergence is achieved, the news recommendation model is considered trained and an evaluation metric is output; otherwise, training continues based on the hyperparameters.

[0107] Set an iteration limit η, specifically 10. Training ends when the number of training iterations reaches the limit η. For all predicted click rates and actual click rates, the accuracy of the experimental prediction results is evaluated using the TensorFlow library's AUC (Area Under Curve), mAP (mean Average Precision), MSE (Mean Square Error), and F1 (F1-score) metrics, and the results are output.

[0108] Experimental Description and Results: The training set was the Microsoft MIND News dataset described in steps (1-2), and the test set was also the Microsoft MIND News dataset. The training set included 1,453,276 interaction records, 114,066 different anonymous users, 32,047 different news articles, and 51,203 different words. The test set included 363,318 interaction records, 35,877 different anonymous users, 17,003 different news articles, and 35,664 different words. Table 1 shows the comparison results of different training methods on the test set.

[0109] Table 1 Test Results

[0110]

[0111] Among them, the AUC metric, which is the area enclosed by the ROC curve and the X-axis, reflects the ranking ability of samples and is the most commonly used evaluation metric in the recommendation field; the mAP metric reflects the average precision of the entire recommendation system; the MSE metric represents the mean squared error; the F1 metric combines precision and recall; NPA uses word-level and sentence-level attention mechanisms respectively in the representation of news vectors and user vectors; NAML, compared to NPA, considers the influence of news category in the representation of news vectors; NRMS, compared to NPA, uses a multi-head attention mechanism instead of a single attention mechanism; DKN uses a multi-channel approach similar to RGB three primary colors in news representation; KGCN, based on the idea of ​​GCN, uses a neighbor set of a specific size as the receptive field in the news representation process; RippleNet uses knowledge graphs to spontaneously expand user interests in the representation of user vectors; the method described in this invention aggregates word and entity vectors in a convolutional neural network based on activation diffusion theory to represent news vectors, and uses Ebbinghaus forgetting theory to improve the attention mechanism to obtain the vector representation of users. As can be seen, the method described in this invention achieves the highest accuracy in terms of AUC and mAP evaluation metrics. The AUC is 8.9% higher than the last-ranked method and 4.0% higher than the second-ranked method. Furthermore, the mean squared error of MSE is good, and it also ranks first in the F1 score, which balances precision and recall.

[0112] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A news recommendation method integrating activation diffusion theory and Ebbinghaus forgetting theory, characterized in that, The news recommendation method includes the following steps: Obtain the knowledge graph dataset and the news dataset, and preprocess the data in the knowledge graph dataset and the news dataset; Obtain vector representations of news words; The activation diffusion theory is used to aggregate word and entity vectors in a convolutional neural network (CNN) to obtain the feature vector of the news article; the process of obtaining the entity vector is as follows: Calculate semantic distance: Locate the entity corresponding to the news word in the knowledge graph, calculate the semantic association distance between the entity and its nth-order neighbor nodes according to the activation diffusion theory, and construct the context neighbor node set of the entity. Context vector: the entity obtained from activation diffusion The feature vector of the context neighbor node set is represented as The context vectors of the entities corresponding to all words in the news are concatenated into a context feature vector matrix. ; in, Representing entities The set of all neighboring nodes after activation and diffusion. L The total number of neighbors of the current entity. For entities in a knowledge graph Entity feature vectors trained using TransD graph embedding. Represents a matrix, n Total number of words m For vector dimensions, if a word does not have a corresponding entity in the knowledge graph, Set to 0; The attention mechanism is improved by utilizing Ebbinghaus's forgetting theory. The user's feature vector is obtained based on their click history in news articles. The process is as follows: The forgetting matrix is ​​constructed using the Ebbinghaus forgetting curve: The formula based on the Ebbinghaus forgetting curve is as follows: in, Indicates the most recent Click time Indicates the memory retention rate; Based on the actual number of times users clicked on news articles, the memory retention rate for each click was obtained and normalized into a forgetting matrix: in, for The memory retention ratio after softmax normalization is used as the forgetting weight. The actual number of times a user clicked on a news article; Improve attention mechanisms: select candidate news and users u Historical news records The attention weight coefficients are obtained by inner product and normalization, and the specific formula is as follows: Among them, the method This represents an operation that takes two feature vectors as input, performs an inner product along a specified dimension, and then sums them. and Represents user records and candidate news Vector representation; The corresponding weights are summed and then normalized using softmax to obtain the user's result. u Historical records The final weight of the corresponding vector ; Calculate the user vector: Calculate the final weights. News vectors clicked by users Multiply and sum to obtain the user's vector representation. ; Based on the feature vectors of candidate news articles and user feature vectors, click-through rates are predicted and news recommendations are made.

2. The news recommendation method according to claim 1, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The process of obtaining the knowledge graph and news recommendation datasets, and preprocessing the data in the knowledge graph and news datasets is as follows: Knowledge graph preprocessing: Obtaining the knowledge graph dataset, entities After training using the TransD knowledge graph representation learning algorithm, entity feature vectors are obtained. ; News dataset preprocessing: Each news record clicked by a user is processed into the following information: user ID, sampling time, specific news title, and actual click details.

3. The news recommendation method according to claim 1, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The process of obtaining the vector representation of news words is as follows: For a word appearing in a news article Using BERT to perform vectorization, we obtain the feature vectors of the words. Then, the feature vectors of all words are concatenated into a feature vector matrix for each word. ; in, m For word vector dimensions, express Word vectors, Represents a matrix, n This represents the total number of words.

4. The news recommendation method according to claim 1, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The process for calculating the semantic distance is as follows: The entity's first-order neighbors are used as the initial context set. Then, second-order neighbors are expanded outwards from these first-order neighbors. During this expansion, new nodes with smaller semantic association distances replace the original first-order neighbors, updating the entity's context set. This process is iterated until no new nodes with smaller semantic association distances are found, at which point the expansion stops. The formula for calculating the semantic association distance is: in, and They are entities and The vector in the th The value of dimension, and It is a physical entity and The average value of the vectors, For vector dimensions.

5. The news recommendation method according to claim 3, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The process of aggregating word and entity vectors in a convolutional neural network (CNN) to obtain the feature vector of the news article is as follows: The result Feature vector matrix of a word and Context feature vector matrix of a dimensional entity Stacked into A 3D vector matrix as input : in, express Word vectors, Representing entities The feature vector of the set of context neighbor nodes; Using multiple lengths in the convolutional layer A sliding window with a stride of 1 is used, and the initial weight matrix of the convolution kernel is... The convolution kernel weight matrix in the submatrix The sliding window moves with a step size, and the number of convolutional kernels is... k The local corresponding eigenvectors of the submatrix with respect to the convolution kernel The calculation formula is: in, For the bias matrix, all The features are combined into a feature matrix obtained by convolution with a single convolution kernel, and the ReLU function is used as the activation function. The matrix is ​​dimensionality reduced in the pooling layer to obtain a vector of fixed size, and then max pooling is used to filter the feature vectors. ; The feature vector obtained after max pooling is processed in the output layer. By concatenating these features, a fixed-length news feature vector is obtained. The final representation.

6. The news recommendation method according to claim 1, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The click-through rate prediction is based on the obtained user vector. and candidate news vectors The formula for predicting click-through rate is: Among them, the method This is a combination of the reduce_sum function from the TensorFlow library and the matrix inner product operation. It takes two feature vectors as input, performs an inner product along a specified dimension, and sums them.

7. The news recommendation method according to claim 1, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The training process of the news recommendation method is as follows: Set iteration limit The news recommendation method is trained using the input dataset. Compare the news recommendation prediction results with the actual results and calculate the loss function; If the iteration limit is reached and convergence is achieved, the training is complete. The accuracy of all predicted click rates and actual click rates is evaluated and the results are output. Otherwise, training continues based on the hyperparameters.

8. The news recommendation method according to claim 7, which integrates activation diffusion theory and Ebbinghaus forgetting theory, is characterized in that... The loss function is calculated in the following specific steps: Based on actual results and forecast results By using the cross-entropy function in the TensorFlow library and The loss is calculated using regularization and then optimized using the Adam Optimizer from the TensorFlow library.

Images