Energy technology news recommendation method based on deentanglement representation learning

Abstract

The invention belongs to the technical field of data mining and recommendation, and discloses an energy science and technology news recommendation method based on deentanglement representation learning, which comprises the following steps: firstly, obtaining a user behavior data set, and constructing a user feature sequence and global energy science and technology news co-occurrence graph data; secondly, constructing a prediction model DISNER which comprises an SMPRedict model, a CMPRedict model, a dual deentanglement module and an interest stability adaptive fusion module, and then training the DISNER model by utilizing the constructed data to obtain a weight; finally, the weight is loaded back to the model, user behavior data is input, prediction of the probability that the user clicks the candidate energy science and technology news is achieved, and the problems that in an existing energy science and technology news recommendation system, recommendation content is single, user interest modeling is insufficient, and semantic and collaborative information integration is unbalanced are solved.

Description

Technical Field

[0001] The present invention belongs to the field of data mining and recommendation technology, and in particular relates to an energy science and technology news recommendation method based on disentangled representation learning. Background Art

[0002] With the advancement of high technologies such as communications, information, embedded, and intelligent technologies, the energy technology sector has entered a period of rapid growth. With an increasing amount of news related to energy technology development, users are faced with a massive amount of news information. Given the complex nature of the energy sector, including its multidisciplinary, multi-professional, and multi-scenario nature, how to filter news relevant to user interests and needs has become a critical issue. Energy technology news recommendation systems analyze user browsing history, click behavior, and other data to build user profiles, thereby providing users with personalized energy technology news recommendations. This not only helps users quickly access news of interest, saving significant time, but also improves user experience and satisfaction. More importantly, for researchers, practitioners, and policymakers in the energy technology field, accurate news recommendations help them stay informed about industry dynamics, technological advances, and market trends, thereby maintaining their competitive edge in the fierce market.

[0003] Current news recommendation systems are primarily based on two approaches: content-based and user behavior-based. Content-based approaches analyze the features of energy and technology news text (such as titles, text, and keywords), extract semantic information using natural language processing techniques, and construct user interest models to achieve personalized recommendations. User behavior-based approaches rely on collaborative filtering algorithms, analyzing historical user behaviors such as clicks, comments, and sharing to mine similarities between users and construct a user-news interaction matrix for recommendation prediction. Current approaches either rely solely on content-based approaches, failing to fully leverage the complementary advantages of the two approaches. Alternatively, when collaborative information is used to enhance semantic information, overlapping portions are often overemphasized, while the importance of other unique portions is diminished. This enhancement comes at the expense of other potentially valuable information.

[0004] In recent years, disentangled representation learning has been gradually applied to the field of recommendation systems. Its core advantage lies in its ability to decompose complex and intertwined user behaviors into multiple independent and clear latent factors, thereby more accurately capturing the diversity of user interests. By decoupling the multiple factors behind user behavior, such as long-term interests and short-term interests, explicit preferences and latent preferences, it is possible to more flexibly understand the real needs of users, effectively alleviate the problem of monotonous recommended content, and significantly improve the novelty and diversity of recommendation results. Therefore, using disentangled representation learning to decouple collaborative information and semantic information can more effectively integrate the unique contributions of the two, avoid over-emphasizing the overlapping parts, thereby improving the accuracy and diversity of recommendations, enhancing the interpretability and robustness of the model, and providing users with better services. Summary of the Invention

[0005] The purpose of the present invention is to provide an energy technology news recommendation method based on disentangled representation learning, which solves the problems of single recommended content, insufficient user interest modeling, and unbalanced integration of semantic and collaborative information in existing energy technology news recommendation systems.

[0006] To solve the above technical problems, the present invention provides an energy technology news recommendation method based on disentangled representation learning, which includes: S1: obtaining a user behavior dataset and constructing user feature information sequence data, obtaining records of the order in which users click on energy technology news, and constructing a global energy technology news co-occurrence graph;

[0007] S2: Constructing a prediction model DISNER for users clicking on energy technology news. The prediction model DISNER includes the SMPredict model and the CMPredict model, a dual disentanglement module, and an interest stability adaptive fusion module.

[0008] The SMPredict model includes an embedding layer, a word-level attention machine layer, and a news-level attention layer connected in series. The word-level attention layer includes a multi-head attention mechanism and a word attention mechanism connected in series. The embedding layer is a pre-trained GloVe word vector model.

[0009] The CMPredict model includes a series of gated graph neural networks and news-level attention layers;

[0010] The dual disentanglement module includes a parallel proxy contrast loss module and a cross-space exclusion loss module;

[0011] The interest stability adaptive fusion module includes an adaptive window cutting module and an interest stability calculation module connected in series;

[0012] S3: Use the data constructed in S1 to train the DISNER model and obtain the weights of the DISNER model after training;

[0013] S4: Load the DISNER model weights into the DISNER model, input user behavior data into the DISNER model, and obtain the probability of users clicking on candidate energy technology news.

[0014] Furthermore, the customer behavior dataset in S1 includes user ID, news ID, whether clicked, news title, news text, news keywords, and news category.

[0015] Furthermore, the process of constructing the global energy technology news co-occurrence map described in S1 specifically includes the following steps:

[0016] S1.1: Collect all users' historical news click records and treat each clicked energy technology news article as a node in the graph;

[0017] S1.2: Traverse each user's news reading sequence. For any two consecutively read news articles A and B, establish a directed edge from news A to news B. Integrate all user behavior data to obtain a global energy technology news co-occurrence graph containing all nodes and directed edges. Furthermore, the SMPredict model processing described in S2 specifically includes the following steps:

[0018] S2.1a: Input the user feature information sequence data obtained in S1 into the pre-trained GloVe word vector model for initialization, and obtain the word embedding vector sequence x corresponding to the user feature information sequence data n ;

[0019] S2.2a: embed the word vector sequence x n Input into the multi-head attention mechanism to obtain word-level feature representation X n ;

[0020] S2.3a: Represent word-level features as X n Input into the word-level attention mechanism to obtain semantic-based news representation n e , the semantic-based news representation n e The function expression of is shown in formula (1), which represents the news based on semantics as n e Input into the news-level attention layer to obtain the user's semantic interest representation u e , the user's semantic interest representation u e The function expression of is shown in formula (2):

[0021]

[0022] Where: n e is a semantic-based news representation, M is the number of words in a news article, is the attention weight corresponding to the i-th word, is the representation vector of the i-th word, exp is the exponential function, is the query vector, i is the index of the word in the news text, tanh is the activation function, W ω is the parameter matrix, is the jth word vector, is the i-th word vector;

[0023]

[0024] Where: N is the number of energy technology news in the user's historical click records, u e Represents the user's interest in semantics, is the attention weight of the i-th historical click news, is the semantic-based energy technology news representation vector of user u’s historical click records, W e is the parameter matrix and T is the transpose.

[0025] Furthermore, the processing of the CMPredict model in S2 specifically includes the following steps:

[0026] S2.1b: Extract the local energy technology news co-occurrence map from the global energy technology news co-occurrence map obtained in S1;

[0027] S2.2b: Use the gated graph neural network to encode the local energy technology news co-occurrence map obtained in S2.1b to obtain the collaboration-based news representation n p , the function expressions of the specific process are shown in formulas (3) to (5):

[0028] n( 0 )=n e (3)

[0029]

[0030] Where: n (0) is the initialization representation of the graph node, n e is a semantic-based news representation, W g is the parameter matrix, represents the set of adjacent nodes of node i in the local energy technology news co-occurrence graph, l represents the number of layers, represents the node representation of the lth layer at the node i position, represents an intermediate representation of node i at the l+1 layer, GRU is a gated recurrent unit, is the node representation of the l+1th layer at the node i position, n p For collaborative-based news representation, GRU stands for gated recurrent unit;

[0031] S2.3b: Representing collaborative news p Input into the news-level attention layer to obtain the user's collaborative interest representation u p .

[0032] Furthermore, the processing of the dual de-entanglement module in S2 specifically includes the following steps:

[0033] S2.1c: Compute semantically based news representation n e Average representation of user interaction history and collaborative news representation p The average representation of the user interaction history in the semantic space is used to obtain the user's interest proxy in the semantic space and user interest agents in collaborative spaces Its function expressions are shown in formula (6) and formula (7):

[0034]

[0035] Where: N u A news collection of user u's interaction history;

[0036] S2.2c: Leveraging User Interest Proxies in Semantic Space and user interest agents in collaborative spaces and the user's semantic interest representation u e Collaborative interest representation u with the user p Calculating the proxy loss function Its function expression is shown in formula (8):

[0037]

[0038] in: is the proxy loss function, B is the set of users in the training batch, τ is the temperature coefficient, sim() is the cosine similarity function, and v is the alias of other users in the same training batch except u;

[0039] S2.3c: Calculate the user's semantic interest representation u e Collaborative interest representation u with the user p The cosine similarity of the user's semantic interest representation u is calculated by comparing the calculated cosine similarity with the set threshold. e Collaborative interest representation u with the user p If the cosine similarity between the two is greater than the set threshold, the orthogonal constraint loss function is used to penalize the input sample data. Its function expression is shown in formula (9):

[0040]

[0041] in: is the orthogonal constraint loss function, u is the user alias, γ is the threshold, and max is the maximum value.

[0042] Furthermore, the processing of the interest stability adaptive fusion module in S2 specifically includes the following steps:

[0043] S2.1d: Dynamically adjust the window granularity according to the actual number of user behaviors N, and calculate the window size size and the number of windows Ta. The function expression of the window size size is shown in formula (10), and the function expression of the number of cut windows Ta is shown in formula (11):

[0044]

[0045] Where: size is the window size, β is the window hyperparameter, N is the number of news in the user interaction history, Ta is the number of cutting windows, is the ceiling symbol;

[0046] S2.2d: Use the window size and the number of cut windows Ta obtained in S2.1d to perform adaptive window cutting on the user history records, and obtain the window set {{M1,M2,…,M size}1…{M1,M2,…,M size} Ta}, where M2 is the second news vector in the window, and Ta is the number of cut windows;

[0047] S2.3d: Calculate the semantic embedding mean of historical news in each window Its function expression is shown in formula (13):

[0048]

[0049] in: is the semantic embedding mean of historical news in each window, n i is the i-th news vector in the window set, |.| is the absolute value symbol, and t is the index variable;

[0050] S2.4d: Calculate the cosine similarity sequence s between adjacent windows t , its function expression, as shown in formula (14), is consistent with the user's interest stability index μ s , its function expression is shown in formula (15), and the interest stability index α, its function expression is shown in formula (16):

[0051]

[0052] Where: s t is the cosine similarity sequence between adjacent windows, is the semantic embedding mean of historical news in the t+1th window, and Ta is the number of cutting windows;

[0053] Where: μ s It is an indicator of user interest stability;

[0054]

[0055] Furthermore, the process of extracting the local energy technology news co-occurrence map from the global energy technology news co-occurrence map obtained in S1 by the CMPredict model in S2.1b specifically includes the following steps:

[0056] S2.1.1b: Obtain each piece of news that the user has browsed historically from the user feature sequence data and use it as the initial point set for the local energy technology news co-occurrence graph;

[0057] S2.1.2b: From the global energy technology news co-occurrence graph constructed in S1, for each node in the initial node set, select its news neighbors within the multi-hop range. In each hop selection process, select the top M news neighbors according to the weight of the edge. n Neighbor news, and the weighted ranking is the top M n The neighbor news of the local energy technology news is added to the node set of the local energy technology news co-occurrence graph to obtain the node set of the local energy technology news co-occurrence graph. The beneficial effects of the present invention are:

[0058] 1) This paper addresses the technical bottlenecks of existing energy technology news recommendation systems, such as the single recommended content, insufficient user interest modeling, and unbalanced integration of semantic and collaborative information. By introducing disentangled representation learning technology and constructing the DISNER recommendation model, the paper improves the overall performance of the energy technology news recommendation system.

[0059] 2) The SM Predict model proposed in this paper combines word-level attention layer and news-level attention layer through deep semantic modeling, which can accurately analyze the semantic features of energy technology news and improve the relevance of recommendations.

[0060] 3) The CM Predict model proposed in this paper, with the help of a gated graph neural network and a news-level attention layer, can fully explore the potential synergistic relationships between news and enhance the diversity of recommendations.

[0061] 4) The dual disentanglement module proposed in this paper effectively separates users' semantic interests and collaborative interests through proxy contrast loss and cross-space exclusion loss, avoiding the implicit coupling between the two in the vector space, and ensuring that the model can independently and accurately represent users' diverse interests.

[0062] The interest stability adaptive fusion module proposed in the present invention dynamically adjusts the weights of deep interests and shallow interests through adaptive window cutting and interest stability calculation, further improving the personalization and accuracy of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 It is a flow chart of the present invention;

[0065] Figure 2 It is a network structure diagram of the SMPredict model in the present invention;

[0066] Figure 3 It is a network structure diagram of the dual disentanglement module in the present invention. DETAILED DESCRIPTION

[0067] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0068] like Figure 1-Figure 3 As shown, this embodiment discloses a method for recommending energy science and technology news based on disentangled representation learning, including the following steps: The task of this embodiment is to use the user's sequence data N u And the global energy technology news co-occurrence map G n =(V n ,E n ) to predict the probability of a click The specific definitions involved in the problem are as follows: For the user's sequence data, let the set of energy technology news browsed by user u be defined as N u ={n1,n2,…,n N}, where N urepresents the historical click sequence of user u, n represents the news number, the number represents the click order, N represents the number of news in the history of user u, and energy technology news is defined as n=(D n ,E n ,c n ), each energy technology news contains a title D n =[w1,w2,...,w M ], where w represents a word, M represents the number of words in the news title, and E n Represents a collection of entities E n =[e1,e2,...,e T ], where e represents an entity word, T is the number of entities in the news title, and c n is a category, d is the vector dimension.

[0069] S1: Obtain user behavior datasets and construct user feature information sequence data, obtain records of the order in which users click on energy technology news, and construct a global energy technology news co-occurrence map;

[0070] In this embodiment, the customer behavior data set includes user ID, news ID, whether clicked, news title, news text, news keywords, and news category.

[0071] In this embodiment, if a user reads the news article v i Then read the news article v j , then add a line from v to the graph i Point to v j The directed edge (v i ,v j ), the weight of the edge of the news co-occurrence graph is determined based on the frequency of this situation in the reading history of all users, that is, the edge weight reflects the possibility of one news being read immediately after another news.

[0072] In this embodiment, constructing a global energy technology news co-occurrence map specifically includes the following steps:

[0073] S1.1: Collect all users' historical news click records and treat each clicked energy technology news article as a node in the graph;

[0074] S1.2: Traverse each user's news reading sequence. For any two consecutively read news articles, A and B, establish a directed edge from news A to news B. Integrate all user behavioral data to obtain a global energy technology news co-occurrence graph containing all nodes and directed edges. S2: Build a DISNER prediction model for user clicks on energy technology news. The DISNER prediction model includes the SMPredict and CMPredict models, a dual disentanglement module, and an interest stability adaptive fusion module.

[0075] The SMPredict model consists of an embedding layer, a word-level attention layer, and a news-level attention layer, which are connected in series. The word-level attention layer includes a multi-head attention mechanism and a word attention mechanism, which are connected in series. The embedding layer is a pre-trained GloVe (Global Vectors for Word Representation) word vector model.

[0076] The processing of the S2SMPredict model is as follows:

[0077] S2.1a: Input the user feature information sequence data obtained in S1 into the pre-trained GloVe word vector model for initialization, and obtain the word embedding vector sequence x corresponding to the user feature information sequence data n ;

[0078] In this embodiment, the word embedding vector sequence can capture the semantic information of each word and provide a basis for subsequent modeling.

[0079] S2.2a: embed the word vector sequence x n Input into the multi-head attention mechanism to obtain word-level feature representation X n ;

[0080] In some specific implementations, the Multi-Head Self-Attention (MHSA) mechanism allows the model to simultaneously focus on multiple relationships between different words in the text, thereby more comprehensively understanding the semantic structure of energy technology news content. The specific formulas are shown in Equations (1) to (5):

[0081] Q = x n W i Q (1)

[0082]

[0083] MHSA(Q,K,V)=Concat(head1,...,head h )W O (4)

[0084] head i =Attention(Q,K,V) (5) where: and W O is the parameter matrix, Concat is the connection function, head i is the head of the attention mechanism, MHSA(Q,K,V) is the multi-head attention function, Q is the query vector, K is the key vector, and V is the value vector;

[0085] S2.3a: Represent word-level features as X n Input into the word-level attention mechanism to obtain semantic-based news representation n e , the semantic-based news representation n e The function expression of is shown in formula (6), where the word-level feature is represented by X n Input into the news-level attention layer to obtain the user's semantic interest representation u e , the user's semantic interest representation u e The function expression of is shown in formula (7)

[0086]

[0087] Where: n e is a semantic-based news representation, M is the number of words in a news article, is the attention weight corresponding to the i-th word, is the representation vector of the i-th word, exp is the exponential function, is the query vector, i is the index of the word in the news text, tanh is the activation function, W ω is the trainable parameter matrix, is the jth word vector, is the i-th word vector.

[0088]

[0089] Where: N is the number of energy technology news in the user's historical click records, u e Represents the user's interest in semantics, is the attention weight of the i-th historical click news, is the semantic-based energy technology news representation vector of user u’s historical click records, W e is the parameter matrix, T is the transpose; the CMPredict model consists of a series of gated graph neural networks and news-level attention layers;

[0090] The processing of the CMPredict model includes the following steps:

[0091] S2.1b: Extract the local energy technology news co-occurrence map from the global energy technology news co-occurrence map obtained in S1;

[0092] The processing process of the CMPredict model to extract the local energy technology news co-occurrence map from the global energy technology news co-occurrence map obtained in S1 specifically includes the following steps:

[0093] S2.1.1b: Obtain each piece of news that the user has browsed historically from the user feature sequence data and use it as the initial point set for the local energy technology news co-occurrence graph;

[0094] S2.1.2b: From the global energy technology news co-occurrence graph constructed in S1, for each node in the initial node set, select its news neighbors within the multi-hop range. In each hop selection process, select the top M news neighbors according to the weight of the edge. n Neighbor news, and the weighted ranking is the top M n Add the neighbor news of the local energy technology news co-occurrence graph to the node set of the local energy technology news co-occurrence graph to obtain the node set of the local energy technology news co-occurrence graph. S2.2b: Use the gated graph neural network to encode the local energy technology news co-occurrence graph obtained in S2.1b to obtain the news representation based on collaboration. p , the function expressions of the specific process are shown in formulas (8) to (10):

[0095] n (0) =n e (8)

[0096]

[0097] Where: n (0) is the initialization representation of the graph node, n e is a semantic-based news representation, W g is the parameter matrix, represents the set of adjacent nodes of a node in the local energy technology news co-occurrence graph, l represents the number of layers, represents the representation of the jth node in the lth layer located at the node i position, It represents an intermediate representation of node i at the l+1 layer, GRU (Gated Recurrent Unit) is a gated recurrent unit, is the node representation of the l+1th layer located at the node i;

[0098] S2.3b: Representing collaborative news p Input into the news-level attention layer to obtain the user's collaborative interest representation u p .

[0099] The dual disentanglement module includes a parallel proxy contrast loss module and a cross-space exclusion loss module;

[0100] The processing of the dual de-entanglement module specifically includes the following steps:

[0101] S2.1c: Compute semantically based news representation n e Average representation of user interaction history and collaborative news representation p The average representation of the user interaction history in the semantic space is used to obtain the user's interest proxy in the semantic space and user interest agents in collaborative spaces Its function expressions are shown in formula (11) and formula (12):

[0102]

[0103] Where: N u A news collection of user u's interaction history;

[0104] S2.2c: Leveraging User Interest Proxies in Semantic Space and user interest agents in collaborative spaces and the user's semantic interest representation u e Collaborative interest representation u with the user p Calculating the proxy loss function Its function expression is shown in formula (13):

[0105]

[0106] in: is the proxy loss function, B is the set of users in the training batch, τ is the temperature coefficient, sim() is the cosine similarity, and v is the alias of other users in the same training batch except u;

[0107] S2.3c: Calculate the user's semantic interest representation u e Collaborative interest representation u with the user p The cosine similarity of the user's semantic interest representation u is calculated by comparing the calculated cosine similarity with the set threshold. e Collaborative interest representation u with the user p If the cosine similarity between the two is greater than the set threshold, the input data is penalized by the orthogonal constraint loss function. Its function expression is shown in formula (14):

[0108]

[0109] in: is the orthogonal constraint loss function, u is the user's alias, γ is the threshold, and max is the maximum value;

[0110] The interest stability adaptive fusion module includes a serially connected adaptive window cutting module and an interest stability calculation module;

[0111] The processing process of the interest stability adaptive fusion module specifically includes the following steps:

[0112] S2.1d: Dynamically adjust the window granularity according to the actual number of user behaviors N, and calculate the window size size and the number of cut windows Ta. The function expression of the window size size is shown in formula (15), and the function expression of the number of cut windows Ta is shown in formula (16):

[0113]

[0114] Where: size is the window size, β is the window hyperparameter, N is the number of news in the user interaction history, is the ceiling symbol;

[0115] S2.2d: Use the window size and the number of cut windows Ta obtained in S2.1d to perform adaptive window cutting on the user history records, and obtain the window set {{M1,M2,…,M size}1…{M1,M2,…,M size}T a}, where M2 is the second news vector in the window, and Ta is the number of cut windows;

[0116] S2.3d: Calculate the semantic embedding mean of historical news in each window Its function expression is shown in formula (17):

[0117]

[0118] in: is the semantic embedding mean of historical news in each window, n i is the i-th news vector in the window set, |.| is the absolute value symbol, and t is the window index variable;

[0119] S2.4d: Calculate the cosine similarity sequence s between adjacent windows t , its function expression is shown in formula (18), and the user's interest stability index μ is calculated s , its function expression is shown in formula (19), and the interest stability index α, its function expression is shown in formula (20):

[0120]

[0121] Where: s t is the cosine similarity sequence between adjacent windows, is the semantic embedding mean of historical news in the t+1th window, and Ta is the number of cutting windows;

[0122]

[0123] Where: μ s It is an indicator of user interest stability;

[0124]

[0125] The interest stability index α is used in the prediction score calculation, and the weights of different interest prediction scores are dynamically adjusted according to the interest stability. The function expression of is shown in formula (21).

[0126]

[0127] Then use BPR (Bayesian Personalized Ranking Loss) loss to learn the recommendation task, and the recommendation loss The function expression of is shown in formula (22).

[0128]

[0129] Where B is the set of users in the training batch, φ(·) is the activation function, and are the matching scores of positive and negative samples respectively.

[0130] S3: Use the data constructed in S1 to train the DISNER model and obtain the weights of the DISNER model after training;

[0131] S4: Load the DISNER model weights into the DISNER model, input user behavior data into the DISNER model, and obtain the probability of users clicking on candidate energy technology news.

[0132] Example 1

[0133] This embodiment discloses a method for recommending energy technology news based on disentangled representation learning. The SM Predict model proposed in this embodiment combines word-level attention mechanism and news-level attention mechanism through deep semantic modeling, which can accurately analyze the semantic features of energy technology news and improve the relevance of recommendations; the CMPredict model uses gated graph neural network and news-level attention mechanism to fully explore the potential collaborative relationship between news and enhance the diversity of recommendations. The application of the method of this embodiment solves the problems of single recommended content, insufficient user interest modeling, and unbalanced integration of semantic and collaborative information in existing energy technology news recommendation systems.

[0134] In order to verify the effectiveness of the method of the present invention, based on the above embodiments, a comparative test was conducted between the method of the present invention and several commonly used target detection algorithms:

[0135] When training the model constructed in this application, the hyperparameters are set as follows: the learning rate is 1×10 -3 The batch size is 32, the training epochs are 20, and an early stopping strategy is used. The word embedding dimension is 300; the news and user embedding dimensions are 400; and the maximum length of a news headline is 30. To ensure consistency and stability in the training process, the SMPredict model, CMPredict model, dual disentanglement module, and interest stability adaptive fusion module use the same hyperparameter configuration, including key parameters such as learning rate, batch size, and number of training epochs.

[0136] The hardware parameters used in the model training constructed by the present invention are shown in Table 1:

[0137] Table 1 Hardware configuration for model training

[0138] CPU GPU Memory harddisk Intel Core i7-13700K GeForce RTX 3090 32GB 1TB

[0139] Table 2 Experimental comparison table of the method of the present invention and the existing method

[0140]

[0141]

[0142] Note: AUC (Area Under ROC Curve), the area under the ROC curve, measures the probability that the classification model ranks positive samples before negative samples. The larger the value, the better the classification effect.

[0143] MRR (Mean Reciprocal Rank) is the average reciprocal ranking, which takes the reciprocal of the ranking of the first relevant item in each query result and calculates the average to evaluate the efficiency of the retrieval and recommendation system in finding relevant content.

[0144] NDCG@5 and NDCG@10 (Normalized Discounted Cumulative Gain at 5 / 10) consider the top 5 or 10 items in the recommendation list and evaluate the relevance and ranking quality of the top positions in the recommendation system. Higher values ​​indicate better recommendations.

[0145] DKN stands for Deep Knowledge-aware Network, NPA stands for Neural News Recommendation with Personalized Attention, NAML stands for Neural news recommendation with attentivemulti-view learning, NRMS stands for Neural news recommendation with multi-head self-attention, LSTUR stands for Neural news recommendation with long-and short-term userrepresentations, FIM stands for Fine-grained interest matching for neural newsrecommendation, HieRec stands for Hierarchical User Interest Modeling for PersonalizedNews Recommendatio, GERL stands for Graph Enhanced Representation Learning, GLORY stands for Global-Local news Recommendation sYstem, and GRMR stands for Graph neural newsrecommendation method based on multi-view representation learning

[0146] Table 2 shows an experimental comparison of the present invention with existing methods. The AUC evaluation index represents the model's ability to distinguish between positive and negative samples (positive and negative samples refer to news that the user is interested in and news that the user is not interested in). The MRR evaluation index indicates the degree to which news that the user is interested in is ranked high in the recommendation list. The NDCG evaluation index indicates that the recommendation list contains more news related to the user's interests, and the ranking of these news is more in line with the user's preferences. The four evaluation indicators of the present invention on both datasets are all higher than those of other methods, demonstrating that the present method can provide more accurate and personalized recommendations.

[0147] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0148] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A method for recommending energy technology news based on disentangled representation learning, characterized in that: The following steps are involved: S1: Obtain user behavior datasets and construct user feature information sequence data, obtain records of the order in which users click on energy technology news, and construct a global energy technology news co-occurrence map; S2: Constructing a prediction model DISNER for users clicking on energy technology news. The prediction model DISNER includes the SMPredict model and the CMPredict model, a dual disentanglement module, and an interest stability adaptive fusion module. The SMPredict model includes an embedding layer, a word-level attention layer, and a news-level attention layer connected in series. The word-level attention mechanism includes a multi-head attention mechanism and a word attention mechanism connected in series. The embedding layer is a pre-trained GloVe word vector model. The CMPredict model includes a series of gated graph neural networks and news-level attention layers; The dual disentanglement module includes a parallel proxy contrast loss module and a cross-space exclusion loss module; The interest stability adaptive fusion module includes an adaptive window cutting module and an interest stability calculation module connected in series; S3: Use the data constructed in S1 to train the DISNER model and obtain the weights of the DISNER model after training; S4: Load the DISNER model weights into the DISNER model, input user behavior data into the DISNER model, and obtain the probability of users clicking on candidate energy technology news.

2. The energy technology news recommendation method based on disentangled representation learning according to claim 1 is characterized in that: The customer behavior dataset described in S1 includes user ID, news ID, whether clicked, news title, news text, news keywords, and news category.

3. The energy technology news recommendation method based on disentangled representation learning according to claim 1 is characterized in that: The process of constructing the global energy technology news co-occurrence map described in S1 specifically includes the following steps: S1.1: Collect all users' historical news click records and treat each clicked energy technology news article as a node in the graph; S1.2: Traverse each user's news reading sequence. For any two consecutively read news articles A and B, establish a directed edge from news A to news B. Integrate all users' behavioral data to obtain a global energy technology news co-occurrence graph containing all nodes and directed edges.

4. The energy technology news recommendation method based on disentangled representation learning according to claim 1 is characterized in that: The SMPredict model processing process described in S2 specifically includes the following steps: S2.1a: Input the user feature information sequence data obtained in S1 into the pre-trained GloVe word vector model for initialization, and obtain the word embedding vector sequence x corresponding to the user feature information sequence data n ; S2.2a: Embed the word vector sequence x n Input into the multi-head attention mechanism to obtain word-level feature representation X n ; S2.3a: Represent word-level features as X n Input into the word-level attention mechanism to obtain semantic-based news representation n e , the semantic-based news representation n e The function expression of is shown in formula (1), which represents the news based on semantics as n e Input into the news-level attention layer to obtain the user's semantic interest representation u e , the user's semantic interest representation u e The function expression of is shown in formula (2): Where: n e is a semantic-based news representation, M is the number of words in a news article, is the attention weight corresponding to the i-th word, is the representation vector of the i-th word, exp is the exponential function, is the query vector, i is the index of the word in the news text, tanh is the activation function, W ω is the parameter matrix, is the jth word vector, is the i-th word vector; Where: N is the number of energy technology news in the user's historical click records, u e Represents the user's interest in semantics, is the attention weight of the i-th historical click news, is the semantic-based energy technology news representation vector of user u’s historical click records, W e is the parameter matrix and T is the transpose.

5. The energy technology news recommendation method based on disentangled representation learning according to claim 1 is characterized in that: The processing of the CMPredict model described in S2 specifically includes the following steps: S2.1b: Extract the local energy technology news co-occurrence map from the global energy technology news co-occurrence map obtained in S1; S2.2b: Use the gated graph neural network to encode the local energy technology news co-occurrence map obtained in S2.1b to obtain the collaboration-based news representation n p , the function expressions of the specific process are shown in formulas (3) to (5): n (0) =n e (3) Where: n (0) is the initialization representation of the graph node, n e For semantic-based news representation, W g is the parameter matrix, represents the set of adjacent nodes of node i in the local energy technology news co-occurrence graph, l represents the number of layers, represents the representation of the jth node in the lth layer located at the node i position, represents an intermediate representation of node i at the l+1 layer, GRU is a gated recurrent unit, is the node representation of the l+1th layer at the node i position, n p For collaborative-based news representation, GRU stands for gated recurrent unit; S2.3b: Representing collaborative news p Input into the news-level attention layer to obtain the user's collaborative interest representation u p .

6. The energy technology news recommendation method based on disentangled representation learning according to claim 1 is characterized in that: The processing process of the dual deentanglement module in S2 specifically includes the following steps: S2.1c: Compute semantically based news representation n e Average representation of user interaction history and collaborative news representation p The average representation of the user interaction history in the semantic space is used to obtain the user's interest proxy in the semantic space and user interest agents in collaborative spaces Its function expressions are shown in formula (6) and formula (7): Where: N u A news collection of user u's interaction history; S2.2c: Leveraging User Interest Proxies in Semantic Space and user interest agents in collaborative spaces and the user's semantic interest representation u e Collaborative interest representation u with the user p Calculating the proxy loss function Its function expression is shown in formula (8): in: is the proxy loss function, B is the set of users in the training batch, τ is the temperature coefficient, sim() is the cosine similarity function, and v is the alias of other users in the same training batch except u; S2.3c: Calculate the user's semantic interest representation u e Collaborative interest representation u with the user p The cosine similarity of the user's semantic interest representation u is calculated by comparing the calculated cosine similarity with the set threshold. e Collaborative interest representation u with the user p If the cosine similarity between the two is greater than the set threshold, the input sample data is penalized using the orthogonal constraint loss function. Its function expression is shown in formula (9): in: is the orthogonal constraint loss function, u is the user alias, γ is the threshold, and max is the maximum value.

7. The energy technology news recommendation method based on disentangled representation learning according to claim 1 is characterized in that: The processing process of the interest stability adaptive fusion module described in S2 specifically includes the following steps: S2.1d: Dynamically adjust the window granularity according to the actual number of user behaviors N, and calculate the window size size and the number of windows Ta. The function expression of the window size size is shown in formula (11), and the function expression of the number of cut windows Ta is shown in formula (12): Where: size is the window size, β is the window hyperparameter, N is the number of news in the user interaction history, Ta is the number of cutting windows, is the ceiling symbol; S2.2d: Use the window size and the number of cut windows Ta obtained in S2.1d to perform adaptive window cutting on the user history records, and obtain the window set {{M1,M2,…,M size }1…{M1,M2,…,M size } Ta }, where M2 is the second news vector in the window, and Ta is the number of cut windows; S2.3d: Calculate the semantic embedding mean of historical news in each window Its function expression is shown in formula (13): in: is the semantic embedding mean of historical news in each window, n i is the i-th news vector in the window set, |.| is the absolute value symbol, and t is the window index variable; S2.4d: Calculate the cosine similarity sequence s between adjacent windows t , its function expression, as shown in formula (14), is consistent with the user's interest stability index μ s , its function expression is shown in formula (15), and the interest stability index α, its function expression is shown in formula (16): Where: s t is the cosine similarity sequence between adjacent windows, is the semantic embedding mean of historical news in the t+1th window, and Ta is the number of cutting windows; Where: μ s It is an indicator of user interest stability; 8. The energy technology news recommendation method based on disentangled representation learning according to claim 4 is characterized in that: The process of extracting the local energy technology news co-occurrence map from the global energy technology news co-occurrence map obtained in S1 by the CMPredict model described in S2.1b specifically includes the following steps: S2.1.1b: Obtain each piece of news that the user has browsed historically from the user feature sequence data and use it as the initial point set for the local energy technology news co-occurrence graph; S2.1.2b: From the global energy technology news co-occurrence graph constructed in S1, for each node in the initial node set, select its news neighbors within the multi-hop range. In each hop selection process, select the top M news neighbors according to the weight of the edge. n Neighbor news, and the weighted ranking is the top M n The neighbor news of is added to the node set of the local energy technology news co-occurrence graph to obtain the node set of the local energy technology news co-occurrence graph.

Images