CN111738447A - A mobile social network user relationship inference method based on spatiotemporal relationship learning - Google Patents

Abstract

The invention provides a method for inferring the user relationship of a mobile social network based on spatio-temporal relationship learning, which considers the mobility and sociability of individuals at the same time. Considering the effectiveness of the social network structure on social connection prediction, the model firstly constructs a preliminary social graph based on the mobility of users, then extracts the social network structure characteristics of the user pairs from the preliminarily constructed social graph, and finally carries out friend relation inference by integrating the characteristics of the mobility and the sociability. Once the training of the model is completed, different scenes can be well migrated to predict the friendship between users. Experiments on two real-world data sets have shown that our method is consistently superior to existing methods. Furthermore, our model is also valid for relationships with little check-in data and no face-to-face events.

Description

A mobile social network user relationship inference method based on spatiotemporal relationship learning

technical field

The invention relates to a method for inferring user relationship in a mobile social network based on time-space relationship learning, which utilizes the user's mobile information.

Background technique

In recent years, with the popularity of mobile social networking applications such as Facebook, Twitter and Weibo, users can timely publish the places of interest they are visiting (an Internet celebrity restaurant, a tourist attraction, etc.) to share with friends. Although this type of social networking brings great convenience to people's making friends, there is a risk of revealing the user's social relationship. Users of mobile social networks are also becoming aware of this. For example, a large-scale study of Facebook users showed that the proportion of Facebook users hiding their friend lists increased from 17.2% in 2010 to 56.2% in 2011. However, few users know that their friends can be inferred from the check-in records with temporal and spatial relationships, thereby accurately revealing the hidden social relationships between users.

Existing social relationship inference methods based on spatiotemporal relationships are mainly divided into two categories: the first is a heuristic method based on feature selection, which selects effective features such as the number of encounters, the number of shared locations, and the popularity of shared locations as a measure of user popularity. A measure of friendship. But these methods impose many assumptions and limitations on the required movement data, which greatly narrows their applicability. For example, almost all existing effective methods can only be used when two individuals share a location. Another type of method is the method based on feature learning, which uses the method of machine learning to vectorize the user's mobile features, and uses the similarity between the vectors as the criterion to infer whether the user has a friend relationship. However, this kind of method is aimed at individual modeling and loses the time information in the mobile data, so it cannot directly obtain more accurate relationship inference.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a model that can accurately infer the user's social relationship by using the user's movement information.

In order to achieve the above object, the technical solution of the present invention is to provide a method for inferring user relationship in a mobile social network based on spatiotemporal relationship learning, which is characterized in that it includes the following steps:

Step 1. Extract the interactive behavior feature between pairs of users, and use the feature to infer whether there is a friend relationship between the two users, including the following steps:

Step 101: Divide the POIs involved in the check-in of all users of the mobile data into an I×J grid according to the latitude and longitude, and at the same time divide the time into M time segments to construct an I×J×M space-time matrix STD, wherein, The time dimension is divided into equal time slices of length τ, the space dimension is evenly divided into grids of equal size, and each grid is recursively divided into four equal grids until POIs of interest points The number of in each grid is less than the threshold σ;

式中三元组

表示第m时间段内用户u_a和u_b在第i行第j列的位置网格中移动信息的统计量；Step 102: Project the trajectory of each user pair (u _a , u _b ) into the space-time matrix STD, and each user's check-in can be projected into a specific square. For each square, calculate: the The number of interest points _na visited by the user _{u a} in the time period; the number of interest points n _{b visited by} the user _{ub a} , u _b ) space-time matrix

The triples in the formula

Represents the statistics of the movement information of users u _a and u _b in the position grid of the i-th row and the j-th column in the m-th time period;

Step 103: Encode the space-time matrix O _(a,b) of each pair of users (u _a , u _b ) into a low-dimensional vector, and use the low-dimensional vector to calculate the probability that user u _a and user _ub have a friend relationship , obtain the initial social relationship graph G=(U, E), U is the set of vertices in the graph, representing all users with mobile information; E is the set of edges, representing the friendship between two users, where, The size of the space-time matrix O _(a,b) is adjusted by the parameters σ and τ;

的步骤如下：Step 2. Extract a k-reachable subgraph for each user to describe the network structure between users. For a given initial social relationship graph G=(U, E), define the extracted user pair (u _a , u The k-reachable subgraph of _b )

The steps are as follows:

初始化为空图；Step 201, set the path length to 2, set the

initialized to an empty map;

然后删除社交关系图G中和

中均出现的点和边删除，除了用户u_a和用户u_b本身；Step 202: Find all paths with lengths of 2 between (u _a , u _b ) in the initial social relationship graph G, and express all the paths found as

Then delete the social graph G and neutralize

The points and edges that appear in all are deleted, except for user u _a and user u _b itself;

Step 203: Gradually increase the path length, and repeat step 202 until the path length exceeds k;

进行编码，基于累加的原则对相同长度路径的编码，对不同长度的路径的编码结果进行拼接，实现对用户对k-可达子图的向量化，获得用户对的综合特征向量；Step 3. According to the initial social relationship graph G, for each pair of users (u _a , u _b ) k-reachable subgraph

Encoding, encoding paths of the same length based on the principle of accumulation, and splicing the encoding results of paths of different lengths, realizing the vectorization of the k-reachable subgraph of the user, and obtaining the comprehensive feature vector of the user pair;

Step 4. Use the classifier to perform 0/1 classification according to the comprehensive feature vector of the user pair, where 1 is a friend and 0 is not a friend, and the latest predicted social graph is obtained.

Step 5: Use the latest social graph to update the structural features of the user pair, and then re-predict until the prediction result no longer changes, that is, the final predicted social network graph is obtained.

使其与编码器的输入O_(a,b)接近，该训练过程的优化目标为：Preferably, in step 103, the space-time matrix O _{(a, b) is} input into an auto-encoder with R hidden layers, and the auto-encoder encodes it into a d-dimensional vector to obtain a reconstructed space-time matrix

To make it close to the input O _(a,b) of the encoder, the optimization objective of this training process is:

表示混合网络中自编码器网络的损失函数，即尽可能地使得解码后的重构时空矩阵

与编码前的时空矩阵O_(a,b)相同。；In the formula,

Represents the loss function of the autoencoder network in the hybrid network, that is, the reconstructed spatiotemporal matrix after decoding as much as possible

The same as the space-time matrix O _(a,b) before encoding. ;

为：The autoencoder uses supervised training to reconstruct and distinguish the encoding process, that is, add a classification network to the autoencoder to monitor its encoding process, and the loss function of the process

for:

表示预测结果，即分类网络的输出结果；y表示样本标签；n表示训练样本的个数，即训练数据集中涉及到的用户对的个数。In the formula,

represents the prediction result, that is, the output result of the classification network; y represents the sample label; n represents the number of training samples, that is, the number of user pairs involved in the training data set.

In order to obtain a more discriminative vector representation, the following constraints are imposed on the integrated hybrid network:

表示整个混合网络的总体损失函数；In the formula,

represents the overall loss function of the entire hybrid network;

Once trained, the encoder is taken from the auto-encoder network for encoding and preliminary relation inference for any pair of user spatiotemporal relation matrices in the user set.

The present invention provides a new social relationship inference model, which simultaneously considers mobility and sociality between individuals. Considering the effectiveness of social network structure for social connection prediction, the model first constructs a preliminary social graph based on the user's mobility, then extracts the social network structure features of user pairs from the preliminary constructed social graph, and finally integrates the mobility and social Friend relationship inferences were carried out based on the characteristics of two aspects of sex. Once our model is trained, it can better transfer different scenarios to predict friend relationships between users. Experiments on two real-world datasets demonstrate that our method consistently outperforms existing methods. Furthermore, our model is also effective for relations with little check-in data and no encounter events.

Description of drawings

Fig. 1 is a schematic diagram of user mobile relationship mapping friend relationship;

Figure 2 is a system structure diagram of a social relationship inference model;

3 is a schematic diagram of an interactive behavior feature extraction process;

4 is a schematic diagram of a k-reachable subgraph encoding process;

Figure 5(a) is a graph showing the effect of spatial granularity on the accuracy of friend relationship inference in the Brightkite dataset;

Figure 5(b) is a graph showing the effect of spatial granularity on the accuracy of friend relationship inference in the Gawalla dataset;

Figure 6(a) is a graph showing the effect of time granularity on the accuracy of friend relationship inference in the Brightkite dataset;

Figure 6(b) is a graph showing the effect of time granularity on the accuracy of friend relationship inference in the Gawalla dataset;

Figure 7(a) is a graph showing the effect of the feature vector dimension on the accuracy of friend relationship inference in the Brightkite dataset;

Figure 7(b) is a graph showing the effect of the feature vector dimension on the accuracy of friend relationship inference in the Gawalla dataset;

Figure 8(a) is a graph showing the effect of the number of iterations on the accuracy of friend relationship inference in the Brightkite dataset;

Figure 8(b) is a graph showing the effect of the number of iterations on the accuracy of friend relationship inference in the Gawalla dataset;

Figure 9(a) is a graph showing the effect of the number of user check-ins on the accuracy of friend relationship inference in the Brightkite dataset;

Figure 9(b) is a graph showing the effect of the number of user check-ins on the accuracy of friend relationship inference in the Gawalla dataset;

Figure 10(a) shows the result of the accurate influence of users on the number of shared locations on the inference of friend relationships in the Brightkite dataset;

Figure 10(b) shows the result of the accurate influence of the number of shared locations on the friend relationship inference in the Gawalla dataset.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments. It should be understood that these examples are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

The present invention infers whether there is a friend relationship between users by utilizing the similarity of movement trajectories between users and the dissemination of social networks. Figure 1 is an abstraction of this real problem. The inference model is divided into two stages: 1. Construction of an initial social network graph: Based on the observation that movement trajectories among friends generally have higher similarity, an initial social network graph is constructed. 2. Obtaining hidden social relationships: Based on the phenomenon that the network topology between friends is more similar, mining hidden friend relationships that have similar preferences but do not have similar moving trajectories.

FIG. 2 shows the specific implementation content of a method for inferring user relationships in a mobile social network based on spatiotemporal relationship learning provided by the present invention, which mainly includes the following two stages:

Stage 1: Building a preliminary social graph

At this stage, we extract the interactive behavior features between user pairs, and use this feature to infer whether there is a friend relationship between two users. Usually, the characteristics of interaction behavior can be characterized by the statistical properties of the interaction between pairs of users, such as the number of encounters and the number of joint visits. However, there may be no common visit events or encounter events for friends, and not all common visits are of equal importance in inferring a user's friend relationship. Therefore, we propose a more comprehensive approach to learn complex interaction behavior features. We embed both the common and individual behaviors of the two users into the interaction behavior features. Furthermore, we argue that different locations and time periods have different predictive importance, and design a hybrid model to simultaneously generate compressed interaction behavior feature vectors and initial classification results, where each location's influence range and time interval are the model's parameter. This method can ensure that the compressed interactive behavior feature retains the user's original action feature as much as possible, and this feature has an important indicative role in the task of inferring friend relationships. The first stage includes the following steps:

Step 1. Build the space-time matrix

式中三元组

表示第m时间段内用户u_a和u_b在第i行第j列的位置网格中移动信息的统计量。Figure 3 shows a schematic diagram of constructing a user pair spatiotemporal matrix. Firstly, the points of interest (POIs) involved in the check-in of all users of the mobile data are divided into an I×J grid according to the latitude and longitude, and the time is divided into M time segments to construct an I×J×M space-time matrix (STD). Among them, the time dimension is divided into equal time slices of length τ; the space dimension is evenly divided into grids of equal size. Considering the large difference in the density of POIs in geographical areas, we recursively divide each The grids are divided into four equal grids until the number of POIs in each grid is less than a threshold σ. For each user pair (u _a , u _b ), its trajectory can be projected into the STD, and each user's check-in can be projected into a specific square, for each square, we calculate three indicator factors, are the number n _a of POIs visited by user u _a , the number n _b of POIs visited by user u _b , and the number n _{a,b of} POIs visited jointly by user u _a and user ub in the time period, respectively. From this, the space-time matrix of the user pair (u _a , u _b ) is obtained

The triples in the formula

It represents the statistics of the movement information of users u _a and u _b in the position grid of the i-th row and the j-th column during the m-th time period.

Step 2. Encoding of space-time matrix

This step encodes the space-time matrix O _{(a,b) of} each pair of users (u _a , ub ) into a low-dimensional vector, and uses this vector to calculate the probability that user u _a and user _ub have a friend relationship, where space-time The size of the matrix can be adjusted by the parameters σ and τ.

使其与编码器的输入O_(a,b)接近。该训练过程的优化目标为：We feed the spatiotemporal matrix into an autoencoder, and the encoder with R hidden layers encodes it into a d-dimensional vector, the output of the Rth layer of the encoder. The decoder with R hidden layers outputs the reconstructed spatiotemporal matrix from the d-dimensional vector

Make it close to the input O _(a,b) of the encoder. The optimization objective of this training process is:

表示混合网络中自编码器网络的损失函数，即尽可能地使得解码后的重构时空矩阵

与编码前的时空矩阵O_(a,b)相同，U表示训练样本中的所有用户。In the formula,

Represents the loss function of the autoencoder network in the hybrid network, that is, the reconstructed spatiotemporal matrix after decoding as much as possible

Same as the spatiotemporal matrix O _(a,b) before encoding, U represents all users in the training sample.

表示该对用户对(u_a,u_b)具有朋友关系的概率)来监控其编码过程。该过程的损失函数

为：The training of the autoencoder is unsupervised, which means that it only learns the input structure and does not have the function of feature selection or inclination, so the finally obtained d-dimensional feature vector only retains the user's original movement information, which cannot be guaranteed in friends. useful in relational inference tasks. To avoid this problem, we use supervised training to reconstruct and differentiate the encoding process, i.e. we add a classification network to the autoencoder (output

represents the probability that the pair of users (u _a , u _b ) has a friend relationship) to monitor its encoding process. The loss function of the process

for:

表示预测结果，即分类网络的输出结果；y表示样本标签；n表示训练样本的个数，即训练数据集中涉及到的用户对的个数。In the formula,

represents the prediction result, that is, the output result of the classification network; y represents the sample label; n represents the number of training samples, that is, the number of user pairs involved in the training data set.

To obtain a more discriminative vector representation, we impose the following constraints on the ensemble hybrid network:

表示混合网络的综合损失函数。In the formula,

Represents the synthetic loss function of the hybrid network.

Once trained, the encoder is taken from the auto-encoder network for encoding and preliminary relation inference for any pair of user spatiotemporal relation matrices in the user set.

Stage 2: Gaining Hidden Social Relationships

In the second stage, we consider the non-direct link features between user pairs, i.e. network structure features. Since the mobile data does not have a social topology map, we extract structural features based on the prediction results of the first stage (a semi-real social map). Considering that simple heuristic structural features are not universally applicable, we propose a new feature to describe the network structure among target users—K reachable subgraph, and encode the subgraph to obtain a structural feature vector. Finally, the encoded structure vector is combined with the corresponding interaction features to improve the accuracy of inference. Furthermore, we design an iterative process that simultaneously refines the structural feature vectors and the resulting final social relationship graph.

The second stage includes the following steps:

Step 3. Extract the local graph structure - K reachable subgraph

的步骤如下：Katz is commonly used to measure the network proximity between users, which considers the global view of user pairs in the social graph, that is, considers all possible paths between user pairs. However, we think this is unreasonable because longer paths are not only computationally complex but also negatively impact relational inference. Therefore, we adopt a local viewpoint on network proximity, that is, extract a k-reachable subgraph for each user to describe the network structure between users. For a given initial social relationship graph G=(U, E), we define a k-reachable subgraph that extracts user pairs (u _a , u _b )

The steps are as follows:

初始化为空图；Step 301, set the path length to 2, set the

initialized to an empty map;

然后删除社交关系图G中和

中均出现的点和边删除(除了用户u_a和用户u_b本身)。Step 302: Find out all paths with a length of 2 between (u _a , u _b ) in the initial social relationship graph G constructed in the first stage, and express all the paths found as

Then delete the social graph G and neutralize

The vertices and edges that appear in all are deleted (except for user u _a and user u _b itself).

Step 303 , gradually increase the path length, and repeat step 302 until the path length exceeds k.

Step 4. Iteratively update the predicted social graph

进行编码。基于累加的原则对相同长度路径的编码，对不同长度的路径的编码结果进行拼接，实现对用户对k-可达子图的向量化。图4详细地说明了K可达子图编码过程。According to the initial social relationship graph G constructed in the first stage, we have k-reachable subgraphs for each user pair (u _a , u _b )

to encode. Based on the principle of accumulation, the coding results of paths of the same length are spliced, and the vectorization of the k-reachable subgraph by the user is realized. Figure 4 illustrates the K-reachable subgraph encoding process in detail.

Use the classifier to perform 0/1 classification (where 1 is a friend, 0 is not a friend) according to the comprehensive feature vector of the user pair (synthesizing the interactive behavior features obtained in the first stage and the local structure feature vector obtained in the second stage), and obtain the latest The predicted social graph of .

Using the latest social graph, update the structural features of the user pair, and then re-predict until the prediction result does not change, that is, the final predicted social network graph is obtained.

The verification process of the present invention is as follows:

The experiment uses two public real data, Brightkite and Gowalla, both of which are location-based social networks, through which users can publish their time-location information and share them with their friends. This kind of dataset is divided into two parts: the check-in dataset and the user's social relationship graph. Table 1 shows the statistical information of the experimental dataset.

Table 1. Dataset Statistics

The experiments are compared with four popular inference methods based on moving distance, number of shared locations, random walks, and graph embeddings.

• Relationship inference based on the number of shared locations: This method extracts and utilizes the number of shared locations visited by a pair of users as a feature to infer whether there is a friend relationship between a pair of users. The experiment adopts svm classifier for supervised learning.

• Relationship inference based on travel distance. The method calculates the center position of each user's activity according to all the mobile information, and uses the Euclidean distance of the center position of each pair of users as a feature to infer whether they have a friend relationship. The experiment adopts svm classifier for supervised learning.

• Relational inference based on random walks. This method maps the user's mobile data into a user-location bipartite graph, walks out its neighbors in the bipartite graph for each user according to certain rules, compresses the neighbor information into vectors, and finally compares the user pairs between The similarity of feature vectors is used for friend relationship inference.

• Relational inference based on graph embeddings. This method constructs a huddle graph according to the user's mobile information, that is, two users have appeared together (appearing in the same position within a certain time interval), that is, there is an edge, and each user walks out a certain number of friends on the huuuuuuuuuu. , and encode these friend nodes into vectors, and finally infer the friend relationship by comparing the similarity of feature vectors between user pairs.

In the experiment, we use the f1-score as the evaluation index, which is not sensitive to the distribution of the class, and can more accurately describe the accuracy of the inference results even when the class distribution is unbalanced. It is defined as:

f1-score=(2×precision rate×recall rate)÷(precision rate+recall rate)

where precision precision is the number of true positives divided by the number of predicted positives, and recall (aka true positive rate) is the number of true positives divided by the number of label positives. The value of f1-score is between 0 and 1. The larger the value, the higher the accuracy of the inference.

In our experiments, we use a fully connected autoencoder network, the number of nodes in each layer of the network structure is half of the number of nodes in the previous layer, and the last layer is set by the user. On the other hand, in the structure of the decoder, we use the same inverse network structure as the encoder. The experiment adjusts the number of layers according to the size of the spatiotemporal matrix of the input encoder. We use a simple KNN and SVM as the classifier in stage 1 and binary classifier in stage 2, respectively, and use RBF as the kernel function of the support vector machine. The learning rate of all networks is set to 0.005, and the iterations are terminated when the change of the social network graph edges between two rounds of prediction is less than 1%.

1) The influence of spatial granularity σ

We experimented the accuracy of the inference results when σ was 500, 750, 1000, 1250 and 1500. As shown in Fig. 5(a), (b), in the Brightkite dataset (Fig. 5(a)), when σ increases from 500 to 1000, its f1-score increases by 4.56%, and then as σ increases Accuracy gradually decreases. A similar trend is observed for the Gawalla dataset (Fig. 5(b)), when σ = 750, the f1-score achieves the maximum value, which is due to the fact that the POIs in the Gawalla dataset are more scattered than those in the Brightkite dataset, which means that in Gawalla A grid covers a larger geographic area than Brightkite, so the optimal σ value for Gawalla dataset is smaller than Brightkite dataset.

2) The influence of time granularity τ

We experimented the accuracy of the inference results when τ was 1 day, 7 days, 14 days, 30 days and 60 days. As shown in Fig. 6(a), (b), the Brightkite dataset (Fig. 6(a)), when τ = 7 days, the F1-Score is the maximum value. Similar experimental results are also confirmed in the Gawalla dataset (Fig. 6(b)). This is due to the fact that human activity tends to exhibit periodicity, generally on a weekly basis.

3) The influence of encoding dimension d

We experimented the accuracy of the inference results when d was 16, 32, 64, 128 and 256. As shown in Fig. 7(a), (b), the accuracy changes of Brightkite dataset (Fig. 7(a)) and Gawalla dataset (Fig. 7(b)) have the same trend, which is due to the variation of the spatiotemporal relationship vector The higher the dimension, the more information it contains, which makes the attack performance better; however, the high-dimensional spatiotemporal relationship vector will also generate too much noise, which reduces the accuracy of the attack.

4) Influence of the number of iterations

Based on the above exploration of parameters, we conduct experiments on two datasets using the best value of each parameter.

Figure 8(a), (b) depict the Brightkite dataset (Fig. 8(a)) and Gawalla dataset (Fig. 8(b)), the effect of the number of iterations on our inference accuracy. It can be seen that the inference accuracy keeps improving as the number of iterations increases, in Gowalla and Brightkite the number of iterations required to satisfy the termination condition is 4 and 5, respectively.

5) The impact of the number of user check-ins

Figure 9(a), (b) plots the effect of users on the number of shared locations on the accuracy of inference results. It can be seen that with the increase of the number of shared locations of users, the accuracy of the inference results also continues to improve. The experimental results also show that our inference model in both Brightkite dataset (Fig. 9(a)) and Gawalla dataset (Fig. 9(b)) outperforms the feature extraction-based method (relation inference based on moving distance) in the f1-score. One indicator is about 10% higher.

6) The impact of users on the number of shared locations

Figure 10(a) and (b) plot the effect of the number of user check-ins on the quasi-accuracy of the inference results. Experimental results on Brightkite dataset (Fig. 10(a)) and Gawalla dataset (Fig. 10(b)) show that our method is robust to users with different check-in numbers. Obviously, the more users check in, the more accurate the modeling of user behavior patterns and the more accurate the user relationship inference will be.

Claims (2)

1. a mobile social network user relationship inference method based on spatiotemporal relationship learning, is characterized in that, comprises the following steps: Step 1. Extract the interactive behavior feature between pairs of users, and use the feature to infer whether there is a friend relationship between the two users, including the following steps: Step 101: Divide the POIs involved in the check-in of all users of the mobile data into an I×J grid according to the latitude and longitude, and at the same time divide the time into M time segments to construct an I×J×M space-time matrix STD, wherein, The time dimension is divided into equal time slices of length τ, the space dimension is evenly divided into grids of equal size, and each grid is recursively divided into four equal grids until POIs of interest points The number of in each grid is less than the threshold σ; Step 102: Project the trajectory of each user pair (u _a , u _b ) into the space-time matrix STD, and each user's check-in can be projected into a specific square. For each square, calculate: the The number of interest points _na visited by the user _{u a} in the time period; the number of interest points n _{b visited by} the user _{ub a} , u _b ) space-time matrix

The triples in the formula

Represents the statistics of the movement information of users u _a and u _b in the position grid of the i-th row and the j-th column in the m-th time period; Step 103: Encode the space-time matrix O _(a,b) of each pair of users (u _a , u _b ) into a low-dimensional vector, and use the low-dimensional vector to calculate the probability that user u _a and user _ub have a friend relationship , obtain the initial social relationship graph G=(U, E), U is the set of vertices in the graph, representing all users with mobile information; E is the set of edges, representing the friendship between two users, where, The size of the space-time matrix O _(a,b) is adjusted by the parameters σ and τ; Step 2. Extract a k-reachable subgraph for each user to describe the network structure between users. For a given initial social relationship graph G=(U, E), define the extracted user pair (u _a , u The k-reachable subgraph of _b )

The steps are as follows: Step 201, set the path length to 2, set the

initialized to an empty map; Step 202: Find all paths with lengths of 2 between (u _a , u _b ) in the initial social relationship graph G, and express all the paths found as

Then delete the social graph G and neutralize

The points and edges that appear in all are deleted, except for user u _a and user u _b itself; Step 203: Gradually increase the path length, and repeat step 202 until the path length exceeds k; Step 3. According to the initial social relationship graph G, for each pair of users (u _a , u _b ) k-reachable subgraph

Encoding, encoding paths of the same length based on the principle of accumulation, and splicing the encoding results of paths of different lengths, realizing the vectorization of the k-reachable subgraph of the user, and obtaining the comprehensive feature vector of the user pair; Step 4. Use the classifier to perform 0/1 classification according to the comprehensive feature vector of the user pair, where 1 is a friend and 0 is not a friend, and the latest predicted social graph is obtained. Step 5: Using the latest social graph, update the structural features of the user pair, and then re-predict until the prediction result does not change, that is, the final predicted social network graph is obtained. 2. a kind of mobile social network user relationship inference method based on space-time relationship learning as claimed in claim 1, is characterized in that, in step 103, the space-time matrix O _{(a, b) is} inputted into an automatic with R hidden layers. The encoder, the autoencoder encodes it into a d-dimensional vector, obtaining the reconstructed spatiotemporal matrix

To make it close to the input O _(a,b) of the encoder, the optimization objective of this training process is:

In the formula,

Represents the loss function of the autoencoder network in the hybrid network, that is, the reconstructed spatiotemporal matrix after decoding as much as possible

The same as the space-time matrix O _(a,b) before encoding, U represents all users in the training sample; The training of the auto-encoder uses supervised training to reconstruct and distinguish the encoding process, that is, add a classification network to the auto-encoder to monitor its encoding process, and the loss function of the process

for:

In the formula,

represents the prediction result, that is, the output result of the classification network; y represents the sample label; n represents the number of training samples, that is, the number of user pairs involved in the training data set. . In order to obtain a more discriminative vector representation, the following constraints are imposed on the integrated hybrid network:

In the formula,

Represents the synthetic loss function of the hybrid network. Once trained, the encoder is taken from the auto-encoder network for encoding and preliminary relation inference for any pair of user spatiotemporal relation matrices in the user set.

Images