Intention-aware heterogeneous feature collaborative fusion personalized news content recommendation method

Abstract

The invention discloses an intention-aware personalized news content recommendation method based on heterogeneous feature collaborative fusion, and belongs to the technical field of news recommendation. Aiming at the problem that current personalized news recommendation mainly depends on semantic interest representation modeling of a user, heterogeneous semantic features of headline contents and abstract contents of news clicked by the user are considered, and a global semantic correlation graph of a similar relationship of the headline contents and a global semantic correlation graph of a similar relationship of the abstract contents are constructed; a graph convolutional network is used for capturing collaborative title semantic features and collaborative abstract semantic features of similar news, and meanwhile heterogeneous type semantic features of the similar news are subjected to collaborative fusion, so that user interest representation is achieved; and finally, calculating an interest score for the intention news, and generating a personalized recommendation list. According to the method, the collaborative characteristics of similar news titles and abstracts related to semantics are fused, modeling representation of user interests is achieved, the accuracy of collaborative recommendation is improved, and the method has application value in the aspect of personalized content recommendation.

Description

Technical Field

[0001] The present invention belongs to the technical field of news recommendation, and particularly relates to a personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception. Background Art

[0002] Personalized news content recommendation mainly models the interest behavior preferences of users based on the historical clicked news of users, and realizes precise recommendation of content relevance. Traditional news recommendation technologies mainly recommend based on the semantic similarity between homogeneous type title contents of historical clicked news content to meet the personalized needs of users. This collaborative filtering recommendation of homogeneous type title contents is prone to the problem of information cocoons, and due to the length limitation of the title content, the accuracy of the recommended similar semantic content needs to be improved. However, news resources such as titles, abstracts, and category tags can form semantic correlation structure diagrams of different types, reflecting the heterogeneous semantics of different types of resources. Based on these semantic correlation structure diagrams of different types of heterogeneous resources, the potential similar semantics of cross-collaboration between titles and abstracts, and titles and tags can be mined. For example, although there is no explicit semantic association between the titles of two news items, their title and abstract contents may describe the same event and have semantic relevance. By mining the collaborative semantics between the title content and the abstract content, it helps to expand the interest representation of users. Therefore, based on the semantic association information between heterogeneous type resource contents of news, precise collaborative interest information can be provided for users, and the quality of personalized news content recommendation can be improved. Summary of the Invention

[0003] Aiming at the problem that current personalized news recommendation mainly relies on modeling the semantic interest representation of users, the present invention provides a personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception, comprising the following steps:

[0006] Step S1, input the title content and abstract content of the news clicked by the user into the Bert pre-trained language representation model, and learn the semantic feature representation of the title content and the semantic feature representation of the abstract content for any one of the clicked news;

[0007] Step S2, based on the news clicked by the user, calculate the semantic feature similarity between the title contents of the news, and construct a global semantic correlation graph of the title contents of the news clicked by the user;

[0008] Based on the news clicked by the user, calculate the semantic feature similarity between the abstract contents of the news, and construct a global semantic correlation graph of the abstract contents of the news clicked by the user;

[0009] Step S3: Based on the global semantic correlation graph of the title content and the global semantic correlation graph of the abstract content, use the graph convolutional network to learn the collaborative feature representation of the title content and the collaborative feature representation of the abstract content for the collaborative fusion of heterogeneous type features respectively;

[0010] Step S4: Fuse and model the user's click news representation for the collaborative feature representation of the title content and the collaborative feature representation of the abstract content of the news;

[0011] Step S5: Based on all the user's click news representations, model the user's interest content representation;

[0012] Step S6: Calculate the interest scores for the candidate news according to the user's interest content representation, perform negative sampling on the candidate news, and optimize the parameters of the user's interest content representation through the NCE loss function;

[0013] Step S7: According to the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news, use the optimized parameters to calculate the perception score of the user's interest content representation for the intent news, and recommend the Top-k personalized news content for the user.

[0014] Further, in step S1, input the title content and the abstract content of the user's click news into the Bert pre-trained language representation model, and learn the semantic feature representation of the title content and the semantic feature representation of the abstract content for any news. The specific steps are as follows:

[0015] Step 1.1: Use the bidirectional LSTM to input the word embeddings w of the words in the title content of the i-th clicked news respectively, and then extract the single-word semantic representation h of the title content. The calculation process is as follows: i,t for input, and then extract the single-word semantic representation h of the title content. The calculation process is as follows: i,t The calculation process is as follows:

[0016] h i,t = BiLSTM(w i,t ) (1)

[0017] Step 1.2: Since the importance of words in the news title content is different, calculate the importance weight α of the single-word semantic representation h of the title content of the i-th clicked news through the self-attention network: i,t The importance weight α of the single-word semantic representation h of the title content of the i-th clicked news is calculated as follows: i,t :

[0018] α i,t = softmax(v T tanh(Wh i,t + b)) (2)

[0019] In formula (2), W is a learnable parameter, b is a bias, v is an attention query vector, and T represents the transpose;

[0020] Step 1.3: Fuse the semantic representations of all words in the title content of the \(i\)-th clicked news to generate the semantic feature representation \(r\) of the title content of the \(i\)-th news i,t which is:

[0021]

[0022] In formula (3), is the semantic representation of the \(j\)-th word in the title content of the \(i\)-th news and the importance weight of \(l\) is the total number of \(l\) layers;

[0023] Step 1.4: Repeat Steps 1.1 to 1.3 to calculate the semantic feature representation \(r\) of the summary content of the \(i\)-th news i ,a .

[0024] Furthermore, in the said Step S2, based on the news clicked by the user, calculate the semantic feature similarity between the title contents of the news, and construct a global semantic correlation graph of the title contents of the news clicked by the user;

[0025] Based on the news clicked by the user, calculate the semantic feature similarity between the summary contents of the news, and construct a global semantic correlation graph of the summary contents of the news clicked by the user. The specific steps are as follows:

[0026] Step 2.1: According to the semantic feature representations \(r\) of the title contents of the \(p\)-th and \(q\)-th news in the news clicked by the user p,t and \(r\) q ,t , use cosine similarity to calculate the similarity \(sim\) between the two t (p,q) which is:

[0027] sim t (p,q) = (r p,t ·r q,t ) / |r p,t | * |r q,t | (4)

[0028] Step 2.2: For the set of news clicked by the user, calculate the semantic feature similarity between the title contents of any two news. When the similarity is not less than the threshold \(\alpha\), record that there is a semantic correlation edge between the title contents of the two news, denoted as 1, otherwise there is no semantic correlation edge, denoted as 0. The process is expressed as:

[0029]

[0030] Step 2.3: According to the semantic correlation edges of the title contents, construct a global semantic correlation graph \(G\) of the title contents of the news clicked by the user t ;

[0031] Repeat steps 2.1 - 2.3 to construct the global semantic correlation graph G of the user's click on the news abstract content a .

[0032] Furthermore, in step S3, according to the global structural semantic correlation graph of the title content and the global structural semantic correlation graph of the abstract content, use the graph convolutional network to learn the collaborative feature representations of the title content and the abstract content with heterogeneous type features respectively. The specific steps are as follows:

[0033] Step 3.1, based on the global structural semantic correlation graph of the title content and the global structural semantic correlation graph of the abstract content, take the semantic feature representation r of the title content of all clicked news t and the semantic feature representation r of the abstract a , and respectively use them as the initial inputs of the global semantic correlation graph G t of the user's click on the news abstract content and the global semantic correlation graph G a of the user's click on the news abstract content and Use the graph convolutional network for learning to obtain the semantic latent graph embedding representation of the title content at the l + 1 layer and the semantic latent graph embedding representation of the abstract content The update process is as follows:

[0034]

[0035] In Equation (6), is the learnable parameter of the l - th layer convolutional network of the global structural semantic correlation graph of the title content, is the adjacency matrix, is the degree matrix. In Equation (7), is the learnable parameter of the l - th layer convolutional network of the global structural semantic correlation graph of the abstract content; is the semantic latent graph embedding representation of the abstract content; is the semantic latent graph embedding representation of the title content;

[0036] Step 3.2, during the propagation process of the l - th layer of the semantic correlation graph, simultaneously use the gating network vector to fuse the semantic latent graph embedding representation of the title content and the semantic latent graph embedding representation of the abstract content to obtain Realize the collaborative propagation of heterogeneous type features, and the semantic feature representation of the title content of all clicked news. The gating network vector m l is calculated as:

[0037]

[0038] In Equation (8), is the learnable parameter of the gating network vector in the propagation of the l-th layer graph convolutional network; is the bias parameter of the l-th layer graph convolutional network;

[0039] Step 3.3, the title content feature representation and the abstract content feature representation after the collaborative fusion of heterogeneous type features are respectively used to enhance the title content semantic latent graph embedding representation and the abstract content semantic latent graph embedding representation which is defined as:

[0040]

[0041] Step 3.4, the enhanced representations and are used as the input of the next layer of the graph convolutional network, denoted as:

[0042]

[0043] Step 3.5, repeat Step 3.1. After L-layer propagation, the title content collaborative feature representation of the collaborative fusion of heterogeneous type features and the abstract content collaborative feature representation

[0044] Furthermore, in step S4, the news representation is modeled by fusing the title content collaborative feature representation and the abstract content collaborative feature representation of the news. The specific steps are as follows:

[0045] According to the title content collaborative feature representation and the abstract content collaborative feature representation the representation of the i-th news is obtained as the title content collaborative feature representation and the abstract content collaborative feature representation The two are concatenated to model the news representation as:

[0046]

[0047] In formula (13), || is the concatenation operation.

[0048] Furthermore, in step S5, the user interest content representation is modeled based on all the user-clicked news representations. The specific steps are as follows:

[0049] Average pooling is performed on all the user-clicked news representations to obtain the user interest content representation u:

[0050] u = Meanpooling(r1, …, r |V| ) (15)

[0051] In Equation (14), |V| is the number of news items clicked by the user, and Meanpooling is average pooling.

[0052] Furthermore, in step S6, the interest score for the clicked news is calculated based on the user interest content representation, and negative sampling is performed on the clicked news. The parameters of the user interest content representation are optimized using the NCE loss function. The specific steps are as follows:

[0053] Step 6.1: Concatenate the semantic feature representation r c,a of the title content and the semantic feature representation r c,a of the abstract content of the candidate news to obtain the candidate news representation r c ;

[0054] Step 6.2: Based on the user interest content representation u and the candidate news representation r c , calculate the interest score through inner product as

[0055]

[0056] Step 6.3: Calculate the interest scores of the user for the positive sample y + of the clicked news and the negative sample y - according to the negative sampling strategy. Define the loss function using the NCE loss as follows:

[0057]

[0058] In Equation (16), is the click probability score of the s-th positive sample, is the click probability score of the x-th negative sample corresponding to the s-th positive sample, and S is the number of positive samples.

[0059] Furthermore, in step S7, based on the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news, and using the optimized parameters, calculate the perception score of the user interest content representation for the intent news, and recommend the top-k personalized news content for the user. The specific steps are as follows:

[0060] Step 7.1: Fusion the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news to obtain the representation of the intent news;

[0061] Step 7.2: In the set of intent news, calculate the inner product of the user interest content representation and the intent news representation according to step 6.2, perform interest score prediction, and then recommend the top k news content with higher interest scores for the user.

[0062] Compared with the prior art, the present invention has the following advantages:

[0063] The remarkable feature that distinguishes the method provided in the present invention from the existing methods lies in designing a semantic correlation graph of the title content and abstract content of the user's historical clicked news, and learning a user interest representation that synergistically fuses heterogeneous features of the title and abstract. By perceiving the title and abstract of the user's historical clicked news through the intent news, personalized news content recommendation with synergistic fusion of heterogeneous features for intent perception is achieved. Based on the semantic features of the title content and abstract content, the present invention integrates the semantic features of heterogeneous type resources of similar semantic neighbors from the structural aspect, jointly models and enhances the semantic representation learning process of the title content and abstract content, improves the accuracy of news recommendation, and can be widely applied in the personalized news recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the overall model architecture of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0065] To understand the present invention in depth, we will describe it comprehensively and meticulously. However, the present invention has multiple implementation manners and is not limited to the specific examples listed herein. The presentation of these examples aims to deepen the comprehensive understanding of the disclosed content of the present invention.

[0066] As Figure 1 shown, the personalized news content recommendation method with synergistic fusion of heterogeneous features for intent perception of the present invention includes the following steps:

[0067] Step S1, input the title content and abstract content of the news clicked by the user into the Bert pre-trained language representation model, and learn the semantic feature representation of the title content and the semantic feature representation of the abstract content for any clicked news;

[0068] Step 1.1, use bidirectional LSTM to input the word embeddings w i,t of the title content of the i-th clicked news respectively, and then extract the single-word semantic representation h i,t of the title content. The calculation process is as follows:

[0069] h i,t = BiLSTM(w i,t ) (1)

[0070] Step 1.2, since the importance of words in the news title content is different, calculate the importance weight α i,t of the single-word semantic representation h i,t of the title content of the i-th clicked news through the self-attention network:

[0071] α i,t = softmax(v T tanh(Whi,t + b)) (2)

[0072] In formula (2), W is a learnable parameter, b is a bias, v is an attention query vector, and T represents transpose;

[0073] Step 1.3: Fuse the semantic representations of all words in the content of the i-th clicked news title to generate the semantic feature representation r of the content of the i-th news title i,t as follows:

[0074]

[0075] In formula (3), is the semantic representation of the j-th word in the content of the i-th news title is the importance weight, and l is the total number of l layers;

[0076] Step 1.4: Repeat Steps 1.1 to 1.3 to calculate the semantic feature representation r of the content of the summary of the i-th news i ,a .

[0077] Step S2: Based on the news clicked by the user, calculate the semantic feature similarity between news titles, and construct a global semantic correlation graph of the news titles clicked by the user;

[0078] Based on the news clicked by the user, calculate the semantic feature similarity between news summaries, and construct a global semantic correlation graph of the news summaries clicked by the user;

[0079] Step 2.1: According to the semantic feature representations r p,t , r q ,t of the p-th and q-th news titles in the news clicked by the user, calculate the similarity sim t (p, q) between them using cosine similarity as follows:

[0080] sim t (p, q) = (r p,t · r q,t ) / |r p,t | * |r q,t | (4)

[0081] Step 2.2: For the set of news clicked by the user, calculate the semantic feature similarity between any two news titles. When the similarity is not less than the threshold α, it is recorded that there is a semantic correlation edge between the two news titles, denoted as 1; otherwise, there is no semantic correlation edge, denoted as 0. The process is expressed as:

[0082]

[0083] Step 2.3, connect edges according to the semantic relevance of the title content, and construct the global semantic relevance graph G of the user's click on the news title content t ;

[0084] Repeat steps 2.1 - 2.3 to construct the global semantic relevance graph G of the user's click on the news summary content a 。

[0085] Step S3, according to the global semantic relevance graph of the title content and the global semantic relevance graph of the summary content, use the graph convolutional network to learn the collaborative feature representations of the title content and the summary content with heterogeneous type features fused respectively;

[0086] Step 3.1, based on the global structural semantic relevance graph of the title content and the global structural semantic relevance graph of the summary content, take the semantic feature representations r t of all clicked news titles a and the semantic feature representations r t of the summaries respectively as the initial inputs of the global semantic relevance graph G a of the user's click on the news summary content and Use the graph convolutional network for learning to obtain the semantic latent graph embedding representations of the title content at the l + 1 layer and the semantic latent graph embedding representations of the summary content The update process is as follows:

[0087]

[0088] In Equation (6), is the learnable parameter of the l - th layer convolutional network of the global structural semantic relevance graph of the title content, is the adjacency matrix, is the degree matrix. In Equation (7), is the learnable parameter of the l - th layer convolutional network of the global structural semantic relevance graph of the summary content; is the semantic latent graph embedding representation of the summary content; is the semantic latent graph embedding representation of the title content;

[0089] Step 3.2, during the propagation process of the l - th layer of the semantic relevance graph, simultaneously use the gating network vector to fuse the semantic latent graph embedding representation of the title content and the semantic latent graph embedding representation of the summary content to obtain Realize the collaborative propagation of heterogeneous type features. The semantic feature representations of all clicked news titles, and the gating network vector m l is calculated as:

[0090]

[0091] In Equation (8), is the learnable parameter of the gating network vector in the propagation of the l-th layer graph convolutional network; is the bias parameter of the l-th layer graph convolutional network;

[0092] Step 3.3, the title content feature representation and the abstract content feature representation after the collaborative fusion of heterogeneous type features are respectively used to enhance the title content semantic latent graph embedding representation and the abstract content semantic latent graph embedding representation which is defined as:

[0093]

[0094] Step 3.4, the enhanced representations and are used as the input of the next layer of the graph convolutional network, denoted as:

[0095]

[0096] Step 3.5, repeat Step 3.1. After L-layer propagation, the collaborative feature representation of the title content with collaborative fusion of heterogeneous type features and the collaborative feature representation of the abstract content

[0097] Step S4, fuse and model the user click news representation based on the collaborative feature representation of the title content and the collaborative feature representation of the abstract content of the news;

[0098] Based on the collaborative feature representation of the title content and the collaborative feature representation of the abstract content the representation of the i-th news is obtained as the collaborative feature representation of the title content and the collaborative feature representation of the abstract content The two are concatenated to model the news representation as:

[0099]

[0100] In Equation (13), || is the concatenation operation.

[0101] Step S5, model the user interest content representation based on all the user click news representations;

[0102] Perform average pooling on all the user click news representations to obtain the user interest content representation u:

[0103] u = Meanpooling(r1,…,r |V| ) (15)

[0104] In formula (14), |V| is the number of news items clicked by the user, and Meanpooling is average pooling.

[0105] Step S6: Calculate the interest scores for candidate news based on the user interest content representation, perform negative sampling on the candidate news, and optimize the parameters of the user interest content representation through the NCE loss function;

[0106] Step 6.1: Concatenate the semantic feature representation r c,a of the title content and the semantic feature representation r c,a of the abstract content of the candidate news to obtain the candidate news representation r c ;

[0107] Step 6.2: Based on the user interest content representation u and the candidate news representation r c , calculate the interest score through inner product as

[0108]

[0109] Step 6.3: Calculate the interest scores of the user for the positive sample y + of the clicked news and the negative sample y - according to the negative sampling strategy, and define the loss function through the NCE loss as follows:

[0110]

[0111] In formula (16), is the click probability score of the s-th positive sample, is the click probability score of the x-th negative sample corresponding to the s-th positive sample, and S is the number of positive samples.

[0112] Step S7: Calculate the perception score of the user interest content representation for the intent news based on the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news, and use the optimized parameters, and recommend the top-k personalized news content for the user;

[0113] Step 7.1: Fuse the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news to obtain the representation of the intent news;

[0114] Step 7.2: In the set of intent news, calculate the inner product of the user interest content representation and the intent news representation according to Step 6.2 to predict the interest score, and then recommend the top k news content with higher interest scores for the user.

[0115] To verify the effectiveness of the present invention, we conducted experiments on the news dataset (MIND (msnews.github.io)), which contains users' personalized historical click information and candidate news information. User-related news information includes log data of the title content and abstract content of the news browsed and clicked by users. User log data contains the candidate intention news of users and the interest tags for the candidate intention news, where 1 indicates that the user is interested in the news and 0 indicates that the user is not interested in the news. The dataset information is shown in Table 1:

[0116] Table 1 Dataset situation

[0117]

[0118] The heterogeneous type resources we used have two aspects, including titles and abstracts. The evaluation metrics are AUC, MRR, and NDCG@5. To verify the effectiveness and advancement of the technical solution proposed by the present invention, several existing recommendation prediction model methods are selected for comparison: DKN, NAML, NPA, and FIM. The experimental results are shown in Table 2:

[0119] Table 2 Experimental results

[0120]

[0121] As can be seen from the results in Table 2, when the technical solution of the present invention is used for the interest prediction of personalized content news recommendation for users, it can obtain recommendation results with higher accuracy and ranking metrics than the existing methods.

[0122] The content not described in detail in the specification of the present invention belongs to the prior art well-known to those skilled in the art. Although the illustrative specific embodiments of the present invention are described above for the understanding of those skilled in the art in the technical field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those ordinary skilled in the technical field, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

Claims

1. A personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception, characterized in that It includes the following steps: Step S1: Input the title content and abstract content of the news clicked by the user into the Bert pre-trained language representation model to learn the semantic feature representation of the title content and the semantic feature representation of the abstract content for any clicked news. Step S2: Based on the news clicked by the user, calculate the semantic feature similarity between the title contents of the news, and construct a global semantic correlation graph of the title contents of the news clicked by the user. Based on the news clicked by the user, calculate the semantic feature similarity between the abstract contents of the news, and construct a global semantic correlation graph of the abstract contents of the news clicked by the user. Step S3: According to the global semantic correlation graph of the title content and the global semantic correlation graph of the abstract content, use the graph convolutional network to learn the collaborative feature representation of the title content and the collaborative feature representation of the abstract content with the collaborative fusion of heterogeneous type features respectively. Step S4: Fuse and model the collaborative feature representation of the title content and the collaborative feature representation of the abstract content of the news to represent the news clicked by the user. Step S5: Based on all the representations of the news clicked by the user, model the representation of the user's interested content. Step S6: Calculate the interest scores for the candidate news according to the representation of the user's interested content, perform negative sampling on the candidate news, and optimize the parameters of the representation of the user's interested content through the NCE loss function. Step S7: According to the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intended news, use the optimized parameters to calculate the perception score of the representation of the user's interested content for the intended news, and recommend the Top-k personalized news content for the user.

2. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 1, characterized in that: In the said Step S1, input the title content and abstract content of the news clicked by the user into the Bert pre-trained language representation model to learn the semantic feature representation of the title content and the semantic feature representation of the abstract content for any clicked news. The specific steps are as follows: Step 1.1, use bidirectional LSTM to respectively input the word embeddings w of the content of the i-th clicked news title, and then extract the semantic representation h of the words in the title content. The calculation process is as follows: i,t i,t ​​ h i,t = BiLSTM(w i,t ) (1) Step 1.2, calculate the importance weight α of the semantic representation h of the words in the title content of the i-th clicked news through the self-attention network i,t as i,t : α i,t = softmax(v T tanh(Wh i,t + b)) (2) In formula (2), W is a learnable parameter, b is a bias, v is an attention query vector, and T represents transpose. Step 1.3, fuse the semantic representations of all words in the content of the news title clicked for the i-th article to generate the semantic feature representation r of the news title content for the i-th article i,t It is as follows: In formula (3), is the semantic representation of the j-th word in the title content of the i-th news is the importance weight, and l is the total number of l layers; Step 1.4, repeat Steps 1.1 to 1.3 to calculate the semantic feature representation r of the summary content of the i-th news i,a 。 3. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 2, wherein: In the said Step S2, based on the news clicked by the user, calculate the semantic feature similarity between the title contents of the news, and construct a global semantic correlation graph of the title contents of the news clicked by the user. Based on the news clicked by the user, calculate the semantic feature similarity between the abstract contents of the news, and construct a global semantic correlation graph of the abstract contents of the news clicked by the user. The specific steps are as follows: Step 2.1, represent the semantic features r of the titles of the p-th and q-th news items clicked by the user p,t and r q,t , and use cosine similarity to calculate the similarity sim t (p, q) as follows: sim t (p, q) = (r p,t ·r q,t ) / |r p,t | * |r q,t | (4) Step 2.2: For the set of news clicked by the user, calculate the semantic feature similarity between the title contents of any two news. When the similarity is not less than the threshold α, it is recorded that there is a semantic correlation edge between the title contents of the two news, denoted as 1, otherwise there is no semantic correlation edge, denoted as 0. The process is expressed as: Step 2.3, connect edges according to the semantic relevance of the title content to construct a global semantic relevance graph G for the user to click on the news title content t ; Repeat steps 2.1 - 2.3 to construct the global semantic correlation graph G for the user to click on the news abstract content a 。 4. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 3, characterized in that: In the said Step S3, according to the global semantic correlation graph of the title content and the global semantic correlation graph of the abstract content, use the graph convolutional network to learn the collaborative feature representation of the title content and the collaborative feature representation of the abstract content with the collaborative fusion of heterogeneous type features respectively. The specific steps are as follows: Step 3.1, based on the global structural semantic correlation graph of the title content and the global structural semantic correlation graph of the abstract content, represent the semantic features r of the title content of all click news t and the semantic feature representation r of the abstract a , and use them as the global semantic correlation graph G of the abstract content of the news clicked by the user t and the global semantic correlation graph G of the abstract content of the news clicked by the user a as the initial input and Use the graph convolutional network for learning to obtain the semantic latent graph embedding representation of the l+1 layer of the title content and the semantic latent graph embedding representation of the abstract content The update process is as follows: In formula (6), is the learnable parameter of the l-th layer convolutional network of the global structural semantic correlation graph of the title content, is the adjacency matrix, is the degree matrix. In formula (7), is the learnable parameter of the l-th layer convolutional network of the global structural semantic correlation graph of the abstract content; is the semantic latent graph embedding representation of the abstract content; is the semantic latent graph embedding representation of the title content; Step 3.2, during the propagation of the l-th layer of the semantic relevance graph, the semantic potential graph of the title content is embedded and represented by using the gated network vector and the semantic potential graph of the abstract content is embedded and represented are fused to obtain The co-propagation of heterogeneous type features represents the semantic features of the title content of all clicked news, and the gated network vector m l is calculated as: In Equation (8), is the learnable parameter of the gating network vector in the propagation of the l-th layer graph convolutional network; is the bias parameter of the l-th layer graph convolutional network; Step 3.3, the title content feature representation and the abstract content feature representation after the collaborative fusion of heterogeneous type features are respectively subjected to latent graph embedding representation of the title content semantics and latent graph embedding representation of the abstract content semantics for enhancement, which is defined as: Step 3.4, use the enhanced representation and as the input of the next layer of the graph convolutional network, denoted as: Step 3.5, repeat Step 3.

1. After L-layer propagation, obtain the collaborative feature representation of the title content with heterogeneous type features synergistically fused and the collaborative feature representation of the abstract content 5. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 4, wherein: In the said Step S4, fuse and model the collaborative feature representation of the title content and the collaborative feature representation of the abstract content of the news to represent the news clicked by the user. The specific steps are as follows: Collaborative feature representation according to the title content and collaborative feature representation of the abstract content Obtain the representation of the i-th news as the collaborative feature representation of the title content and collaborative feature representation of the abstract content Concatenate the two to model the news representation as: In formula (13), || is a concatenation operation.

6. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 5, wherein: In the said Step S5, based on all the representations of the news clicked by the user, model the representation of the user's interested content. The specific steps are as follows: Average pooling is performed on all user click news representations to obtain the user interest content representation u: u = Meanpooling(r1, …, r |V| ) (15) In Equation (14), |V| is the number of user click news, and Meanpooling is average pooling.

7. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 6, characterized in that: In step S6, the interest score for candidate news is calculated based on the user interest content representation, and negative sampling is performed on the candidate news. The parameters of the user interest content representation are optimized through the NCE loss function. The specific steps are as follows: Step 6.1, represent the semantic features r of the title content of the candidate news c,a and the semantic feature representation r of the abstract content c,a , and splice them to obtain the candidate news representation r c ; Step 6.2, based on the user interest content representation u and the candidate news representation r c , calculate the interest score through the inner product as Step 6.3, calculate the interest scores of the user for the positive sample y of the clicked news + and the negative sample y - using the negative sampling strategy, and define the loss function through the NCE loss as follows: In formula (16), is the click probability score of the s-th positive sample, is the click probability score of the x-th negative sample corresponding to the s-th positive sample, and S is the number of positive samples.

8. The personalized news content recommendation method for collaborative fusion of heterogeneous features with intention perception according to claim 7, characterized in that: In step S7, based on the semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news, and using the optimized parameters, the perception score of the user interest content representation for the intent news is calculated, and the top-k personalized news content is recommended for the user. The specific steps are as follows: Step 7.1: The semantic feature representation of the title content and the semantic feature representation of the abstract content of the intent news are fused to obtain the representation of the intent news; Step 7.2: In the set of intent news, the inner product of the user interest content representation and the intent news representation is calculated according to step 6.2 to predict the interest score, and then the top k news content with higher interest scores is recommended for the user.

Images