CN112084427A - Interest point recommendation method based on graph neural network - Google Patents

Abstract

The invention discloses a method for recommending interest points based on a graph neural network. The steps are as follows: constructing a user-interest point interaction graph and a user social graph, the graph neural network learns graph structure information, and integrates collaboration information and a user's embedding vector into the user's embedding vector. Social information; use the k-means algorithm to cluster the points of interest by geographic location, embed the clustering results into a vector, connect the embedded vector obtained in the user-point of interest interaction graph, and input it into a neural network to get the interest Point embedding vector; build a neural network model to simulate the matrix factorization method in machine learning, and input the embedding vectors of users and interest points into the neural network model to predict the score according to the user's historical scores. The present invention embeds the collaboration information and the information in the social network in the vector representation of the user, embeds the collaboration information and the location information of the interest point in the vector representation of the interest point, and inputs the vector representation of the user and the interest point into the neural network for recommendation .

Description

A method for recommendation of points of interest based on graph neural network

technical field

The invention belongs to the technical field of neural networks and recommendation systems, and in particular relates to a method for recommending points of interest based on graph neural networks.

Background technique

In recent years, with the rapid development of mobile Internet technology and smart devices, users can easily obtain real-time personal location information. At the same time, location-based social networks (LBSNs) and point-of-interest recommendation technology (POI recommendation) have also received extensive attention. Collaborative filtering is a widely used method in recommender systems. Its basic idea is to find other users who are similar to a user's preferences by analyzing the user's interest preferences, and then synthesize the evaluation of these similar users' points of interest to predict the user's interest points. degree of preference. Learning vector representations of users and points of interest is a research hotspot in collaborative filtering recommendation. With the development of word embedding technology, many methods embed feature information describing users and POIs as vector representations of users and POIs, which capture the user's preferences and the features of POIs, but this method does not capture user-POI Collaboration information in the interaction graph, the resulting embedding vector is not enough to capture the impact of collaborative filtering. The graph neural network is proposed to learn the vector representation in the graph data, which can integrate the node information, edge information and topology structure in the graph. Its main idea is to use the neural network to iteratively transform and aggregate the feature information from the neighbor nodes. , node information can also be propagated on the graph after transformation and aggregation. Graph neural networks have been successful in learning vector representations due to their ability to integrate node information and topology. For example, using a graph neural network to learn user and item embedding vectors in a user-item interaction graph.

Location information is the basic attribute of POIs and a key factor that must be considered in POI recommendation algorithms. For example, users may tend to select points of interest that are physically close to themselves. Social relations have attracted a lot of attention in POI recommendation, and users can obtain or spread information through his friends, colleagues, etc., which means that users' potential social relations play a certain role when users filter information. Through research, Wang et al. found that the similarity of check-in behavior between two users has a strong relationship with their relevance in social networks. The Gowalla and Foursquare datasets report that more than 5%, 20%, and 40% of social, neighbor, and geographic friends have a similarity greater than 0.2, proving that users' friends have a certain influence on users' decision-making behavior. Many existing models only consider the friends in the user's social network, ignoring that the user's decision-making behavior may be affected by the user's friends' friends, and the information transfer between users is not limited to the user's friends. There is a phenomenon of information diffusion, and only considering the friends in the user's social network is not enough to capture the impact of the information transmitted in the social network.

SUMMARY OF THE INVENTION

In view of the deficiencies of the prior art, the technical problem solved by the present invention is to provide a method for recommending points of interest based on a graph neural network, which is constructed into a social graph and a user-point of interest interaction graph, and the collaboration information is embedded in the vector representation of the user. At the same time, the collaboration information and the location information of the interest point are embedded in the vector representation of the interest point, and then the vector representation of the user and the interest point is input into the neural network for recommendation.

In order to solve the above-mentioned technical problems, the present invention realizes through the following technical solutions:

The present invention provides a method for recommending points of interest based on a graph neural network, comprising the following steps:

Step S1: construct a user-interest point interaction graph and a user social graph, and the graph neural network learns the graph structure information and integrates collaboration information and social information in the user's embedding vector;

Step S2: Use the k-means algorithm to cluster the points of interest by geographic location, embed the clustering results into a vector, connect the embedded vectors obtained in the user-point-of-interest interaction graph, and input it into a neural network to obtain Interest point embedding vector;

Step S3: Construct a neural network model, simulate the matrix decomposition method in machine learning, and input the embedded vectors of users and points of interest into the neural network model to perform score prediction according to the user's historical scores.

Optionally, in step S1, an interaction aggregator is used to learn on the user-POI interaction graph and generate a latent vector of the POI space, and the social aggregator learns to generate a latent vector of the social space on the social graph, and the POI space The latent vector of , and the latent vector of the social space are combined to obtain the embedding vector of the user.

Further, in the step S2, the interest point clustering based on k-means first randomly selects k interest points as the initial cluster centers, then calculates the distance between each interest point and each cluster center, and calculates the distance between each interest point and each cluster center. Each interest point is assigned to the cluster center closest to it.

Optionally, in the step S2, the processing process of the interest point clustering based on the k-means algorithm is as follows:

1) Randomly select k interest points from the dataset as initial cluster centers;

2) Calculate the Euclidean distance ρ from the remaining interest points to the cluster center, and put the closest interest points into the category to form a new category. The calculation method of ρ is:

where v _i = (lat _i , lon _i ) and v _j = (lat _j , lon _j ) are two of the interest point datasets V = {v ₁ , v ₂ ,..., v _n } Points of Interest;

3) Take the mean of all interest points of the current class as the center point, and update the interest point closest to the center of the class;

4) Until the objective function converges or the cluster center remains unchanged, otherwise it will transfer to step 2);

5) Output the clustering results.

From the above, the interest point recommendation method based on the graph neural network of the present invention has the following beneficial effects:

1. A new method for POI recommendation based on graph neural network is proposed, which can learn information on user-POI interaction graphs and social graphs.

2. The user's evaluation and interaction are taken into account in the user-point-of-interest interaction graph, which better captures collaboration information.

3. All levels of friends in the user's social relationship and the intimacy between users are considered.

4. The experimental evaluation of the effect and performance of the algorithm is carried out on the Yelp data set, which verifies the effectiveness of the method of the present invention.

The above description is only an overview of the technical solutions of the present invention, in order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the description, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand , the following detailed description is given in conjunction with the preferred embodiments and in conjunction with the accompanying drawings.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the accompanying drawings of the embodiments will be briefly introduced below.

1 is a model diagram of a method for recommending points of interest based on a graph neural network of the present invention;

Fig. 2 is the embedding mode of the third _- order friend of user u1;

Fig. 3 is the overall frame diagram of the present invention;

Fig. 4 is the change situation diagram of MAE;

Figure 5 is a graph of the variation of RMSE.

Detailed ways

The specific embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. As a part of this specification, the principles of the present invention will be described through examples. Other aspects, features and advantages of the present invention will become clear at a glance through the detailed description. In the figures to which reference is made, the same or similar parts are denoted by the same reference numerals in different figures.

Graph embedding-based POI recommendation includes four parts: user embedding vector modeling, POI embedding vector modeling, rating prediction, and model training. The present invention constructs the user-interest point interaction graph and the social graph respectively, and then uses the interaction aggregator to learn on the user-interest point interaction graph and generate the latent vector of the interest point space, and the social aggregator learns to generate the potential vector of the social space on the social graph. Finally, the latent vector of the interest point space and the latent vector of the social space are combined to obtain the user's embedding vector; in the modeling part of the interest point vector, the collaboration information and location information of the interest point existing in the interaction data are considered, and the location information is used The k-means algorithm performs clustering, so that the interest points in each cluster have the same location attribute label, and then embed them into the vector to obtain the latent vector of the location space, and the learning of the latent vector of the user space is the same as the user embedding vector. The learning of the latent vector of the interest point space is the same. Finally, the latent vector of the position space and the latent vector of the user space are integrated to obtain the embedded vector of the interest point; the score prediction part is the integration of the user and the interest point modeling part, which is input to the neural network. The score prediction is performed in the model; the model training part specifies an objective function for optimization and learns the parameters of the model.

The related work of the present invention is mainly divided into POI recommendation based on social information, POI recommendation based on geographic information and recommendation based on graph embedding.

For POI recommendation based on social information, POI itself does not have social attributes, but users who visit POIs have social attributes, so POIs are given social attributes through the social connection between users who visit POIs. The user's behavior has a certain similarity with the behavior of friends, and users may visit the points of interest recommended by friends, so social relations can be used to improve the performance of point of interest recommendation. Li et al. proposed that users' decisions are not only affected by friends in social networks, but also by geographical friends and neighbor friends (geographical friends are friends who have checked in at the same place, and neighbor friends are users who are geographically close to the target user), and then combine linear aggregation and random walk for recommendation. In the prior art, the similarity of users is integrated into the user-based collaborative filtering algorithm, and some of them use the similarity of users as the weight of the regularization term or latent factor model.

For POI recommendation based on location information, the location attribute is a natural attribute of POI, and it has always been a basic factor that must be considered in POI recommendation algorithms. Some use the geographic correlation, social correlation and category correlation between users and POIs to recommend POIs. Existing techniques integrate convolutional neural networks into probabilistic matrix factorization to model geographic influence, social relations, and review information. Wang et al. combined users' personal interests and target area population preferences, used the co-occurrence pattern and content description of points of interest, and smoothed the crowd preferences through the spatial pyramid index structure, so as to alleviate the user's travel problems due to the lack of user travel. The data sparsity problem caused by the behavior data of the location, so as to obtain a more accurate recommendation effect. Xing et al. proposed that user check-in presents a phenomenon of geographic clustering, and the user's preference for several points of interest adjacent to the point of interest is used to replace the user's preference for points of interest, and then combined with matrix decomposition for recommendation.

Based on the recommendation of graph neural network, GC-MC describes the relationship between users and items as a bipartite graph. Two multi-connected graph convolution layers are used to aggregate the feature information of users and items. Since GCN learns high-quality users and items The effectiveness of GC-MC has achieved good recommendation results. However, in order to distinguish each node, GC-MC uses one-hot encoding as the vector input of the node, so that the input vector dimension is proportional to the total number of nodes, so it cannot scale to large graphs. Different from GC-MC, the STAR-GCN model learns low-dimensional embedding vectors of users and items to input into the network. In order to predict the embedding vectors of cold-start nodes, STAR-GCN shields some embedding vectors of users and items, and encodes them by graph- The decoder reconstructs the masked embedding vector, which effectively alleviates the cold-start problem. Some propose to construct the user-item interaction data into a bipartite graph, use a graph neural network to learn the embedding vectors of users and items, and explicitly inject collaboration information in the embedding vectors of users and items in the form of embedding propagation. The existing multi-order neighbor nodes of the nodes in the graph are considered, but only the first-order neighbor nodes of the nodes are considered.

How to combine the information on the user-POI interaction graph and the social graph is the key to modeling the user embedding vector, so the present invention uses two aggregators, the interaction aggregator and the social aggregator, on the user-POI interaction graph and the social graph, respectively. It learns the user latent vector of the interest point space and the user latent vector of the social space, and then integrates these two vectors to obtain the user's embedding vector.

The user-POI interaction graph includes not only the interaction between the user and the POI but also the user's evaluation of the POI. For learning the user latent vector of the POI space, the present invention proposes a method to jointly capture the interaction and evaluation in the user-POI interaction graph, considering the POIs that the user has interacted with and the user's evaluation of the POIs, better The collaboration information in the user-POI interaction graph is aggregated. To mathematically represent this aggregation, the present invention uses the following function:

where C(i) is the set of interest points that user _ui has interacted with (the neighbor nodes of _ui in the user-interest point interaction graph), and _{X ia} represents the evaluation-aware interaction vector between _ui and interest point va , Aggre _items is the point of interest aggregation function. In addition, σ is the nonlinear activation function, and W and b are the weights and biases of the neural network. Next, the evaluation-aware interaction vector X _ia and the aggregation function Aggre _items are introduced in detail.

When users interact with POIs, they can express their own evaluations (reviews or ratings) for POIs, and the evaluation information expresses the user's preference for POIs, which can better capture the collaboration information in the user-POI interaction graph. In order to model the evaluation information, for each evaluation information r, the evaluation embedding vector is used to represent each evaluation information r as a dense vector. For example, in a five-point scoring system, for each rating, er represents the embedding _vector . For the interaction and evaluation information r between the user _ui and the interest point va, the present invention uses the multi-layer perceptron to model and evaluate the perceived interaction vector X _ia as the combination of the interest point initial embedding _{vector q a and} the evaluation embedding vector er. The layer perceptron takes as input the concatenated _vector of q _a and er . The output of the multilayer perceptron is the evaluation-perceptual interaction vector X _ia between _ui and v _a , which is expressed as:

为两个向量的连接操作，g_v代表多层感知器。in

is the concatenation operation of two vectors, g _v represents the multilayer perceptron.

中 向量的元素级上的平均值。基于均值的聚合器是局部谱卷积的线性近似， 如下所示：For Aggre _items , a commonly used aggregation function is the mean operation, that is, taking

The element-wise mean of the vector in . The mean-based aggregator is a linear approximation of the local spectral convolution as follows:

是对于所有兴趣点向量基于均值的聚合操作，其假设所有 的交互对用户u_i的嵌入向量的贡献都是相同的。然而，每个交互对于用户 嵌入向量的贡献都可能有很大的不同。因此，通过给每个交互分配一个权 重，允许每个交互对用户的嵌入向量做出不同的贡献。为了缓解基于均值 的聚合器的局限，受注意机制的启发，调整α_i，分配个性化的权重给每个 (v_a，u_i)：α _i is fixed as

is a mean-based aggregation operation for all interest point vectors, which assumes that all interactions contribute the same to the embedding vector of user _ui . However, each interaction's contribution to the user's embedding vector can vary widely. Therefore, by assigning a weight to each interaction, each interaction is allowed to contribute differently to the user's embedding vector. To alleviate the limitations of mean-based aggregators, inspired by the attention mechanism, adjust α _i , assigning individualized weights to each ( _va , _ui ):

α _ia represents the attention weight of the interest point v _a that _ui has interacted with, that is, the contribution to the latent vector on the space of interest points of u _i . The present invention uses a two-layer neural network to parameterize α _ia , which is called an attention network. Note that the input to the network is X _ia and the initial embedding vector p _i of the target user _ui , and the network is defined as follows:

The final attention weights are obtained by normalizing the above attention scores using the Softmax function, which can be interpreted as the contribution of the interaction to the latent vector on the space of user u _i interest points:

Due to the principle of social relevance, a user's preferences are similar to or influenced by those of friends in a social network. At the same time, the intimacy between users and their friends in the social graph further affects the user's decision-making behavior. Therefore, the present invention introduces an attention mechanism to model the intimacy of users with their friends, and then aggregate their information. This section is described in two parts: first-order aggregation and higher-order aggregation. The first-order aggregation considers the neighbor nodes in the user's social graph. The user may not only be influenced by friends in the social network, but friends of friends may also affect the user's decision-making behavior. Therefore, the user's social graph is considered in the high-order aggregation. The friends of the friends are called high-level friends in the present invention.

In order to represent the latent vector of the user from the social network, the present invention introduces the latent vector of the social space, which aggregates the neighbor user nodes in the social graph. The latent vector expression of the social space of the user _ui is as follows:

中向量的元素级上的均值。如 下所述：N( _i ) is the neighbor node of ui, and Aggre _neigbhors is the aggregation function of the user's neighbor node. Aggregate functions commonly used by Aggre _neigbhors are

The element-wise mean of the vector in . as described below:

是对所有邻居节点的基于均值的聚合器。他假设所 有的邻居节点对于u_i的贡献是相同的。然而在社交网络中用户和朋友之间 的亲密度不同，他们之间的偏好可能也会有所不同，更加亲密的朋友可能 会分享更多的信息，具有更相似的偏好。因此本发明引进具有两层的神经 网络的注意力机制来建模用户之间的亲密度，通过给用户的每个朋友分配 一个亲密度，允许每个朋友对用户的嵌入向量做出不同的贡献，用户与朋 友之间越亲密，对用户的嵌入向量贡献越大，如下所述：where β _i is fixed as

is a mean-based aggregator over all neighbor nodes. He assumes that all neighbor nodes contribute the same to _ui . However, in a social network, the intimacy between users and friends is different, and their preferences may also be different, and more intimate friends may share more information and have more similar preferences. Therefore, the present invention introduces an attention mechanism with a two-layer neural network to model the intimacy between users, by assigning an intimacy to each friend of the user, allowing each friend to make different contributions to the user's embedding vector , the closer the user and friend are, the greater the contribution to the user's embedding vector, as follows:

β _io is the intimacy between user u _o and user _ui .

In order to consider the influence of the user's high-level friends on the user's decision-making, the present invention aggregates the information from the high-level friends. At step l-1, the latent vector iteratively expresses the social space of user _ui as:

As shown in Figure 2, user u ₁ embeds the information of 3rd-order friends, and it can be seen from the figure that user u ₆ has an influence on the embedding vector of user u ₁ .

和

输入到MLP中得到用户嵌入向量，定义如下：In order to better learn the user's embedding vector, it is necessary to consider the latent vector of the user's social space and the latent vector of the interest point space. The user-interest point interaction graph and the social graph provide different aspects of information. Will

and

Input into the MLP to get the user embedding vector, which is defined as follows:

c ₂ =σ(W ₂ ·c ₁ +b ₂ ) (15)

......

h _i =σ(W ₁ ·c _l-1 +b ₁ ) (16)

where l is the index of the hidden layer.

The POI embedding vector modeling considers two parts of information: the collaboration information in the user-POI interaction graph and the location information of the POIs.

For each point of interest v _j , aggregate information from the set B(j) of users who have interacted with v _j . Even for the same POI, each user may express different evaluations during the user-POI interaction. These evaluations from different users can capture the features of the same POI and help to build the embedding vector of POIs. mold. For the interaction of the user ut with _r to the interest point vj, the present invention introduces the interaction vector f _jt of evaluation perception, by inputting the initial embedding vector p _t of the user and the embedding _{vector er} of evaluation information into the multilayer perceptron obtained from g _u , g _u fuses the user’s interaction information and evaluation information on points of interest, as follows:

本发明提出聚合B(j) 中用户的评价感知的交互向量。用户的聚合函数标记为Aggre_users，用来聚 合

中用户的评价感知的交互向量：For the interest point v _j , in order to learn the user-space embedding vector

The present invention proposes to aggregate the user's evaluation-aware interaction vector in B(j). The user's aggregation function is marked as Aggre _users and is used to aggregate

The user's rating-aware interaction vector in :

In addition, an attention mechanism of a two-layer neural attention network is introduced to distinguish the importance weights μ _jt , f _jt and q _j of each user as input:

μ _jt can capture the different influences of different users from the user-POI interaction graph.

The location information modeling of interest points firstly uses the k-means algorithm to cluster the location information of interest points, each interest point can get the corresponding cluster label, and then embeds the cluster label of the interest point into the vector to get the location latent vector of space.

The interest point clustering based on k-means is to randomly select k interest points as the initial clustering center. Then calculate the distance between each interest point and each cluster center, and assign each interest point to its nearest cluster center. The cluster centers and the points of interest assigned to them represent a cluster, and the center of each cluster is recalculated based on the existing points of interest in the cluster. This process will repeat until a certain termination condition is met. Termination conditions can be that no (or a minimum number) of objects are reassigned to different clusters, no (or a minimum number) of cluster centers change again, and the sum of squared errors is locally minimized.

The processing process of interest point clustering based on k-means algorithm is as follows:

i) Randomly select k interest points from the dataset as initial cluster centers.

ii) Calculate the Euclidean distance ρ from the remaining interest points to the cluster center, and put the closest interest points into the category to form a new category. The calculation method of ρ is:

where v _i = (lat _i , lon _i ) and v _j = (lat _j , lon _j ) are two of the interest point datasets V = {v ₁ , v ₂ ,..., v _n } Points of Interest.

iii) Take the mean of all interest points of the current class as the center point, as shown in Figure 2, and update the interest point closest to the center of the class.

iv) Until the objective function converges or the cluster centers do not change, otherwise move to ii).

v) Output the clustering results.

The clustering results are then embedded into a vector. For example, if the clustering result is 10 classes, then k∈{1,2,3,...,10}, introducing a _k as the latent vector of the clustering result.

然后将其输入到多层感知器中得到兴 趣点嵌入向量。如下所述：In order to better learn the embedding vector of the interest point, the present invention integrates the latent vector a _k of the position space of the interest point and the latent vector of the user space

It is then fed into a multilayer perceptron to get the interest point embedding vector. as described below:

d ₂ =σ(W ₂ ·c ₁ +b ₂ )

...

z _j =σ(W·c _l-1 +b ₂ )

The present invention will design recommendation tasks to learn model parameters. Recommendation tasks include a variety of recommendation tasks such as POI ranking and scoring prediction. The present invention applies the proposed graph neural network-based POI recommendation model to the recommendation task of scoring prediction, using the embedding vectors of users and POIs ( _hi and z _j ), first concatenate them and then feed them to MLP for rating prediction as follows:

g ₂ =σ(W ₂ ·g ₁ +b ₂ )

...

g _l-1 =σ(W _l ·g _l-1 +b _l )

r′ _ij = W ^T ·g _l-1

where l is the index of the hidden layer and _r'ij is the predicted score.

In order to learn the parameters of the model, an objective function needs to be specified for optimization. Since the main task of the present invention in this work is scoring prediction, a commonly used objective function is adopted:

The present invention is tested using the Yelp data set, which is one of the largest review websites in the world, which allows users to review or rate businesses. The experimental data of the present invention intercepts data whose longitude is between -112.0 and -111.9 and latitude between 33.3 and 33.45 as experimental data. In order to ensure data quality, the number of user ratings is less than 5, and the number of points of interest is less than rated. on 5 users and points of interest. The vertices in the social graph represent users, and the edges represent the friend relationship between users and users. The vertices in the user-POI interaction graph are composed of users and POIs, and the edges represent the existence of scoring behavior between users and POIs. The final data includes 3870 users, 2301 points of interest, and 57756 rating records, and the social graph contains 78577 edges.

The method for recommending points of interest based on a graph neural network (referred to as Graph-POIR) is compared with five models respectively, and now the model proposed by the present invention and the five models are introduced as follows:

PMF: Probabilistic Matrix Factorization uses only the user-item rating matrix and uses a Gaussian distribution to model latent factors for users and items.

GraphRec: It is a newly proposed graph neural network recommendation model in recent years. The model takes into account first-order friend relationships between users and users in the social graph and ratings and interaction information in the user-POI interaction graph.

NeuMF: This recommendation model is a matrix decomposition implemented by a multi-layer neural network, and is a new classical recommendation model proposed in recent years. The model uses multiple hidden layers to capture their nonlinear feature interactions, which can more effectively capture the implicit/latent associations between users and items.

GCMC: The model uses the GCN encoder to obtain the vector representation of users and items, but only considers the first-order neighbor nodes in the user-item graph.

NGCF: The user-item interaction data is constructed into a bipartite graph structure. By embedding the propagation layer, the collaborative information is explicitly encoded in the form of high-order connections to obtain the vector representation of users and items.

Graph-POIR: The model proposed by the present invention considers the interaction information and rating information in the user-interest point interaction graph in the user embedding vector and the interest point embedding vector. The location information of the POI is also embedded in the POI embedding vector, and the user's high-level friends and intimacy are considered in the user embedding vector.

The recommendation method of points of interest based on graph neural network is implemented using Python language based on Pytorch framework. The computer configuration is CPU i7-8700K 3.7GHz, and the operating system is Ubuntu 18.04.1. In the experiments, the learning rate is set to 0.001 and RMSprop is adopted as the optimizer. Parameter initialization uses a Gaussian distribution (mean and standard deviation of 0 and 0.01, respectively) to initialize the parameters of the model. 80% and 60% of the data are randomly selected as the training set, and the remaining 20% and 40% of the data are used as the test set, and the influence of the training set and the test set in different proportions on the recommended results of the method of the present invention is considered.

In the k-means clustering algorithm, the clustering number k of the location factor is 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, respectively. The size of the embedding layer is set to 16, 32, 64, 128, and 256 for experiments, and the test results show that the recommended performance is the best when the size of the embedding layer is 64. The activation function of the hidden layer in the multilayer perceptron uses the Relu function.

In order to evaluate the prediction accuracy of the recommended method proposed by the present invention, mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators in the experiment.

where _ri ' is the predicted value, and _ri is the actual value.

Tables 1 and 2 describe the RMSE and MAE of NGCF, GCMC, NeuMF, GraphRec, PMF and Graph-POIR on the dataset Yelp when k=40 and the user's second-order friends are considered. 80% of the data in Table 1 is used as the training set, the remaining 20% of the data is the test set, 60% of the data in Table 2 is used as the training set, and the remaining 40% of the data is the test set. Experimental results show that Graph-POIR reduces the rating prediction error and significantly improves the recommendation performance. As shown in Table 1, compared with GraphRec, the RMSE and MAE of Graph-POIR are reduced by 0.59% and 1.38%, respectively, the reason is that Graph-POIR not only considers the intimacy between users and users’ high-order friends, but also considers points of interest. The location information is embedded into the embedding vector of interest points, while GraphRec does not consider the influence of the user's higher-order friends on the user's behavior. Because NGCF, GCMC and NeuMF do not consider the user's social information, the recommendation effect is not as good as GraphRec, which also reflects that the method of considering the user's social information in the present invention can improve the recommendation performance. Compared with PMF and NeuMF, GraphRec, NGCF, GCMC, Graph-POIR use graph embedding technology, which reflects the effectiveness of graph embedding technology in POI recommendation. PMF and NeuMF only use the scores between users and points of interest, and the recommendation performance of NeuMF is better than that of PMF, indicating that neural networks can improve the recommendation performance in recommendation.

In order to further test the performance of the Graph-POIR proposed by the present invention under different parameters k in the k-means algorithm, Fig. 4 and Fig. 5 respectively show the changes of the RMSE and MAE of the Graph-POIR. It can be seen from Figure 4 and Figure 5 that when k=70, the MAE of Graph-POIR achieves the minimum value, and when k=40, the RMSE of Graph-POIR achieves the minimum value. With the change of k value, the change curves of MAE and RMSE are irregular. The reason for this phenomenon is that the initial clustering center of k-means algorithm has a great influence on the clustering results, while the initial clustering center of k-means algorithm has a great influence on the clustering results. The class center is randomly selected, and in future work, we will consider improving the selection of initial nodes in the k-means algorithm.

Table 3 shows that Graph-POIR considers the impact of higher-order friends on the algorithm when users embed vectors. Graph-POIR-1 represents that Graph-POIR only considers first-order friends in the user embedding vector, and Graph-POIR-2 represents that Graph-POIR considers the user's second-order friends in the user embedding vector. The experimental results are shown in Table 3. The RMSE and MAE of Graph-POIR-2 are higher than those of Graph-POIR-1, which reflects the necessity of considering users' multi-order friends when modeling user embedding vectors.

Table 3 Changes in RMSE and MAE

The invention proposes a method for recommending points of interest based on a graph neural network. First, a user-interest point interaction graph and a user social graph are constructed. The graph neural network learns the graph structure information and integrates collaboration information and social information in the user's embedding vector. In addition, the high-level friends in the social graph and the intimacy between friends are considered; at the same time, the k-means algorithm is used to cluster the points of interest by geographic location, and the clustering results are embedded in the vector, and the connection is made through the user-interest. The embedded vector obtained in the point interaction graph is input into the neural network to obtain the embedded vector of interest points; finally, a neural network model is constructed to simulate the matrix decomposition method in machine learning, and the scoring prediction is performed according to the user's scoring behavior. The experimental results show that, compared with the existing point-of-interest recommendation model, the recommendation model proposed by the present invention achieves a better recommendation effect. In the following work, we will consider improving the selection of initial nodes in the k-means algorithm.

The above descriptions are only the preferred embodiments of the present invention, of course, it cannot limit the scope of rights of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, Several improvements and changes are made, and these improvements and changes are also regarded as the protection scope of the present invention.

Claims (4)

1. a kind of interest point recommendation method based on graph neural network, is characterized in that, comprises the following steps: Step S1: construct a user-point of interest interaction graph and a user social graph, and the graph neural network learns the graph structure information and integrates the collaboration information and social information in the user's embedding vector; Step S2: Use the k-means algorithm to cluster the points of interest by geographic location, embed the clustering results into a vector, connect the embedded vectors obtained in the user-point-of-interest interaction graph, and input it into a neural network to get Interest point embedding vector; Step S3: Constructing a neural network model, simulating the matrix decomposition method in machine learning, and inputting the embedded vectors of users and points of interest into the neural network model to perform score prediction according to the user's historical scores. 2. The method for recommending points of interest based on a graph neural network as claimed in claim 1, wherein in step S1, an interaction aggregator is used to learn and generate a latent vector of the point of interest space on the user-point of interest interaction graph, The social aggregator learns to generate the latent vector of the social space on the social graph, and combines the latent vector of the point of interest space and the latent vector of the social space to obtain the embedding vector of the user. 3. The method for recommending interest points based on a graph neural network as claimed in claim 2, wherein in the step S2, k-means-based interest point clustering first randomly selects k interest points as the initial clustering center, and then calculate the distance between each interest point and each cluster center, and assign each interest point to the cluster center closest to it. 4. The method for recommending points of interest based on a graph neural network as claimed in claim 3, wherein in the step S2, the processing process of the point of interest clustering based on the k-means algorithm is as follows: 1) Randomly select k interest points from the dataset as initial cluster centers; 2) Calculate the Euclidean distance ρ from the remaining interest points to the cluster center, and put the closest interest points into the category to form a new category. The calculation method of ρ is:

where v _i = (lat _i , lon _i ) and v _j = (lat _j , lon _j ) are two of the interest point datasets V = {v ₁ , v ₂ , ..., v _n } Points of Interest; 3) Take the mean of all interest points of the current class as the center point, and update the interest point closest to the center of the class; 4) Until the objective function converges or the cluster center remains unchanged, otherwise it will transfer to step 2); 5) Output the clustering results.

Images