Interactive recommendation of social network communication between cities based on GNN and user preferences

2.1 Cross-city POI migration modeling considering user preferences

POI refers to places or locations that users may be interested in within a specific geographic location, such as restaurants, attractions, and shopping centers. In practical urban applications, users usually have stable POI preferences, but when they move between different cities, these POIs may undergo significant changes, known as the POI migration phenomenon. This dynamic change makes it difficult for traditional recommendation models to maintain the accuracy and relevance of recommendations [15,16]. To obtain the importance of each POI, the SA mechanism is introduced for weight allocation. Compared to other methods, the SA mechanism can reduce reliance on external information, thereby enhancing the capture of internal correlations in data [17]. The SA mechanism calculation of POIs within the user’s city is shown in Figure 1.

Figure 1

SA mechanism calculation for POI within a user’s city.

In Figure 1, first, the user’s historical behavior and geographic location data are input into the input layer. Second, in the SA mechanism layer, the model processes the input data through the SA mechanism, identifies users’ potential preferences for different POIs, and calculates relevant weights. Then, this information is transformed into feature vectors in the feature vector representation layer, representing the user’s level of interest in different POIs within the current city and, finally, the output. The feature vector representation of the new POI, the weight coefficients of the SA mechanism, and the fused feature representation of multiple POIs are shown in the following equation:

where Q , K , and V are the query vector, key vector, and value vector, respectively, W v , W k , and W ^q represent weighted matrices, R s is the input feature vector, soft max is the function normalization, and s o f t max and V are attention weights and value vectors, respectively. The feature vector of the new POI is linearly reduced to obtain the true user features, and the calculation formula is shown in the following equation:

where u s ′ is the user’s new POI feature after dimensionality reduction, b 1 is the dimensionality reduction bias vector, and σ is the activation function. Similarly, this method can also calculate the POI preference feature weights of users in other cities. However, how to ultimately make the two converge remains a thorny issue. Therefore, this study introduces the CA mechanism to conduct global feature interaction on the dissemination of interest preferences among users on social networks between cities, aiming to achieve the goal of transferring POI preferences among end-users. Compared to other methods, the CA mechanism can quickly learn the representation of problems and answers and then adjust the weights appropriately, thereby improving the correlation and interactivity between problems and answers within and between cities [18,19]. A schematic diagram of the calculation of the CA mechanism for POI between user cities is shown in Figure 2.

Figure 2

Calculation of the CA mechanism for the user’s intercity POI.

In Figure 2, first, the POI data feature vector of the user in the source city is normalized together with the similar POI data feature vector of the user in the target city as input. The interest preference matching degree is generated by calculating the CA mechanism, and the POI migration trend of users in the target city is captured by adjusting the weighting coefficients. Finally, by summing and pooling multiple POI information, a more comprehensive recommendation feature vector is generated. The POI similarity normalization between the source city and the target city for users is shown in the following equation:

where R r T is the POI similarity calculation value of the user in the target city, W s is the dimension change matrix after CA mechanism feature vector learning, and β r is the normalized user POI similarity calculation coefficient. After weighting, the true POI migration feature vector of the user’s target city is obtained, as shown in the following equation:

where u T ′ is the weighted POI migration feature vector of the user’s target city, M is the M -th similarity coefficient, and r is the r -th POI. Based on the capture of POI feature vectors for users inside and outside the city, this study combines the two to obtain cross-city POI migration results. The process is shown in Figure 3.

Figure 3

The migration process of user preference POI.

In Figure 3, the entire migration fusion process is divided into two channels. First, users input POI information data in both the source and target cities. Second, the SA mechanism is used to perform dimensionality reduction and embedding representation on the POI data of users in the source and target cities, and the feature vectors of POIs are obtained separately. Afterward, the CA mechanism fuses the two types of feature vectors to obtain the final cross-city POI vector, as shown in the following equation:

where W 2 and b 2 are the weight matrix and bias vector of the hidden layer, respectively. The transfer loss function, Transfer Loss, is used to constrain the similarity calculation of feature vectors, as shown in the following equation:

where τ T is the cosine similarity.

2.2 Construction of an interactive POI recommendation model between cities based on GNN

After constructing the inter-city user POI preference migration model, it was found that if POI, such as tourism, medical treatment, employment, and education, are used as the main targets, these users often tend to prefer locations with higher popularity [20]. Traditional methods rely solely on the strength of the POI correlation between two locations for recommendation, which cannot meet user needs with high standards. Therefore, this study uses popularity as a POI selection label and introduces GNN to model the network topology of POI popularity features and geographic locations in the target city. GAT, as a specific model in GNN, introduces an attention mechanism to optimize spatially based GNN [21,22]. GAT overcomes some limitations of traditional GNN, enabling the model to better understand and process node relationships in graph data. The GAT structure diagram is shown in Figure 4 [23].

Figure 4

Schematic diagram of the GAT structure.

Figure 4(a) shows the calculation process of the attention coefficient, and Figure 4(b) shows the node state update process. By weighting the feature vector h j of neighboring nodes and applying an activation function, the attention weight α i j can be obtained and then aggregated to generate the target node h i . In Figure 4(b), the state of the target node h i is updated by combining the features of all neighboring nodes and their corresponding attention coefficients. The formula for the attention coefficient is

where j is the attention score between node i and node e i j , α is a learnable attention weight vector, W is a learnable weighted matrix, and LeakyReLU is the activation function. The calculation of attention weights is shown in the following equation:

where e i k is the attention score between node i and node k , and N i is the set of neighboring nodes of node i . The update of node features is shown in the following equation:

where h i ′ is the updated feature representation of the node i . This study uses GAT to extract the most popular user POIs in the target city and constructs a POI popularity topology framework based on this, as shown in Figure 5.

Figure 5

The topological framework of the heat level of the POI in the target city.

In Figure 5, first, in the embedding layer, the attribute data of POI are embedded as feature vectors. Next, these feature vectors represent the relationship structure between POIs by constructing a POI map. The graph attention layer uses GAT to process nodes in the POI graph and obtain aggregated information about neighboring nodes. Finally, in the POI graph aggregation layer, the processed POI graph is aggregated to form the final POI heat topology structure. The heat aggregation feature representation of POI by GAT is shown in the following equation:

where h i ( 0 ) is the initial feature vector for embedding the target POI, N ( V j ) is the set of domain nodes for the target POI V j , and Agg is an aggregation function. This study combines the above calculation methods and incorporates the attention mechanism into the aggregation function so that the feature vector of a single POI V j can be jointly determined by its connected edge weights, thereby avoiding the problem of incomplete consideration of global information. At this point, the attention coefficient determined by the weights between neighboring points is calculated as shown in the following equation:

where α i j ⁎ is the attention coefficient value determined based on the weights of neighboring points. The POI similarity between two cities is transformed into a weight matrix, and weight fusion is performed through concatenation and vector multiplication. Then, high-dimensional feature mapping is applied to output the optimal attention coefficient value. Finally, the influence coefficient between two POIs can be obtained by normalizing with the softmax function, as shown in the following equation:

where att i j is the coefficient of influence. By substituting the influence coefficient into Eq. (9), the information aggregation update of adjacent solution nodes, namely POIs, is obtained, as shown in the following equation:

where W t is the POI shared transformation matrix. After multiple aggregation updates, a single POI feature matrix representation of the user in the target city can be obtained, as shown in the following equation:

where V T is a single POI feature matrix. This study uses matrix decomposition to obtain the probability of users accessing POI V j in the target city, as shown in the following equation:

In summary, this study proposes a novel intercity social network communication interactive POI recommendation model by combining user preference POI migration and user target city POI popularity aggregation. The structure of the model is shown in Figure 6.

Figure 6

Novel intercity social network communication interaction POI recommendation model structure.

In Figure 6, the model can be divided into three main modules: user preference transfer module, recommendation prediction module, and data filtering module. First, user characteristics are input and processed by the SA mechanism to generate a spatiotemporal feature representation of the user in the source city. Then, the CA mechanism is used to capture the cross-city migration of user characteristics by combining their activity data in different cities. Second, these data are fused and concatenated, and the dimensionality is reduced to construct a POI matrix for the user’s target city. Then, the target POI matrix with the greatest influence is selected through the calculation of POI popularity and feature aggregation in the target city. Finally, the best POI recommendation is obtained by matching based on the similarity of POIs with the source city.

3.1 Performance testing of urban POI interactive recommendation model

The experimental setup has an Intel Core 2.5 Hz dual-core CPU and 16 GB of memory. The GPU is NVIDIA GeForce RTX 1660, the programming language is Python 3.7.15, and the deep learning framework is Python 1.3.2. The size of the model embedding layer is 128. The optimizer uses the classic Adam algorithm with a learning rate set to 0.001. The Yelp dataset contains information on user reviews and ratings of local merchants (e.g., restaurants and stores). This dataset is suitable for testing POI recommender systems based on user feedback. The dataset covers a wide range of types of merchants and their user ratings. The Brightkite dataset is user check-in data from the Brightkite social network. It records the time, geographic coordinates, and location names of users checking in at different geographical locations. This dataset is mainly used for evaluating cross-city POI recommendations and is particularly suitable for analyzing users’ behavioral patterns across different cities. This study first attempts to determine the transfer loss function in POI preference transfer and the values of attention weight coefficients in the recommendation model to maintain the efficient state of the subsequent model for testing. The result is shown in Figure 7.

Figure 7

Transfer Loss function and attention weighting coefficient test results. (a) Transfer loss function and (b) attention weighting coefficient.

Figure 7(a) and (b) shows the test results of the transfer loss function values and attention weight coefficients. In Figure 7(a), although the recommendation accuracy is the highest when transfer loss is 1.0, there is a slight decrease in model performance in the later stage. When the transfer loss value is 0.6, the model’s recommendation accuracy data are more stable, with a maximum recommendation accuracy of 89.91% and 170 iterations. In Figure 7(b), when the attention weight coefficient is set to 1.0, the highest recommendation accuracy of the model is 88.24% after 250 iterations. Therefore, the transfer loss parameter is found to take the value of 0.6 after many experiments to achieve the best balance between recommendation accuracy and stability. This value can achieve a better balance between keeping the model’s high accuracy and stability. Meanwhile, the attention weight coefficient is set to 1.0 to ensure the optimal performance of the model in capturing user interest points and feature fusion. These parameters are chosen not only to improve the recommendation accuracy of the model but also to ensure the model’s adaptability in different datasets and environments. Deviation from these optimal values can lead to a significant decrease in recommendation accuracy, especially in terms of recommendation coverage and convergence speed. Therefore, the impact of these parameters on the algorithm performance has important sensitivities, and reasonable parameter settings are the key to improving model performance. This study compares advanced recommendation models of the same type as GNN, such as heterogeneous GNN (HetGNN), relational graph convolutional network (RGCN), and attention-based GNN (AGNN) that integrate attention mechanisms. The test results are shown in Figure 8.

Figure 8

Preference migration rate test results for different models. (a) Yelp and (b) Birghtkite.

Figure 8(a) and (b) shows preference transfer test data for four models on the Yelp and Brightkite datasets. In the Yelp dataset of Figure 8(a), the preference migration rate of the proposed model shows the highest stability and accuracy in both datasets. In the Brightkite dataset, the preference migration rate reaches 93.7% with an error of only 2.1%. This indicates that the proposed model is able to effectively capture users’ POI preference migration across cities, especially in the recommendation task with different geographic locations. Compared to the other models, the HetGNN and RGCN models have lower preference migration rates because they do not possess sufficiently strong cross-city preference learning capabilities. AGNN, despite being able to capture a certain amount of user preferences, has a high model complexity and is prone to unstable convergence. These results show that the proposed model has a significant advantage in intercity interest point migration, especially in coping with diverse user needs with high adaptability. Tests are conducted using precision, recall, F1 value, and average recommendation time as indicators, as shown in Table 2.

In Table 2, in Yelp, the highest P, R, and F1 values for the research model are 93.52, 91.24, and 92.38%, respectively, with an average recommendation time of 2.19 s. Compared to HetGNN, RGCN, and AGCN, the research model has significant advantages in recommendation accuracy, stability, and recommendation rate. This reflects the high accuracy and consistency of the proposed model in capturing user POI preferences and generating recommendation results. In Brightkite, the performance of each indicator of the research model is the best, with the highest P, R, and F1 values of 93.36, 91.89, and 92.63%, respectively, and an average recommendation time of 2.23 s. This result further shows that the proposed model not only has advantages in recommendation accuracy but also can effectively improve recommendation efficiency, which is suitable for large-scale real-time application scenarios.

3.2 Simulation testing of the POI interactive recommendation model between cities

This study randomly selects two groups of cities with the top four user POI click-through rates in Baidu’s public database and Gaode’s public database from 2012 to 2022 for visual comparison, namely Chengdu and Chongqing, and Shanghai and Beijing. The first step of the simulation process is to preprocess the two datasets to extract the user’s POI features and generate feature vectors. Next, a model based on the PyTorch framework is trained with the parameters set to an embedding layer size of 128, a learning rate of 0.001, and an optimizer of Adam. The transfer loss and attention weight coefficients are adjusted through multiple iterations during model training to ensure the accuracy and convergence of the recommendation system. After completing the training, the model predicts the input user data, recommends the corresponding POI, and evaluates the performance of the recommendation results. Advanced recommendation models have been introduced, such as the deep multi-view sequential model (DeepMove), spatial-temporal attention network (STAN), and geographic information embedding (GeoIE). Seven types of POIs with high user click-through rates are selected for testing, namely leisure and entertainment, transportation facilities, food and beverage, healthcare, business residences, tourist attractions, and science, education, and culture. The test results are shown in Figure 9.

Figure 9

Recommendation coverage test results for different models. (a) Chengdu–Chongqing and (b) Shanghai–Beijing.

Figure 9(a) and (b) shows the recommended coverage of seven types of POI labels for four models between cities in Chengdu and Chongqing and Shanghai and Beijing. In Figure 9(a), the recommendation coverage rate of the research model in the POI of food and beverage is the highest at 96.72%, which is much higher than other models. Chongqing’s hotpot and Chengdu’s snack culture are typical representatives of this type of POI, indicating that the research model has advantages in capturing and recommending local cuisine and dining venues and can maintain high coverage of local specialty foods. In Figure 9(b), the recommended coverage rates for all seven types of POIs are relatively high, especially for commercial residential and educational cultural areas. The reason behind this is that Shanghai is the economic center of the country, and Beijing is the political center of the country. Therefore, the click-through rates for business residential and science, education, and culture in social network POI recommendations in both places are generally high. The research model can achieve a recommendation coverage rate of 95.17%. GeoIE and research models with better performance are selected for visual comparison using a confusion matrix, as shown in Figure 10.

Figure 10

POI recommendation obfuscation results for the two classes of models. (a) GeoIE and (b) our model.

Figure 10(a) and (b) shows the POI recommendation classification results of GeoIE and the research model. In Figure 10(a), among the 7 types of POI recommendations, only 4 types of POI classification recommendations score above 90 points, and 2 types of POIs score above 80 points. Among them, GeoIE has a high similarity in recommending food and transportation facilities, which can lead to recommendation bias. The average recommendation classification of the research model for 7 types of POIs is 95%, and the classification of each type of POI is very accurate, maintaining above 90 points. This shows that the proposed model performs more consistently in the multi-interest point recommendation task, especially in the cross-city recommendation, and is able to match user interest points more accurately. The results demonstrate the efficacy of GNN combined with the attention mechanism for the task of user interest point categorization, thereby substantiating the robust adaptability of the proposed model in complex social network structures. This study compares four models for recommending seven types of POIs based on recommendation time, as listed in Table 3.

In Table 3, in the recommendation test for seven types of POIs, the longest recommendation time for DeepMove is 6.32 s in transportation facilities, the shortest time is 3.15 s in tourist attractions, and the average running time is 4.68 s. GeoIE combines geographic location information to significantly improve recommendation efficiency, with an average recommendation time of 3.49 s for seven types of POIs. Relatively speaking, the average recommendation time for the research model is the shortest at 2.61 s. This value is reduced by 2.07, 1.8, and 0.88 s compared to DeepMove, STAN, and GeoIE, respectively. The proposed model greatly improves computational efficiency while ensuring the recommendation accuracy. Particularly in the recommendation of key POIs such as transportation facilities, food, and beverages, the proposed model can respond to the user’s needs quickly, and the recommendation efficiency is far better than other models.