Old people social network analysis and recommendation system based on graph neural network

Abstract

The invention relates to the technical field of computer technologies, and discloses an old people social network analysis and recommendation system based on a graph neural network, which comprises a data acquisition module, a graph construction module, a graph neural network model module, a social network analysis module, a personalized recommendation module and a user interface module, the data acquisition module is used for acquiring and cleaning social related data of old people; and the graph construction module is used for constructing static and dynamic social network graphs. According to the old people social network analysis and recommendation system based on the graph neural network, a dynamic social network graph is constructed through the graph construction module, sliding time window updating is adopted, dynamic changes of the social relation of old people along with time can be captured, meanwhile, edge weight calculation is combined with the interaction frequency and the interaction depth, and the recommendation efficiency is improved. The interaction depth fuses semantic similarity and emotion scores, the interaction depth is accurately quantified, and a social network analysis module introduces a time decay factor prediction relation.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a social network analysis and recommendation system for the elderly based on graph neural networks. Background Technology

[0002] With the increasing aging of the population, the social needs of the elderly are receiving more and more attention. They not only need to maintain existing social relationships, but also hope to expand new social circles to enrich their later years. Against this backdrop, social networks have become an important platform for the elderly to conduct social interactions. By analyzing the social network structure of the elderly and providing accurate recommendations, their social quality and life satisfaction can be effectively improved. Therefore, related analysis and recommendation systems have gradually become a research hotspot.

[0003] However, existing technologies have significant shortcomings in analyzing and recommending social networks among the elderly: on the one hand, traditional social network analysis methods are mostly based on static graph structures, making it difficult to capture the dynamic changes in social relationships among the elderly over time, and the quantification of the depth of social interaction is not precise enough, leading to discrepancies between the analysis results and the actual social situation; on the other hand, recommendation algorithms lack a dynamic adjustment mechanism for the characteristics of the elderly's social behavior when integrating multiple recommendation strategies, and cannot optimize the recommendation list according to the personalized characteristics of the elderly such as interaction frequency and interest preferences, resulting in insufficient accuracy and applicability of the recommendation results. Therefore, a social network analysis and recommendation system for the elderly based on graph neural networks is proposed. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a graph neural network-based social network analysis and recommendation system for the elderly. This system has advantages such as accurate analysis of social structure and dynamic optimization of recommendation results, and solves the problems of inaccurate quantification of the depth of social interaction among the elderly and insufficient applicability of recommendations in existing technologies when processing social network analysis and recommendations for the elderly.

[0006] (II) Technical Solution

[0007] To achieve the aforementioned objectives of accurately analyzing social structures and dynamically optimizing recommendation results, this invention provides the following technical solution: a graph neural network-based social network analysis and recommendation system for the elderly, comprising a data acquisition module, a graph construction module, a graph neural network model module, a social network analysis module, a personalized recommendation module, and a user interface module;

[0008] The data acquisition module is used to collect and clean social-related data of the elderly;

[0009] The graph construction module is used to construct static and dynamic social network graphs;

[0010] The graph neural network model module is used to learn node features and optimize the model;

[0011] The social network analysis module is used to analyze the structure of social networks and predict social relationships;

[0012] The personalized recommendation module is used to generate personalized social recommendations;

[0013] The user interface module is used to display analysis results and recommended content.

[0014] Preferably, the data collected by the data acquisition module includes basic information of the elderly, social interaction records, and interest tag data; the basic information fields include age, ranging from 60 to 90 years old, gender, and residential address; the social interaction records cover the number of daily interactions, ranging from 1 to 50 times per day, and the duration of each interaction, ranging from 10 to 300 seconds; the number of interest tags is 5 to 15.

[0015] The data acquisition module's data cleaning process includes: removing duplicate data by calculating data fingerprints, correcting erroneous data based on a preset rule base, and supplementing missing values ​​using the mean-filling method, thereby increasing the data integrity rate to 95%-98%.

[0016] The specific steps of the mean-filling method are as follows:

[0017] The first step is to locate the fields in the dataset that have missing values;

[0018] The second step is to filter all non-missing value samples in the field and calculate their arithmetic mean.

[0019] The third step is to replace all missing values ​​in the field with the arithmetic mean. If the number of non-missing value samples in the field is less than 10, the mean of the field for users of the same type will be used to fill the missing values.

[0020] The classification of users into the same category is based on the overlap of user interest tags and the frequency of social interaction. Users with an overlap of interest tags ≥50% and an interaction frequency of ≥3 times per week are judged as users of the same category.

[0021] Preferably, when the graph construction module constructs a static social network graph, the edge weight calculation adopts a weighted summation formula: edge weight = 0.3 × interaction frequency + 0.7 × interaction depth, where the interaction frequency is 1-20 times per week and the semantic similarity of the interaction depth is 0.1-0.9.

[0022] When constructing a dynamic social network graph, the graph construction module adopts a sliding time window update with a sliding time window of 7-30 days. Incremental updates are performed daily from 2-4 AM, and the node attribute update delay is controlled within 10-30 minutes.

[0023] The method for quantifying the interaction depth is as follows: Social interaction text is segmented into words, and the Word2Vec model is used to convert the segmented words into word vectors. The cosine similarity between the word vectors of the interacting parties is calculated. Simultaneously, a sentiment analysis model is used to extract the sentiment tendency of the text. Positive sentiment is assigned a sentiment score of 0.6-0.9, neutral sentiment is assigned a sentiment score of 0.3-0.5, and negative sentiment is assigned a sentiment score of 0.1-0.2. The final interaction depth is calculated as: 0.6 × cosine similarity + 0.4 × sentiment score.

[0024] The cosine similarity is calculated by the ratio of the vector dot product to the product of the magnitudes.

[0025] Preferably, the graph neural network model module adopts a 3-layer graph convolutional network structure, with an input layer dimension of 64-128 dimensions, a hidden layer node count of 128-256, and an output layer dimension of 32-64 dimensions;

[0026] The graph neural network model module includes a self-attention mechanism. The attention coefficient of this mechanism is calculated using the scaled dot product formula, and the number of attention heads is set to 4-8. The batch size of the model training is 32-128, the number of iterations is 500-1000, the initial learning rate is 0.0005-0.002, and an early stopping strategy is adopted. The patience value of the early stopping strategy is 10-20 to prevent overfitting.

[0027] The self-attention mechanism is used to dynamically weight the feature contributions of different neighboring nodes.

[0028] Preferably, when the social network analysis module performs centrality analysis, the degree centrality calculation range is 10-50 neighbor nodes, the betweenness centrality path calculation adopts Dijkstra's algorithm, and the average path length for close centrality is controlled within 3-8 steps.

[0029] When performing community detection, the social network analysis module uses the Louvain algorithm, with the number of communities ranging from 3 to 8, and the module degree value optimized to 0.3-0.7.

[0030] When the social network analysis module predicts social relationships, it incorporates a time decay factor, which is 0.1-0.5. The prediction error of the probability of relationship establishment in the next 30-90 days is ≤15%.

[0031] The modularity is calculated by the difference between the number of edges within the community and the expected number of edges in the random network.

[0032] Preferably, the personalized recommendation module employs a hybrid recommendation strategy, specifically including:

[0033] Collaborative filtering: User similarity is calculated using the Pearson correlation coefficient, which ranges from -0.8 to 0.8. The top 10-30 users with the highest similarity are selected as similar users.

[0034] Graph Neural Network Part: Set the embedding vector similarity threshold to 0.6-0.8, and select items that meet this threshold to be included in the recommendation candidate set;

[0035] Recommendation list composition: The list length is 5-20 items, of which social activity recommendations account for 40%-60% and new social contacts recommendations account for 40%-60%;

[0036] List length adjustment rules: Dynamically adjusted based on the user's average daily interaction count over the past 30 days. If the average daily interaction count is ≥20 times, the recommended list length is 15-20 items; if the average daily interaction count is 5-19 times, the recommended list length is 10-14 items; if the average daily interaction count is ≤4 times, the recommended list length is 5-9 items.

[0037] Collaborative filtering implementation: Recommendation results are generated by weighted sum of the preferences of the target user and similar users, where the weighted sum is calculated based on the sum of the products of the preference scores of similar users and their corresponding similarity scores.

[0038] Preferably, the font size of the user interface module is set to 16-24pt, and the color contrast is ≥4.5:1;

[0039] When the user interface module visualizes the social network, it adopts a force-oriented layout, where the node size is positively correlated with the degree centrality, the node diameter is 10-30 pixels, the edge thickness is positively correlated with the weight, and the edge width is 1-5 pixels.

[0040] The recommended content display area of ​​the user interface module has an image resolution of 300-600dpi and a text description length of 50-200 characters.

[0041] The force-oriented layout visualizes the network structure by simulating the balance of attraction and repulsion between nodes.

[0042] Preferably, it also includes a data storage module, which uses MySQL to store structured data, with the number of records in a single table controlled between 100,000 and 1 million; it uses Neo4j to store graph data, with the maximum number of nodes between 100,000 and 500,000 and the maximum number of edges between 1 million and 5 million; the data backup cycle is 1-7 days, and the backup file is kept for 30-90 days;

[0043] The structured data uses the user ID as the primary key to associate and store basic information and interaction records.

[0044] Preferably, the training data of the graph neural network model module adopts a stratified sampling method, with the training set accounting for 70%-80%, the validation set accounting for 10%-15%, and the test set accounting for 10%-15%.

[0045] The performance evaluation metrics for the graph neural network model include accuracy, recall, and F1 score, with accuracy ranging from 80% to 95%, recall from 75% to 90%, and F1 score from 78% to 92%. The evaluation datasets include a rating dataset containing 23,367 users and 4,580 items, a review dataset containing 401,230 users and 125,160 items, and a comment dataset containing 1,508 users and 2,071 items.

[0046] Preferably, the personalized recommendation module includes a group recommendation function. When the group size is 3-11 people, a weighted average method is used to aggregate individual preferences, with the weight allocation ranging from 0.05 to 0.3, and the group satisfaction rate of the recommendation result is ≥80%.

[0047] The association rule between the weighting and the social preferences of the elderly is as follows: if a user's interaction frequency with other members in the group is ≥10 times / week in the past 90 days, the preference weight is set to 0.2-0.3; if the interaction frequency is 3-9 times / week, the weight is set to 0.1-0.19; if the interaction frequency is ≤2 times / week, the weight is set to 0.05-0.09.

[0048] The weighted average method calculates group preferences by summing the products of individual preference scores and their corresponding weights.

[0049] (III) Beneficial Effects

[0050] Compared with existing technologies, this invention provides a social network analysis and recommendation system for the elderly based on graph neural networks, which has the following beneficial effects:

[0051] 1. This graph neural network-based social network analysis and recommendation system for the elderly constructs a dynamic social network graph through a graph construction module and uses a sliding time window for updating. This allows the system to capture the dynamic changes in the social relationships of the elderly over time. At the same time, the edge weight calculation combines interaction frequency and interaction depth. The interaction depth integrates semantic similarity and sentiment score to accurately quantify the interaction depth. The social network analysis module introduces a time decay factor to predict relationships, which solves the problems of traditional static graph structures and inaccurate quantification, making the analysis results more realistic.

[0052] 2. This graph neural network-based social network analysis and recommendation system for the elderly employs a hybrid strategy through a personalized recommendation module. It uses collaborative filtering to select similar users, while the graph neural network part filters items that meet thresholds. The length of the recommendation list is dynamically adjusted according to the average number of daily interactions, and group recommendations are weighted according to interaction frequency. This solves the problem of the lack of dynamic adjustment in recommendations and improves the accuracy and applicability of recommendations. Attached Figure Description

[0053] Figure 1 This is a diagram of the social network analysis and recommendation system for the elderly according to the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 A social network analysis and recommendation system for the elderly based on graph neural networks includes a data acquisition module, a graph construction module, a graph neural network model module, a social network analysis module, a personalized recommendation module, and a user interface module.

[0056] The data acquisition module is used to collect and clean social-related data of the elderly;

[0057] The graph construction module is used to construct static and dynamic social network graphs;

[0058] The graph neural network model module is used to learn node features and optimize the model;

[0059] The social network analysis module is used to analyze the structure of social networks and predict social relationships;

[0060] The personalized recommendation module is used to generate personalized social recommendations;

[0061] The user interface module is used to display analysis results and recommended content.

[0062] Example 1:

[0063] This embodiment details the specific operation process of the data acquisition module, including data sources, acquisition content, and cleaning process.

[0064] Data collection sources include registration information, interaction logs, and interest tag libraries on social media platforms for seniors. Specific data collected is as follows:

[0065] Basic information is obtained through a user registration form, including age (60-90 years old, accurate to the year), gender (male / female), and residential address (accurate to the street level).

[0066] Social interaction records are collected through the platform's log system, covering the number of daily interactions (1-50 times, recording the initiation time and target of each interaction) and the duration of a single interaction (10-300 seconds, obtained through the system's timing function);

[0067] Interest tags (5-15) are generated through user-initiated annotation and behavior analysis. For example, if a user frequently participates in calligraphy exchanges or square dancing activities, the corresponding tags will be automatically added.

[0068] The data cleaning process is performed according to the following steps:

[0069] The first step is to calculate the data fingerprint to remove duplicate data, generate a unique hash value for each data, compare the data with the same hash value and keep the latest one;

[0070] The second step is to correct erroneous data based on a preset rule base. For example, if the age field shows 50 years old which does not fall within the 60-90 age range, it will be corrected to the actual age based on the ID card information filled in during registration.

[0071] The third step is to use the mean imputation method to fill in missing values. After locating the missing field, the arithmetic mean of the non-missing samples in that field is calculated. If the sample size is less than 10, users are divided into the same category. The criteria for this are that the overlap of interest tags is not less than 50% and the number of interactions per week is not less than 3. The mean of that field of users in the same category is used to fill in the missing values. After cleaning, the data integrity rate is stable at 95%-98%.

[0072] The data collection module addresses the issue of incomplete social data collection among the elderly by accurately segmenting users of similar types and supplementing missing data, thereby improving data usability.

[0073] Example 2:

[0074] This embodiment illustrates the construction methods of static and dynamic social network graphs, focusing on the edge weight calculation and dynamic update mechanism.

[0075] When constructing a static social network graph, edge weight calculation incorporates interaction frequency and interaction depth. Interaction frequency is the number of interactions per week (1-20 times, based on statistics from the past 30 days of interaction records); interaction depth calculation requires two steps:

[0076] First, the social interaction texts, including chat logs and comments, are segmented into words. The Word2Vec model is used to convert the segmented words into word vectors, and the training corpus covers social vocabulary commonly used by the elderly. The cosine similarity between the word vectors of the two parties in the interaction is calculated, ranging from 0.1 to 0.9.

[0077] Secondly, the sentiment tendency of the text is extracted through a sentiment analysis model. Positive sentiment is assigned a sentiment score of 0.6-0.9, neutral sentiment is assigned 0.3-0.5, and negative sentiment is assigned 0.1-0.2. Finally, the interaction depth is obtained by weighting the cosine similarity and the sentiment score.

[0078] The edge weights are obtained by weighted summation of interaction frequency and interaction depth, thereby quantifying the closeness of social relationships among the elderly.

[0079] The dynamic social network graph uses a sliding time window for updates, with the window size set to 14 days, and performs incremental updates daily at 3 AM.

[0080] The edge weights of new interaction records in the past 14 days, including new friends and interaction content, are calculated and added to the graph. Historical data that exceeds the window and is older than 14 days is archived.

[0081] When node attributes, including interest tags and interaction frequency, are updated, an asynchronous processing mechanism is used to control the delay to within 20 minutes to ensure data timeliness.

[0082] The graph construction module captures changes in social relationships through dynamic windows and optimizes weight calculations by combining sentiment analysis of interactive text, making the social network graph more closely reflect the actual social status of the elderly.

[0083] Example 3:

[0084] This embodiment illustrates the process of constructing, training, and optimizing a graph neural network model.

[0085] The graph neural network model adopts a 3-layer graph convolutional network structure. The input layer has a dimension of 128, which includes encoding user basic information, embedding interest tags, and interaction feature vectors. The hidden layer has 256 nodes, and the output layer has a dimension of 64, which is used for node feature representation.

[0086] The model incorporates a self-attention mechanism with 6 attention heads. This mechanism dynamically weights the feature contributions of different neighboring nodes, focusing on nodes with high trust levels in the social interactions of the elderly, such as social objects with frequent interactions and positive emotions.

[0087] The training process for a graph neural network model is as follows:

[0088] The dataset was divided using stratified sampling, with the training set accounting for 80%, the validation set for 10%, and the test set for 10%. The training data included ratings from 23,367 users on 4,580 items and comments from 1,508 users on 2,071 items. This data covered user social interactions and preference information, making it suitable for social analysis scenarios involving the elderly.

[0089] The batch size was set to 64, the number of iterations to 1000, the initial learning rate to 0.001, and the Adam optimizer was used. The early stopping strategy was enabled, and the patience value was set to 15, which means that training was terminated when the validation set loss did not decrease for 15 consecutive iterations to prevent overfitting.

[0090] During training, the model accuracy and recall are output every 100 iterations, and the final F1 score of the model stabilizes at around 85%.

[0091] The graph neural network model module focuses on core social relationships through a self-attention mechanism and trains the model by combining multi-dimensional user features, thereby improving the learning accuracy of social node features of the elderly.

[0092] Example 4:

[0093] This example illustrates the centrality analysis, community discovery, and social relationship prediction operations of the social network analysis module.

[0094] In the centrality analysis, degree centrality is calculated by selecting the top 30 neighboring nodes of the target user, sorting them by interaction frequency, and counting and normalizing the number of connections between nodes; betweenness centrality uses Dijkstra's algorithm to calculate the proportion of all shortest paths that pass through the target node, and the path search range is limited to within 500 nodes to improve efficiency; proximity centrality calculates the average path length from the target node to other nodes, and the reasonableness of the results is ensured by limiting the number of path steps to 3-8.

[0095] Community detection uses the Louvain algorithm, with the following steps:

[0096] Each node is initialized as an independent community, and the current modularity is calculated by the difference between the number of edges within the community and the expected number of edges in the random network.

[0097] Iteratively merge communities, merging the two communities that result in the largest increase in modularity each time, until the modularity no longer increases and is optimized to around 0.5. Finally, the network is divided into 5 communities, each containing 200-500 nodes.

[0098] When predicting social relationships, the fusion time decay factor is set to 0.3, and the weights of historical interaction data are adjusted according to time decay. A prediction model is built using user similarity and trust value to predict the probability of establishing a relationship in the next 60 days. Through testing with 1,000 sets of samples, the prediction error is controlled within 12%.

[0099] The social network analysis module optimizes the modularity calculation of community segmentation and introduces a time decay factor, making the social network analysis results more consistent with the dynamic evolution of social relationships among the elderly.

[0100] Example 5:

[0101] This embodiment illustrates the specific implementation of the hybrid recommendation strategy and group recommendation function of the personalized recommendation module.

[0102] In a hybrid recommendation strategy, the collaborative filtering component is executed as follows:

[0103] Similarity is calculated by users rating their preferences for the project, and the top 20 users with the highest similarity are selected as similar users.

[0104] The graph neural network part calculates similarity using the 64-dimensional embedding vectors output by the model, and selects items with a similarity of 0.7 or higher as the candidate set.

[0105] The recommended list is composed of:

[0106] The length is dynamically adjusted based on the user's average daily interaction frequency over the past 30 days. There are 18 items when there are no less than 20 interactions, 12 items when there are 5-19 interactions, and 8 items when there are no more than 4 interactions. Among them, social activity recommendations and new social object recommendations each account for 50%.

[0107] The group recommendation feature targets groups of 3-11 people and uses a mean-based strategy combined with the social characteristics of older adults to adjust the weights.

[0108] If a user interacts with group members at least 10 times per week in the past 90 days, the weight is set to 0.25; 3-9 times is set to 0.15; and no more than 2 times is set to 0.08. The group preference is calculated by summing the products of individual preferences and corresponding weights, so that the group satisfaction reaches 82%.

[0109] The personalized recommendation module improves the matching degree to the personalized and group social needs of the elderly by dynamically adjusting the length of the recommendation list and the weight of group preferences, combined with a hybrid recommendation strategy.

[0110] Example 6:

[0111] This embodiment illustrates the design details of the user interface module and the architecture of the data storage module.

[0112] In the user interface module, the font size is set to 20pt, the color contrast is 5:1, and a dark blue background with white text is used to suit the visual characteristics of the elderly. The social network visualization adopts a force-oriented layout, achieving balance by simulating the positive correlation between attraction and edge weight and the negative correlation between repulsion and node distance. The node diameter is set according to degree centrality, with a maximum of 30 pixels and a minimum of 10 pixels. The edge width is set according to weight, with 1-5 pixels (5 pixels for weights above 0.8 and 1 pixel for weights below 0.3). In the recommended content display area, the image resolution is set to 450dpi, and the text description is controlled to around 100 words, including the event time, location, and participant tags to highlight group preference information.

[0113] The data storage module adopts a hybrid architecture:

[0114] MySQL stores structured data, with user ID as the primary key, and basic information stored in related tables, with 500,000 records per table. Interaction records are also stored in time-partitioned data.

[0115] Neo4j stores graph data. Node attributes include user ID and interest tags, while edge attributes include weight and interaction time. The maximum number of nodes is 300,000, and the maximum number of edges is 3 million.

[0116] Data backup is performed daily at 3:00 AM, using a combination of incremental and full backups. Backup files are stored for 60 days, and off-site storage ensures data security.

[0117] The user interface module and data storage module are optimized to suit the physiological characteristics of the elderly, and adopt a hybrid storage architecture to ensure data integrity and access efficiency, thus providing support for the stable operation of the system.

[0118] In summary, this graph neural network-based social network analysis and recommendation system for the elderly constructs a dynamic social network graph through a graph construction module and uses a sliding time window for updates, which can capture the dynamic changes in the social relationships of the elderly over time. At the same time, the edge weight calculation combines interaction frequency and interaction depth, and the interaction depth integrates semantic similarity and sentiment score to accurately quantify the interaction depth. The social network analysis module introduces a time decay factor to predict relationships, which solves the problems of inaccurate quantification in traditional static graph structures and makes the analysis results more realistic.

[0119] Furthermore, this graph neural network-based social network analysis and recommendation system for the elderly employs a hybrid strategy through a personalized recommendation module. It uses collaborative filtering to select similar users, while the graph neural network part filters items that meet thresholds. The length of the recommendation list is dynamically adjusted according to the average number of daily interactions, and group recommendations are weighted according to interaction frequency. This solves the problem of the lack of dynamic adjustment in recommendations, improves the accuracy and applicability of recommendations, and addresses the issues of inaccurate quantification of the depth of social interactions among the elderly and insufficient applicability of recommendations in existing technologies when processing social network analysis and recommendations for the elderly.

[0120] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A social network analysis and recommendation system for the elderly based on graph neural networks, characterized in that, It includes a data acquisition module, a graph construction module, a graph neural network model module, a social network analysis module, a personalized recommendation module, and a user interface module; The data acquisition module is used to collect and clean social-related data of the elderly; The graph construction module is used to construct static and dynamic social network graphs; The graph neural network model module is used to learn node features and optimize the model; The social network analysis module is used to analyze the structure of social networks and predict social relationships; The personalized recommendation module is used to generate personalized social recommendations; The user interface module is used to display analysis results and recommended content.

2. The graph neural network-based social network analysis and recommendation system for the elderly according to claim 1, characterized in that, The data collection module collects data including basic information of the elderly, social interaction records, and interest tag data; the basic information fields include age, ranging from 60 to 90 years old, gender, and residential address; the social interaction records cover the number of daily interactions, ranging from 1 to 50 times per day, and the duration of each interaction, ranging from 10 to 300 seconds. The number of interest tags should be 5-15; The data acquisition module's data cleaning process includes: removing duplicate data by calculating data fingerprints, correcting erroneous data based on a preset rule base, and supplementing missing values ​​using the mean-filling method, thereby increasing the data integrity rate to 95%-98%. The specific steps of the mean-filling method are as follows: The first step is to locate the fields in the dataset that have missing values; The second step is to filter all non-missing value samples in the field and calculate their arithmetic mean. The third step is to replace all missing values ​​in the field with the arithmetic mean. If the number of non-missing value samples in the field is less than 10, the mean of the field for users of the same type will be used to fill the missing values. The classification of users into the same category is based on the overlap of user interest tags and the frequency of social interaction. Users with an overlap of interest tags ≥50% and an interaction frequency of ≥3 times per week are judged as users of the same category.

3. The graph neural network-based social network analysis and recommendation system for the elderly according to claim 1, characterized in that, When the graph construction module constructs a static social network graph, the edge weight is calculated using a weighted summation formula: edge weight = 0.3 × interaction frequency + 0.7 × interaction depth, where the interaction frequency is 1-20 times per week and the semantic similarity of the interaction depth is 0.1-0.

9. When constructing a dynamic social network graph, the graph construction module adopts a sliding time window update with a sliding time window of 7-30 days. Incremental updates are performed daily from 2-4 AM, and the node attribute update delay is controlled within 10-30 minutes. The method for quantifying the interaction depth is as follows: Social interaction text is segmented into words, and the Word2Vec model is used to convert the segmented words into word vectors. The cosine similarity between the word vectors of the interacting parties is calculated. Simultaneously, a sentiment analysis model is used to extract the sentiment tendency of the text. Positive sentiment is assigned a sentiment score of 0.6-0.9, neutral sentiment is assigned a sentiment score of 0.3-0.5, and negative sentiment is assigned a sentiment score of 0.1-0.

2. The final interaction depth is calculated as: 0.6 × cosine similarity + 0.4 × sentiment score. The cosine similarity is calculated by the ratio of the vector dot product to the product of the magnitudes.

4. The social network analysis and recommendation system for the elderly based on graph neural networks according to claim 1, characterized in that, The graph neural network model module adopts a 3-layer graph convolutional network structure, with an input layer dimension of 64-128 dimensions, a hidden layer node count of 128-256, and an output layer dimension of 32-64 dimensions; The graph neural network model module includes a self-attention mechanism. The attention coefficient of this mechanism is calculated using the scaled dot product formula, and the number of attention heads is set to 4-8. The batch size of the model training is 32-128, the number of iterations is 500-1000, the initial learning rate is 0.0005-0.002, and an early stopping strategy is adopted. The patience value of the early stopping strategy is 10-20 to prevent overfitting. The self-attention mechanism is used to dynamically weight the feature contributions of different neighboring nodes.

5. The social network analysis and recommendation system for the elderly based on graph neural networks according to claim 1, characterized in that, When performing centrality analysis, the degree centrality calculation range is 10-50 neighboring nodes, the betweenness centrality path calculation adopts Dijkstra's algorithm, and the average path length for close centrality is controlled within 3-8 steps. When performing community detection, the social network analysis module uses the Louvain algorithm, with the number of communities ranging from 3 to 8, and the module degree value optimized to 0.3-0.

7. When the social network analysis module predicts social relationships, it incorporates a time decay factor, which is 0.1-0.

5. The prediction error of the probability of relationship establishment in the next 30-90 days is ≤15%. The modularity is calculated by the difference between the number of edges within the community and the expected number of edges in the random network.

6. The social network analysis and recommendation system for the elderly based on graph neural networks according to claim 1, characterized in that, The personalized recommendation module employs a hybrid recommendation strategy, specifically including: Collaborative filtering: User similarity is calculated using the Pearson correlation coefficient, which ranges from -0.8 to 0.

8. The top 10-30 users with the highest similarity are selected as similar users. Graph Neural Network Part: Set the embedding vector similarity threshold to 0.6-0.8, and select items that meet this threshold to be included in the recommendation candidate set; Recommendation list composition: The list length is 5-20 items, of which social activity recommendations account for 40%-60% and new social contacts recommendations account for 40%-60%; List length adjustment rules: Dynamically adjusted based on the user's average daily interaction count over the past 30 days. If the average daily interaction count is ≥20 times, the recommended list length is 15-20 items; if the average daily interaction count is 5-19 times, the recommended list length is 10-14 items; if the average daily interaction count is ≤4 times, the recommended list length is 5-9 items. Collaborative filtering implementation: Recommendation results are generated by weighted sum of the preferences of the target user and similar users, where the weighted sum is calculated based on the sum of the products of the preference scores of similar users and their corresponding similarity scores.

7. The social network analysis and recommendation system for the elderly based on graph neural networks according to claim 1, characterized in that, The font size of the user interface module is set to 16-24pt, and the color contrast is ≥4.5:1; When the user interface module visualizes the social network, it adopts a force-oriented layout, where the node size is positively correlated with the degree centrality, the node diameter is 10-30 pixels, the edge thickness is positively correlated with the weight, and the edge width is 1-5 pixels. The recommended content display area of ​​the user interface module has an image resolution of 300-600dpi and a text description length of 50-200 characters. The force-oriented layout visualizes the network structure by simulating the balance of attraction and repulsion between nodes.

8. The social network analysis and recommendation system for the elderly based on graph neural networks according to claim 1, characterized in that, It also includes a data storage module, which uses MySQL to store structured data, with the number of records in a single table controlled between 100,000 and 1 million; it uses Neo4j to store graph data, with the maximum number of nodes between 100,000 and 500,000 and the maximum number of edges between 1 million and 5 million; the data backup cycle is 1-7 days, and the backup file is kept for 30-90 days. The structured data uses the user ID as the primary key to associate and store basic information and interaction records.

9. The social network analysis and recommendation system for the elderly based on graph neural networks according to claim 1, characterized in that, The training data for the graph neural network model module is divided into hierarchical sampling, with the training set accounting for 70%-80%, the validation set accounting for 10%-15%, and the test set accounting for 10%-15%. The performance evaluation metrics for the graph neural network model include accuracy, recall, and F1 score, with accuracy ranging from 80% to 95%, recall from 75% to 90%, and F1 score from 78% to 92%. The evaluation datasets include a rating dataset containing 23,367 users and 4,580 items, a review dataset containing 401,230 users and 125,160 items, and a comment dataset containing 1,508 users and 2,071 items.

10. A graph neural network-based social network analysis and recommendation system for the elderly according to claim 1, characterized in that, The personalized recommendation module includes a group recommendation function. When the group size is 3-11 people, a weighted average method is used to aggregate individual preferences, with the weight allocation ranging from 0.05 to 0.

3. The group satisfaction rate of the recommendation results is ≥80%. The association rule between the weighting and the social preferences of the elderly is as follows: if a user's interaction frequency with other members in the group is ≥10 times / week in the past 90 days, their preference weight is set to 0.2-0.

3. If the interaction frequency is 3-9 times / week, the weight is set to 0.1-0.19; if the interaction frequency is ≤2 times / week, the weight is set to 0.05-0.

09. The weighted average method calculates group preferences by summing the products of individual preference scores and their corresponding weights.

Images