Method for recommending content by fusing social relations of non-heritage successors

Abstract

The application discloses a graph convolution content recommendation method fusing non-inheritance successor social relations, and comprises the following steps: performing feature extraction on users, non-inheritance contents and successors to obtain user interest preference features, non-inheritance content features and successor features; based on the user interest preference features, the non-inheritance content features and the successor features, a vector inner product method is used to obtain a similarity value of a user and non-inheritance content to be recommended; and according to the similarity value, non-inheritance content is recommended to the user. Through the recommendation method, the accuracy of non-inheritance content recommendation can be improved.

Description

Technical Field

[0001] This invention belongs to the field of intangible cultural heritage dissemination technology, and in particular relates to a graph convolutional content recommendation method that integrates the social relationships of intangible cultural heritage inheritors. Background Technology

[0002] The application of intelligent recommendation algorithms to personalize intangible cultural heritage (ICH) content on digital content platforms, thereby promoting the inheritance, development, and industrialization of ICH, has been elaborated and discussed by many scholars in their published papers. Some scholars have mentioned that using intelligent algorithms to accurately push ICH-related digital content can improve ICH's accessibility, compensate for the shortcomings in the scope of ICH information dissemination, and to some extent stimulate the interest of ordinary users. For example, using classic collaborative filtering methods, ICH content that users are likely to browse can be recommended based on similarity. Furthermore, the social networks formed by user interactions and content interactions on digital content platforms can utilize the trust between nodes as a dissemination network for ICH content. Other scholars believe that in the dissemination of ICH content, it is necessary to combine intelligent recommendation algorithms with audience characteristics and the inherent characteristics of ICH itself to conduct broad and accurate recommendations of ICH content in order to expand the influence of ICH digital content. In the dissemination of intangible cultural heritage (ICH), inheritors who publish ICH content play a crucial role. They are guardians of traditional ICH skills, possessing rich experience and knowledge, and are able to impart the essence of ICH culture, ensuring its authentic transmission. In the digital age, the role of inheritors is even more important. They can disseminate ICH culture to users through online platforms, enhancing its global influence. Currently, to achieve effective dissemination of ICH content, empathy has become an effective content creation method and marketing strategy: while showcasing ICH projects, the creators' (mainly ICH inheritors) daily joys and sorrows are incorporated, making the projects more relatable and resulting in diverse styles of ICH content creation by inheritors. Simultaneously, inspired by the good adaptability of ICH to oral transmission in traditional dissemination, user social relationships are also an important part of the digital content dissemination of ICH. User social relationships can transmit ICH information through social media and other channels, especially through sharing and forwarding, expanding the reach of ICH content and enhancing the exchange and inheritance of ICH culture.

[0003] Numerous studies have theoretically explored the benefits of applying recommendation algorithms to the dissemination of intangible cultural heritage (ICH) content, as well as the auxiliary information sources that can be used in such applications. However, there are relatively few studies that conduct specific experiments and technically demonstrate the feasibility and effectiveness of these methods. This invention summarizes and analyzes existing theoretical research on ICH content recommendation and, in conjunction with relevant research in the field of recommendation algorithms, designs a methodological model suitable for ICH content recommendation.

[0004] In existing recommendation algorithm research, user historical interaction data remains the preferred choice for most scholars. However, faced with the explosive growth of data volume, data sparsity has become a hot topic in recommendation method research. Adding users' social data as auxiliary information to model construction is one way to address the data sparsity problem. The SoRec model is considered the first social network recommendation model based on matrix factorization, combining users' direct and indirect social relationships to derive user characteristics, which are then used as the basis for recommendations. The SocialMF model also uses matrix factorization, assuming that users possess certain latent interests, which are influenced by the interest characteristics of their trusted friends. Therefore, even without rating data, it learns user characteristics for recommendations by modeling the trust propagation between users and their friends. Matrix factorization has been applied in recommendations based on social data, but it can only extract simple linear information. Given that social data between users and user-content interaction data can both be represented using graph structures, and that graph neural network (GNN) algorithms perform well in graph data analysis and mining, social recommendation methods based on graph neural networks have become a research hotspot.

[0005] In the GraphRec recommendation model, a graph convolutional neural network-based approach is used to construct a user model using user-content interaction data and social data between users. This user model is then fused with the content model and rating model, achieving good recommendation results. Considering that different friends have varying levels of trust with target users in their social relationships, a point not typically taken into account by general graph convolutional network-based recommendation models, researchers have incorporated an attention mechanism into the model, namely Graph Attention Network (GAT). Weights are assigned in the process of obtaining target user features through friend features to obtain more accurate user features and recommendation results. In the NGCF model, user-content interaction data is integrated into the embedding to model the higher-order connectivity of the user-content bipartite graph. Spatial graph convolution is also applied in similarity calculation. As a lightweight version of the NGCF model, LightGCN employs a simple yet effective user-item interaction matrix weighting strategy, which not only reduces model complexity but also demonstrates better recommendation performance than NGCF on large-scale datasets. Unlike other recommendation methods, the GNN-DSR model believes that user interaction with content not only reflects user interests and preferences, but also the attractiveness of the content to users, and constructs content features through user characteristics.

[0006] In summary, most of the above models only model users and content to complete recommendations. To adapt to the special scenario of recommending intangible cultural heritage content, they still lack the impact of the characteristics of intangible cultural heritage inheritors on content recommendation. Summary of the Invention

[0007] To address the aforementioned technical issues, this invention proposes a graph convolutional content recommendation method (GCHS) that integrates the social relationships of intangible cultural heritage inheritors. By incorporating the social relationships of inheritors into the recommendation model, the accuracy of intangible cultural heritage content recommendations can be improved.

[0008] To achieve the above objectives, this invention provides a graph convolutional content recommendation method that integrates the social relationships of intangible cultural heritage inheritors, including:

[0009] A graph convolutional content recommendation method that integrates the social relationships of intangible cultural heritage inheritors is characterized by including:

[0010] Feature extraction is performed on users, intangible cultural heritage content, and inheritors to obtain user interest and preference features, intangible cultural heritage content features, and inheritor features;

[0011] Based on the user interest and preference features, intangible cultural heritage content features, and inheritor features, the similarity value between the user and the intangible cultural heritage content to be recommended is obtained using the vector inner product method.

[0012] Based on the similarity value, intangible cultural heritage content is recommended to the user.

[0013] Optionally, obtaining the user interest preference features, intangible cultural heritage content features, and inheritor features includes:

[0014] Obtain the required initial user characteristics and initial content characteristics respectively;

[0015] Based on the acquired initial user features and initial content features, a multi-layer spatial graph convolutional neural network with attention mechanism as the aggregation method is used to obtain the user features and content features of each layer.

[0016] The user interest and content features of each layer are connected to obtain the final user interest and preference features and intangible cultural heritage content features. Then, the inheritor features are obtained by using the average aggregation method.

[0017] Optionally, obtaining the user's initial characteristics includes:

[0018] Define the user set as U = {u1, u2, ..., u...} m The set of intangible cultural heritage content is I = {i1, i2, ..., i}. n}, where m and n represent the total number of users and the total number of content items, respectively; the user-content interaction matrix is ​​defined as R. m*n Using r ij Represents matrix R m*n The element in the i-th row and j-th position, r ij =1 indicates that user i has interacted with content j, r ij = 0 indicates that user i and content j did not interact; for the user-content interaction matrix Rm*n One-hot encoding is performed, and the initial user characteristics of the corresponding user nodes are obtained by comparing them according to the index order.

[0019] Optionally, the user characteristics of each layer are:

[0020]

[0021] in, Let W1 and W2 represent the feature vector of user u after the (l+1)th iteration, where W1 and W2 are both trainable weight vectors, and N... u Let α represent the set of first-order neighbors of user u in the social graph, ⊙ represent the element-wise multiplication of two vectors, and α represent the set of first-order neighbors of user u in the social graph. uu This represents the attention weights of user u's first-order neighbor user v when aggregating the feature vector of user u, where v represents user u's first-order neighbor user v. Let v represent the feature vector of user v after the l-th iteration.

[0022] Optionally, the user interest preference features are:

[0023]

[0024] Among them, e u Let be the feature vector of user u's interests and preferences, and l be the number of iteration layers.

[0025] Optionally, the initial features of the content are:

[0026]

[0027] in, As the initial feature of the content, N i This represents the set of user nodes that interact with content i;

[0028] The content characteristics of each layer are as follows:

[0029]

[0030] in, This represents the feature vector of content i after the (l+1)th iteration. This represents the feature vector of user u after the l-th iteration. Then, it represents the feature vector of content u after l iterations, ⊙ represents the element-wise multiplication of two vectors, and α iu The attention weights of user u, the first-order neighbor of aggregated content i, are used to represent the feature vector of i.

[0031] The final characteristics of the intangible cultural heritage content are as follows:

[0032]

[0033] Among them, e i is the final feature vector of intangible cultural heritage content, and l is the number of iteration layers.

[0034] Optionally, the inheritor characteristics are obtained through an average aggregation method based on the characteristics of the intangible cultural heritage content published by the inheritor;

[0035] The characteristics of the inheritor are as follows:

[0036]

[0037] Among them, e h Inheritor feature vector, N h The term "e" represents the collection of content published by the inheritor "h". i Intangible cultural heritage content released by inheritors.

[0038] Optionally, the similarity value is:

[0039]

[0040] Where Sim(u, i) represents the similarity value between the content to be recommended i and the target user u. Let e ​​be the feature vector of the intangible cultural heritage content i to be recommended. u Let u be the feature vector of the target user. Let i be the feature vector of the publisher of the intangible cultural heritage content to be recommended.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] Integrating the characteristics of intangible cultural heritage inheritors into the recommendation model increases auxiliary information in the recommendation process, enabling it to adapt to the special scenario of recommending intangible cultural heritage content and improve the accuracy of model recommendations.

[0043] The characteristics of intangible cultural heritage inheritors are obtained through their social relationships, and an attention mechanism is applied to ensure the accuracy of feature extraction. Attached Figure Description

[0044] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0045] Figure 1 This is a schematic diagram of the recommended method flow according to an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the framework of the graph convolutional content recommendation method that integrates the social relationships of inheritors in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of interactive data in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram comparing the recommendation accuracy of each comparison model in the embodiments of the present invention with that of GCHS;

[0049] Figure 5 This is a schematic diagram comparing the recommendation recall rates of various comparison models in this embodiment of the invention with those of GCHS;

[0050] Figure 6 This is a schematic diagram comparing the F-values ​​of the various comparison models in this embodiment of the invention with the recommended results of GCHS;

[0051] Figure 7 This is a schematic diagram comparing the various comparative models of this invention with the recommended normalized loss cumulative gain of GCHS;

[0052] Figure 8 This is a schematic diagram comparing the recommended accuracy of three ablation experiment models according to an embodiment of the present invention.

[0053] Figure 9 This is a schematic diagram comparing the recommended recall rates of three ablation experiment models according to an embodiment of the present invention.

[0054] Figure 10 This is a schematic diagram comparing the recommended F-values ​​of three ablation experiment models according to an embodiment of the present invention.

[0055] Figure 11 This is a schematic diagram comparing the recommended normalized loss cumulative gain of three ablation experiment models in an embodiment of the present invention. Detailed Implementation

[0056] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0058] The characteristics of users attracted by the content created by inheritors (i.e., the implicit social relationships of the inheritors) can reflect their own characteristics. Using these characteristics as one of the bases for matching users and content similarity can improve the accuracy of recommendations. Referring to existing research, this invention uses spatial graph convolution in feature construction and introduces an attention mechanism in the aggregation layer to obtain more accurate features of inheritors, users, and content for similarity calculation, aiming to more accurately predict user interests and recommend potentially interesting intangible cultural heritage content.

[0059] The overall approach of the intangible cultural heritage content recommendation method proposed in this embodiment is as follows: Figure 1 As shown, user features and intangible cultural heritage (ICH) content features are obtained by using the initial features of users and the initial features of ICH content, respectively. Then, the features of the inheritors are obtained by using the features of users who interact with the ICH-related content created by the inheritors. The three features are then compared to obtain the final user-content similarity result, and the final Top-N recommendation result is obtained by sorting them in descending order.

[0060] The recommended method's detailed framework is as follows: Figure 2 As shown, the initial user and content features are first obtained from the user-content interaction graph and the user-user social graph. Then, the user and content features of each layer are obtained through a multi-layer spatial graph convolutional neural network GraphSAGE with an attention mechanism as the aggregation method. The features of each layer are connected to obtain the final user interest features and intangible cultural heritage content features. Then, the inheritor features are obtained using the Mean aggregation method. Finally, the inheritor features are added to the user-content similarity calculation to obtain the comprehensive similarity. The final Top-N recommendation list is obtained by sorting the comprehensive similarity in descending order.

[0061] The GraphSAGE used in the model differs from traditional graph convolution; its main idea is neighbor sampling and aggregation. After sampling neighbor information, the features of the target node are learned through aggregation operations. Specifically, the operations are as follows:

[0062]

[0063] in, For the (l+1)th layer embedding of node u, W (l+1) The weights of layer 1+1, Let l be the l-th layer embedding of node u. Let σ be the l-th layer embedding of node u, and N be the nonlinear transformation. u Let be the set of neighboring nodes of node u, where u is the neighboring node of node u, and aggregate is the aggregation method.

[0064] S1. Feature Extraction

[0065] In this embodiment, the recommendation model requires features from three parties—users, intangible cultural heritage content, and inheritors—to calculate the final similarity. The inheritor's features are obtained by aggregating the features of users who have interacted with the intangible cultural heritage content they publish, i.e., the inheritor's implicit social relationships on the platform. Therefore, the feature extraction module is divided into three parts: user interest and preference features, intangible cultural heritage content features, and inheritor features.

[0066] (1) User interest and preference characteristics

[0067] Users with similar interests often exhibit the same likes and dislikes towards the same things. To understand the interest characteristics of a target user, one can start by identifying users with similar interests.

[26] This measures the intimacy between a target user and their implicit social friends (both users have interacted with the same content), i.e., the similarity of their interests, and uses this as a reference when obtaining the target user's interest characteristics through social relationship aggregation.

[0068] Define the user set as U = {u1, u2, ..., u...} m The set of intangible cultural heritage content is I = {i1, i2, ..., i}. n}, where m and n represent the total number of users and the total number of content items, respectively; the user-content interaction matrix is ​​defined as R. m*n Using r ij Represents matrix R m*n The element in the i-th row and j-th position, r ij =1 indicates that user i has interacted with content j, r ij = 0 indicates that user i and content j did not interact. According to this rule, the user-content interaction matrix R... m*n One-hot encoding is performed, and the initial preference feature vectors of the corresponding user nodes are obtained by comparing them according to the index order. Regarding user social relationships, user nodes that have interacted with the same content are connected to obtain a social graph, and it is determined that the two have a certain implicit social relationship. It should be noted that the initial user preference feature vectors derived from the interaction matrix are also applicable to the social graph, serving as user feature vectors in the social graph. Within the social relationships of the same user, there is always a strength issue. Some users have more similar interests, while others have less similarity. Therefore, in the aggregation of social relationships, weights are used to distinguish friends with different levels of interest similarity. This embodiment uses an attention mechanism to obtain weights in the aggregation layer; therefore, the expression for constructing the feature vector of user u is:

[0069]

[0070] In this formula, W1 and W2 are both trainable weight vectors. Let represent the feature vector of user u after the l-th iteration. and Similarly. This operation is to add the original features of the target node with a certain weight after aggregating the features of neighboring nodes, in order to appropriately avoid losing the node's own features in subsequent operations. u This represents the set of first-order neighbors of user u in the social relationship graph; Let represent the feature vector of user v after the l-th iteration. ⊙ represents the element-wise multiplication operation between the two vectors, which makes the features of neighboring nodes more similar to those of the target node. α uv The attention weights of user u's first-order neighbor user v when aggregating the feature vector of user u are expressed as follows:

[0071]

[0072]

[0073] After l layers of operations, the feature vectors obtained from each layer are concatenated to obtain the final user feature vector e. u .

[0074]

[0075] S2. Characteristics of Intangible Cultural Heritage Content

[0076] Previous research has shown that the attractiveness of content can be analyzed from user groups who like and comment on it, indicating which user group prefers it. The content's own characteristics can be derived from the comprehensive characteristics of its audience. Therefore, in this embodiment, the characteristics of intangible cultural heritage content are obtained by aggregating the interest features of users who have interacted with it. In the user-content interaction graph, the user feature vector is the same as the user feature vector in the social graph. First, the Mean aggregation method is used to obtain the initial characteristics of the target node through the features of its neighboring nodes (i.e., user nodes that have liked the content).

[0077]

[0078] Where, N i This represents the set of user nodes that interact with content i.

[0079] by Let i represent the feature vector of content i after the (l+1)th iteration, and let its expression be:

[0080]

[0081] Similarly, in calculating the user preference feature vector, W1 and W2 are both trainable weight vectors. This represents the feature vector of user u after the l-th iteration. This represents the feature vector of content u after l iterations, and ⊙ represents the element-wise multiplication of the two vectors. α iu The attention weights of user u, a first-order neighbor of aggregated content i, are represented by the feature vector of i. The expression for this weight is:

[0082]

[0083]

[0084] Similar to the construction of user feature vectors, after l layers of operations, the feature vectors obtained from each layer are concatenated to obtain the final intangible cultural heritage content feature vector e. i .

[0085]

[0086] S3. Characteristics of Successors

[0087] Even among inheritors of similar or identical intangible cultural heritage projects, their creation of digital content such as videos and texts can be influenced by factors such as personality and persona, resulting in different user groups. The users attracted by inheritors mainly fall into two categories: followers who follow their accounts (defined as explicit social friends) and users who have positively interacted with their content (defined as implicit social friends). Inspired by previous research on social recommendation, if users only follow an inheritor's account but haven't interacted with their work, such users have no practical value in constructing the inheritor's characteristics; those explicit social friends who have interacted with the inheritor's content are already included in the implicit social circle. Therefore, in constructing inheritor characteristics, only implicit social friends are considered. Since the characteristics of intangible cultural heritage content are constructed through the characteristics of interacting users, the characteristics of inheritors are directly constructed based on the characteristics of the intangible cultural heritage content they publish. This embodiment uses an average aggregation method to obtain the inheritor's characteristics, the expression of which is:

[0088]

[0089] Among them, e h Inheritor feature vector, N h The term "e" represents the collection of content published by the inheritor "h". i Feature vectors of intangible cultural heritage content released by inheritors.

[0090] S4. Similarity Calculation

[0091] Similarity measurement is the calculation of the similarity between the intangible cultural heritage content to be recommended and the user. Given previous research demonstrating the influence of inheritors on the dissemination and recommendation of intangible cultural heritage content, this embodiment incorporates inheritor characteristics into the similarity calculation to optimize the result. In the recommendation field, similarity calculation methods include cosine similarity, Manhattan distance, vector dot product, and Dice coefficient. Considering the characteristics of the recommendation method in this embodiment, the vector dot product is used to calculate the similarity between the user and the intangible cultural heritage content to be recommended, providing a basis for subsequent recommendations. Therefore, the similarity calculation expression is:

[0092]

[0093] Sim(u, i) represents the similarity value between the content to be recommended i and the target user u, e i Let e ​​be the feature vector of the intangible cultural heritage content i to be recommended. u Let e ​​be the feature vector of the target user u. h Let be the feature vector of the publisher (i.e., inheritor) of the intangible cultural heritage content i to be recommended. After obtaining the similarity value, when recommending content to users, the content is sorted from high to low according to the similarity value, and the Top-N content is selected to recommend to users.

[0094] S5. Model Training

[0095] The model is optimized using the Bayesian personalized ranking loss function, which is widely used in ranking and recommendation tasks. Its expression is as follows:

[0096]

[0097] In this formula, R + Positive feedback data, meaning that the user has interacted with the content, R - This is negative feedback data, and its meaning is the same as R. + Conversely, σ is the sigmoid function, γ is the balance coefficient of the regularization term, and θ represents the quadratic regularization term for all parameters in the model.

[0098] S6. Experiments and Analysis

[0099] S6.1. Data Preparation

[0100] Sina Weibo, as a large-scale content creation platform, has attracted many intangible cultural heritage (ICH) inheritors and teams to join the platform and create ICH-related content. Simultaneously, the platform also hosts numerous related topics and activities such as "Amazing Chinese Intangible Cultural Heritage" and "ICH Celebrates the New Year," providing rich ICH-related content and user-content interaction information. This embodiment uses a web crawler to start with the accounts of ICH inheritors on the platform. It selects 100 certified ICH inheritor accounts and collects the IDs of no more than 50 ICH-related content posts published between May 1st and October 31st, 2022. It also obtains the IDs of no more than 500 users who interacted with the target content. Based on these data acquisition and filtering rules, the final experimental data is obtained. See Table 1 for details.

[0101] Table 1 Description of the experimental dataset

[0102]

[0103] To more intuitively display the experimental data used, heatmaps were used to visualize user-content interaction data: light colors represent no interaction, and dark colors represent interaction. See [link to partial interaction data visualization] for details. Figure 3 As shown.

[0104] S6.2. Experimental Environment

[0105] All experiments in this embodiment were run on a host with an RTX 3050 graphics card with 4GB of RAM, 16GB of RAM, and an Intel i5-12450H CPU. The programming language used was Python 3.9, and the deep learning framework was Tensorflow.

[0106] S6.3. Evaluation Indicators

[0107] When recommending intangible cultural heritage content to users, in addition to considering the overall accuracy of judging the content that interests the user, it is also necessary to consider whether the content of interest can be presented to the user first. Therefore, in this embodiment, the evaluation of the recommendation model's performance is measured by two aspects: classification accuracy and ranking accuracy. Classification accuracy is measured by precision, recall, and F-measure; ranking accuracy is measured by normalized depreciation cumulative gain (NDCG).

[0108] Let P be the number of contents that a user is interested in, N be the number of contents that a user might be interested in, and TP be the number of contents that a user is actually interested in from the prediction results. Then the formulas for calculating recall and precision are:

[0109]

[0110]

[0111] The F-score is a harmonic average of precision and recall, and its calculation formula is as follows:

[0112]

[0113] Normalized loss cumulative gain is an indicator that takes into account the ranking of recommended content within the recommendation list. When sorted in descending order, the higher the ranking, the higher the score. Its expression is:

[0114]

[0115] In this formula, IDCG is the maximum DCG value under ideal conditions, and DCG is the cumulative gain due to loss. Its expression is:

[0116]

[0117] Where, r i The relevance score is the similarity value in the model of this embodiment, N is the number of contents recommended to the target user, and i is the descending order of the contents in the recommendation list.

[0118] S6.4. Experimental Setup and Result Analysis

[0119] In the experiment, for ease of operation, integers from 0 to n were used to number users and content respectively, and the data was divided into training, validation, and test sets in a 7:1:2 ratio for model training and recommendation performance testing. For ease of description, the model in this embodiment is named GCHS. First, ablation experiments were used to examine the heirloom feature component incorporated into GCHS and the attention mechanism used in feature aggregation, and then it was compared with previous similar models. Referring to previous related research, N was set to 5, 10, 15, 20, 25, and 30 for Top-N recommendations. In each experiment, the model's recommendation performance was measured from two aspects: recommendation accuracy and ranking accuracy.

[0120] (1) Comparative Experiment

[0121] After ablation experiments validated the effectiveness of the key components of the model and yielded positive results, to assess the level of the model's recommended performance, a comparative experiment was conducted between the complete GCHS model and several baseline models under the same experimental conditions. A brief overview of the comparative models used is shown in Table 2.

[0122] Table 2 Comparison of experimental models

[0123]

[0124] Using the comparative models mentioned above, recommendation tests were conducted on the dataset of this study. Precision, recall, F-score, and normalized depreciation cumulative gain were calculated for different N values ​​and presented in line graphs. The results are shown in [see figure]. Figures 4-7 As shown:

[0125] As can be seen from the results, the Precision, Recall, and NDCG scores of the model in this embodiment and the comparative models all increase with the increase of the number of recommendations N. Regarding the F-scores, except for BPR and SocialMF, which show a decrease when N increases from 20 to 25, the F-scores of the other models all increase with the increase of N. Looking at all the models involved in the comparative experiment, the model applying graph neural networks outperformed BPR and SocialMF, which did not apply graph neural networks, in all three evaluation metrics.

[0126] When the number of recommended content items N is 5, GCHS achieves a precision of 23.5% on the experimental dataset. When N is 30, it achieves a recall of 27.7%, an F-score of 17.9%, and an NDCG score of 29%. Compared to DiffNet, the best-performing model among all comparison models, GCHS improves the F-score by approximately 5.7% and the NDCG score by approximately 1.6%, demonstrating overall superior performance compared to other models.

[0127] (2) Ablation test

[0128] To verify the effectiveness of the inherited heritage features and the attention mechanism used in feature aggregation incorporated into the model—that is, whether they have a positive impact on the model—the GCHS model was compared with a model that did not incorporate inherited heritage features (GCHS-nh) and a model that did not use the attention mechanism (GCHS-natt) in terms of recommendation accuracy and ranking accuracy. The results are shown in […]. Figures 8-11 As shown.

[0129] As shown in the results, the precision of GCHS and the other two models decreases with increasing N, while the recall increases. The F-score and NDCG also generally increase with increasing N. Among the four evaluation metrics, GCHS-nh and GCHS-natt are both lower than GCHS, indicating that removing the heirloom feature or attention mechanism negatively impacts the model's recommendation performance. This demonstrates that these two components have a positive effect on the model's recommendation effectiveness.

[0130] S6.5. Results of Intangible Cultural Heritage Content Recommendation

[0131] Given the good recommendation performance of the complete GCHS recommendation model in the above ablation and comparison experiments, it was applied to the test dataset to recommend intangible cultural heritage content. The user-content similarity results are shown in Table 3. After sorting the similarity results in descending order, the Top-N recommendation list of users is shown in Table 4.

[0132] Table 3 User-Content Similarity Results

[0133]

[0134] Table 4 Recommended List of Intangible Cultural Heritage Content

[0135]

[0136] This embodiment proposes a method for recommending intangible cultural heritage (ICH) content by integrating the social characteristics of ICH inheritors. It extracts inheritor features from the implicit social relationships of their accounts and incorporates these features into user-content similarity calculation to obtain a comprehensive similarity score for final Top-N recommendations. In the feature construction of users and content, a spatial graph convolution method is used, with an attention mechanism employed in the aggregation layer. Neighbor information is used to obtain the features of corresponding objects for similarity calculation. Finally, ablation experiments are conducted on a real dataset using Recall, Precision, F-score, and NDCG as evaluation metrics. Comparison experiments are performed with baseline models such as BPR, GraphRec, and NGCF. The results show that the proposed model exhibits good performance in both recommendation accuracy and ranking accuracy, improving the accuracy of ICH content recommendations to users.

[0137] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for fusing non-heritage social relations of a graph convolution content recommendation, characterized in that, The method comprises: characteristic extraction of users, intangible cultural heritage content and inheritors, obtaining user interest preference characteristics, intangible cultural heritage content characteristics and inheritor characteristics; based on the user interest preference characteristics, intangible cultural heritage content characteristics and inheritor characteristics, using vector inner product method to obtain the similarity value between the user and the intangible cultural heritage content to be recommended; according to the similarity value, recommending intangible cultural heritage content to the user; obtaining the user interest preference characteristics, intangible cultural heritage content characteristics and inheritor characteristics comprises: obtaining the required user initial characteristics and content initial characteristics respectively; based on the obtained required user initial characteristics and content initial characteristics, using a multi-layer spatial domain graph convolutional neural network with attention mechanism as the aggregation method to obtain the user characteristics and content characteristics of each layer; connecting the user characteristics and content characteristics of each layer to obtain the final user interest preference characteristics and intangible cultural heritage content characteristics, and then using the average aggregation method to obtain the inheritor characteristics. 2.The method of claim 1, wherein, obtaining the user initial characteristics comprises: Define the user set as The collection of intangible cultural heritage content is ,in , Represent the total number of users and the total number of content items, respectively; define the user-content interaction matrix as follows: ,use Representation matrix The Middle Line number One element, Indicates user With content There has been interaction. Indicates user With content No interactive behavior occurred; regarding the user-content interaction matrix One-hot encoding is performed, and the initial user characteristics of the corresponding user nodes are obtained by comparing them according to the index order. 3.The method of claim 1, wherein, the user characteristics of each layer are: in, Indicates user In the The feature vector after the nth iteration and All of them are trainable weight vectors. Indicates user The set of first-order neighbors in a social relationship graph. This indicates that the corresponding elements of two vectors are multiplied. Indicates aggregated users When the feature vector is its first-order neighbor user Attention weights Indicates user First-order neighbor users , Indicates user In the The feature vector after the nth iteration. 4.The method of claim 1, wherein, the user interest preference characteristics are: wherein, is a user interest preference feature vector, is the number of iterations. 5.The method of claim 1, wherein, the content initial characteristics are: wherein, is a content initial feature, represents a set of user nodes that interacted with the content initially; the content characteristics of each layer are: in, Content In the The feature vector after the nth iteration Indicates after the first Users after the next iteration eigenvectors, This indicates that it has passed. Content after the next iteration eigenvectors, This indicates that the corresponding elements of two vectors are multiplied. Indicates aggregated content When the feature vector is its first-order neighbor user Attention weights; the final intangible cultural heritage content characteristics are: wherein, is the final non-heritage content feature vector, is the number of iterations. 6.The method of claim 1, wherein, the inheritor characteristics are obtained based on the characteristics of the intangible cultural heritage content published by the inheritor through the average aggregation method; the inheritor characteristics are: wherein, successor feature vector, representing a successor a set of published content, non-heritage content published by the successor.

7. The method of claim 1, wherein the method of fusing non- genetic heir social relationship with graph convolution content recommendation is characterized in that, the similarity value is: in, This indicates content to be recommended. With target users The similarity value, For intangible cultural heritage content to be recommended eigenvectors, For target users eigenvectors, For intangible cultural heritage content to be recommended The feature vector of the publisher.

Images