CN115510333A - A POI Prediction Method Based on Spatio-Temporal Awareness Combining Local and Global Preferences - Google Patents

Abstract

The present invention relates to a POI prediction method based on spatio-temporal perception combined with local and global preferences, comprising the following steps: S1: select a publicly checked-in POIs data set as a training set, the training set includes user historical POI trajectory sequences; S2: construct POI Prediction model LGSA, LGSA includes local feature module, global feature module and feature fusion module; S3: use local feature module and global feature module to calculate Loc _u and Glo _u of the user's historical POI trajectory sequence; S4: set the initial weight coefficient α, use The feature fusion module combines Loc _u and Glo _u to obtain the total preference feature C _u ; S5: Use the total preference feature C _u to predict and obtain the user's next point of interest. Using the model of the present invention can further improve the accuracy of POI prediction.

Description

A POI Prediction Method Based on Spatio-Temporal Awareness Combining Local and Global Preferences

technical field

The invention relates to the field of POI prediction, in particular to a POI prediction method based on space-time perception and combining local and global preferences.

Background technique

With the growing number of personalized service platforms, location-based social networks (LBSNs) such as Foursquare and Brightkite have become a popular social platform. Users leave a large number of check-in POIs on social platforms. POIs are points of interest. It provides data support for user personalized service research. The next location prediction has always been a long-term problem for location-based personalized services. Predicting people's travel behavior and destinations through historical trajectories can provide individual users with more personalized and intelligent services. optimized location services. For example, by mining the user's historical POI trajectory and obtaining user preferences, the next location can be recommended for the user.

In the POI prediction problem, the user's POI check-in is affected by various factors such as time, geography, POI category, and long-term preference; among them, geographical distance will affect the range of users' continuous check-in locations, because users are more inclined to choose interest A multi-functional area with a short geographical distance between points; in the time dimension, users will check in points of interest with different functions according to the time point. For example, the point of interest checked in at 12:00 noon is generally a restaurant, and from 14:00 to 17:00 in the afternoon The POI categories of 00 check-in are leisure places such as coffee shops and movie theaters. In addition, the user's dynamic preference over time is also crucial, and the trajectory sequence can reflect the dependence between the user's dynamic preference over time and the check-in location.

Considering the above factors, most studies in recent years use recurrent neural networks (RNNs) or its variants to capture users’ long-term and short-term preferences, or use spatio-temporal attention networks to aggregate all visited locations in users’ historical POI trajectories, in order to capture users’ long-term and short-term preferences, Ma et al. proposed a hierarchical gating network combined with Bayesian personalized ranking; subsequently, Ma et al. further proposed a memory-enhanced graph neural network based on the original problem to capture users' long-term and short-run preferences. However, both of these methods rely on shallow methods, which cannot effectively capture the dynamic interests of users over time, and ignore the geographical factors in the trajectory sequence. It is targeted to reduce the prediction effect; while ST-RNN and STAN both consider the spatio-temporal characteristics of the user's check-in POI, but ignore the user's trajectory sequence and personalized spatio-temporal area. This way of predicting the location will reduce the same spatio-temporal area. Strong correlations between locations.

Contents of the invention

In view of the above-mentioned problems existing in the prior art, the technical problem to be solved by the present invention is: how to improve the prediction accuracy of POI.

In order to solve the above technical problems, the present invention adopts the following technical solution: a POI prediction method based on space-time perception and combining local and global preferences, including the following steps:

表示用户在t_s时刻所处的兴趣点；S100: Select the public check-in POIs data set as the training set O, the training set O includes the user's historical POI trajectory sequence, and the historical POI trajectory sequence of each user is expressed as

Indicates the point of interest where the user is at time t _s ;

The user's historical POI trajectory is arranged in chronological order, and the user's historical POI trajectory sequence includes a user label, a time label, a location label, and the latitude and longitude corresponding to the location label;

S200: Construct a POI prediction model LGSA, where the LGSA includes a local feature module, a global feature module, and a feature fusion module;

S300: Randomly select a user's historical POI trajectory sequence from the training set O, and calculate the local feature vector representation and global feature vector representation of the user's historical POI trajectory sequence:

S310: Taking the user's historical POI track sequence as input, and using the local feature module to calculate the local feature vector representation Loc _u of the user in the space-time region, the expression is as follows:

表示细粒度子序列的子序列特征；Among them, T represents the matrix transpose, avg( ) represents the average function, ranh( ) represents the activation function, At ^l represents the attention score matrix,

Represents subsequence features of fine-grained subsequences;

S320: Taking the user's historical POI trajectory sequence as input, and using the global feature module to calculate the user's global feature vector representation Glo _u in the time unit of weeks, the expression is as follows:

Glo _u =LN(A ^h +Dropout(PFFN(A ^h ))); (2)

Among them, A ^h represents the output of the feature after multi-head attention, LN( ) represents the layer normalization in the neural network, Dropout( ) represents the method used to prevent the model from overfitting, PFFN( ) represents the forward feed-forward network;

S400: Set the initial weight coefficient α, use the feature fusion module to combine Loc _u and Glo _u , and fuse the local features and global features through weighted summation to obtain the user's total preference feature C _u , the specific expression is as follows:

C _u =αLoc _u +(1-α)Clo _u ; (3)

Among them, α is the weight coefficient, α∈[0,1];

S500: Using the total preference feature C _u to calculate the predicted POI of the user, the specific steps are as follows:

S510: Calculate the predicted POI representation vector P _u of the user, the expression is as follows:

表示轨迹序列中的地点表示特征，d表示特征维度，t表示时间，Co表示用户时空区域中地点间的特征关系，Co的计算表达式如下：in,

Indicates that the location in the trajectory sequence represents the feature, d represents the feature dimension, t represents the time, Co represents the feature relationship between the locations in the user's space-time area, and the calculation expression of Co is as follows:

表示时空区域的子序列表示向量，W_r是可学习的参数；in,

The subsequence representation vector representing the space-time region, W _r is a learnable parameter;

S520: Map Pu to the location label of the _POI through index mapping to obtain the predicted POI of the user; the index mapping is that a POI location label corresponds to a POI representation vector, and the two are in a one-to-one correspondence;

S600: Using the user's predicted POI representation vector to calculate the objective function K of the LGSA model, the specific expression is as follows:

K＝argmin∑ _{(u,pos,neg)∈O} -log(σ(P _u,pos -P _u,neg )); (6)

Among them, P _u,pos represents the distance between the real POI and the predicted POI representation vector, P _u,neg represents the distance between the POI in the non-current user trajectory sequence and the predicted POI representation vector, u represents the user label, pos represents the real POI label, neg Represents the POI label in the non-current user trajectory sequence, and σ represents the sigmoid function;

S700: Using the objective function K as a loss function to train the LGSA model, and using the gradient descent method to reversely update the parameters of the LGSA model;

S800: Traverse all the historical POI trajectory sequences of users in the training set, repeat steps S300-S700 to train the model, preset the maximum number of iterations for training, stop training when the training reaches the maximum number of iterations, and obtain a trained LGSA model;

S900: Taking the historical POI trajectory sequence of the user to be predicted as the input of the trained LGSA, and outputting a prediction result of the next POI of the user to be predicted.

As a preference, the specific steps of calculating the local feature vector representation Loc _u of each user in the spatio-temporal region in the S310 are as follows:

S311: Randomly select a user historical POI trajectory sequence, and perform time interval processing on the user historical POI trajectory sequence, the specific expression is as follows:

表示用户在访问地点i和地点j之间的时间间隔，t_i表示在地点i的时间戳，t_j表示在地点j的时间戳；in,

Indicates the time interval between the user visiting location i and location j, t _i indicates the timestamp at location i, and t _j indicates the timestamp at location j;

Perform geographical distance processing on the user's historical POI trajectory sequence, the specific expression is as follows:

表示访问地点i和地点j之间的地理距离，r表示半径，lon_i和lat_i表示地点i的GPS_i的经纬度,lon_j和lat_j表示地点j的GPS_j的经纬度，Haversine(·)表示地理距离函数；in,

Indicates the geographical distance between visiting location i and location j, r indicates the radius, lon _i and lat _i indicate the longitude and latitude of GPS _i of location i, lon _j and lat _j indicate the longitude and latitude of GPS _j of location j, Haversine( ) means geographical distance function;

S312: Divide the user's historical POI trajectory sequence according to the time interval and geographical distance to obtain the space-time region set Reg _u , the specific expression is as follows:

Reg _u = {reg ₁ ,reg ₂ ,...,reg _x }; (9)

Among them, reg _x represents the xth space-time region, and x represents the number of space-time regions;

S313: Use the sliding window w to divide each spatio-temporal area in the spatio-temporal area set into fine-grained subsequences, the specific expression is as follows:

Among them, w represents the size of the sliding window;

计算表达式如下：S314: Calculating subsequence features of fine-grained subsequences

The calculation expression is as follows:

Among them, W ₁ represents the learnable parameters, M _a ∈ R ^{w × d} represents the adaptive adjacency matrix, d represents the feature dimension, tanh represents the activation function, and the subsequence embedding in the region is expressed as

的注意力分数矩阵，计算表达式如下：S315: Calculate

The attention score matrix of , the calculation expression is as follows:

Among them, At ^l represents the attention score matrix, W ₂ , W ₃ , b ₁ and b ₂ all represent learnable parameters, and Softmax represents the activation function;

计算得到该用户在时空区域上的局部特征向量表示Loc_u；S316: using At ^l and

Calculate and obtain the local feature vector representation Loc _u of the user in the spatio-temporal region;

S317: Traverse all user historical POI trajectory sequences in the training set, and calculate the local feature vector representation of each user in the spatio-temporal region.

This method can effectively mine the characteristics of the location itself in the subsequence of the spatiotemporal region and the information of its surrounding associated locations, and obtain the user's interest preference within the scope of the local specific spatiotemporal region.

As a preference, the specific steps of calculating the user's global feature vector representation Glo _u in the spatio-temporal region in the S320 are as follows:

S321: Randomly select a user's historical POI trajectory sequence, and fuse the time information in the user's historical POI trajectory sequence, and the calculation expression is as follows:

表示完成时间信息融合后的用户历史POI轨迹序列，W₄表示可学习参数，

表示轨迹序列与特殊时间周期的拼接向量；in,

Indicates the user's historical POI trajectory sequence after time information fusion is completed, W ₄ indicates the learnable parameters,

A concatenated vector representing a sequence of trajectories and a special time period;

的非侵入式的自注意力

计算表达式如下：S322: Calculate

non-intrusive self-attention

The calculation expression is as follows:

和S_u映射得到的可学习矩阵，K^T表示K的转置矩阵，σ表示可学习的参数；Among them, N means that there are N layers of multi-head attention network layers, Attention(·) means the attention function, Q, K, and V are respectively from

The learnable matrix obtained by mapping with S _u , K ^T represents the transpose matrix of K, and σ represents the learnable parameters;

S323: Calculate N-layer multi-head attention, the calculation expression is as follows:

是第y层多头注意力网络层，A^h是多头注意力层的输出，GWLU表示高斯误差线性单元，W₅、W₆、b₅、和b₆表示可学习参数；in,

is the multi-head attention network layer of the y-th layer, A ^h is the output of the multi-head attention layer, GWLU represents a Gaussian error linear unit, and W ₅ , W ₆ , b ₅ , and b ₆ represent learnable parameters;

S324: Perform layer normalization processing and dropout function processing on the output of each sub-layer, and obtain the global feature vector representation Glo _u of the user in the spatio-temporal region;

S325: Traverse all user historical POI trajectory sequences in the training set, and calculate the global feature vector representation of each user in the spatio-temporal region.

This method focuses on the user's dynamic sequence, that is, the global feature, which can effectively capture the dynamic preference and long-term semantics of user behavior over time, and mine relevant locations in the trajectory sequence.

As a preference, the specific steps for performing time information fusion in S321 are as follows:

用户历史POI轨迹序列中的特殊时间模式，

其中，

表示轨迹序列中的地点表示向量，t_i表示特殊时间周期的表示向量，u表示用户，week表示以周为单位的特殊时间周期；S321-1: The historical trajectory sequence of each user is

Special temporal patterns in the sequence of user historical POI trajectories,

in,

Represents the location representation vector in the trajectory sequence, t _i represents the representation vector of a special time period, u represents the user, and week represents the special time period in weeks;

S321-2: According to the time unit conversion, convert the POI check-in time into a special time period in weeks, and obtain the embedded matrix of the special time period

Calculate the embedding matrix E(S _u ) of the trajectory sequence according to the word2vec word embedding method;

和E(S_u)在特征维度上进行拼接，得到拼接矩阵

具体表达式如下：S321-3: will

and E(S _u ) are spliced on the feature dimension to get the splicing matrix

The specific expression is as follows:

Among them, con( ) represents the splicing function;

进行处理，得到完成时间信息融合后的用户历史POI轨迹序列

S321-4: Use the activation function pair

Process to obtain the user's historical POI trajectory sequence after time information fusion

Local features include learning the dependencies between POIs of the user in the spatio-temporal region, and global features can obtain the user's dynamic preference for a special time period, so combining local features and global features can be considered on the basis of trajectory sequences. Capturing user behavior preferences at the spatio-temporal level.

Compared with the prior art, the present invention has at least the following advantages:

1. The present invention discloses a POI prediction method based on spatiotemporal perception combined with local and global preferences for next POI prediction. According to the geographic distance and time interval, each user's trajectory sequence is divided into personalized spatio-temporal regions, and the user's dependencies among POIs in the check-in region are learned from the local view. In addition, time information fusion is used to mine the user's dynamic preferences over time with fine-grained time as a week, because the weekly change period feature can well reflect the specific activities of the user's historical track in each week; using a non-invasive way To fuse the user's trajectory sequence and the time period of the sequence, and mine the user's dynamic preference for the time period from the global view. Finally, the local feature and the global feature are fused to obtain the total preference feature, and finally the user's next POI prediction point is obtained by using the total preference feature.

2. The present invention proposes an LGSA model to learn user preferences from local (i.e., personalized spatio-temporal regions) and global (i.e., trajectory sequences with time periods) views based on user historical POI trajectory sequences.

3. In order to effectively learn the local dependencies in the spatio-temporal region, the present invention proposes a personalized spatio-temporal perception region, divides the personalized spatio-temporal region for each user according to the geographical distance and time interval in the trajectory sequence, and learns from the local view The user checks in the dependency between POIs in the spatiotemporal region, which can reduce the correlation of POIs in different spatiotemporal regions in the trajectory sequence, thereby improving the prediction results.

4. In order to fully integrate the trajectory sequence and the personalized time period, the present invention uses a non-intrusive method to fuse the user's trajectory sequence and the time period of the sequence, and mine the user's dynamic preferences and the POI timing of the trajectory sequence in the time period from the global view , this method will neither cover the trajectory information, but also can extract the relationship between the trajectory sequence and the weekly change period very well.

Description of drawings

Fig. 1 is a frame diagram of the LGSA structure including local features and global features in the present invention.

FIG. 2 is a schematic diagram of dividing space-time regions using time intervals and geographic distances in the present invention.

Fig. 3 is the performance comparison on the NYC dataset in the experiment of the present invention.

Fig. 4 is the performance comparison on the TKY data set in the experiment of the present invention.

Fig. 5 is the performance comparison on the Brightkite data set in the experiment of the present invention.

Fig. 6 is the batch processing parameter experiment in the experiment of the present invention.

Fig. 7 is an experiment of spatiotemporal threshold parameters in the experiment of the present invention.

detailed description

The present invention will be described in further detail below.

The present invention discloses a combination of local and global preferences based on spatio-temporal perception for next POI prediction. A local preference module is proposed according to the dependence between POIs in the personalized spatio-temporal region; then, the global preference is modeled to mine the user's dynamic preference in a specific time period; modeling the preference aggregation can effectively fuse local features and Global features to complete POI prediction.

See Figure 1-Figure 2, a POI prediction method based on space-time perception and combining local and global preferences, including the following steps:

表示用户在t_s时刻所处的兴趣点；S100: Select the public check-in POIs data set as the training set O, the training set O includes the user's historical POI trajectory sequence, and the historical POI trajectory sequence of each user is expressed as

Indicates the point of interest where the user is at time t _s ;

The user's historical POI trajectory is arranged in chronological order, and the user's historical POI trajectory sequence includes a user label, a time label, a location label, and the latitude and longitude corresponding to the location label;

S200: Construct a POI prediction model LGSA, where the LGSA includes a local feature module, a global feature module, and a feature fusion module;

S300: Randomly select a user's historical POI trajectory sequence from the training set O, and calculate the local feature vector representation and global feature vector representation of the user's historical POI trajectory sequence:

S310: Taking the user's historical POI track sequence as input, and using the local feature module to calculate the local feature vector representation Loc _u of the user in the space-time region, the expression is as follows:

表示细粒度子序列的子序列特征；Among them, T represents the matrix transpose, avg( ) represents the average function, tanh( ) represents the activation function, At ^l represents the attention score matrix,

Represents subsequence features of fine-grained subsequences;

The specific steps of calculating the local feature vector representation Loc _u of each user in the space-time region in the S310 are as follows:

S311: Randomly select a user historical POI trajectory sequence, and perform time interval processing on the user historical POI trajectory sequence, the specific expression is as follows:

表示用户在访问地点i和地点j之间的时间间隔，t_i表示在地点i的时间戳，t_j表示在地点j的时间戳；in,

Indicates the time interval between the user visiting location i and location j, t _i indicates the timestamp at location i, and t _j indicates the timestamp at location j;

Perform geographical distance processing on the user's historical POI trajectory sequence, the specific expression is as follows:

表示访问地点i和地点j之间的地理距离，r表示半径，lon_i和lat_i表示地点i的GPS_i的经纬度,lon_j和lat_j表示地点j的GPS_j的经纬度，Haversine(·)表示地理距离函数；in,

Indicates the geographical distance between visiting location i and location j, r indicates the radius, lon _i and lat _i indicate the longitude and latitude of GPS _i of location i, lon _j and lat _j indicate the longitude and latitude of GPS _j of location j, Haversine( ) means geographical distance function;

S312: Divide the user's historical POI trajectory sequence according to the time interval and geographical distance to obtain the space-time region set Reg _u , the specific expression is as follows:

Reg _u = {reg ₁ ,reg ₂ ,...,reg _x }; (9)

Among them, reg _x represents the xth space-time region, and x represents the number of space-time regions;

The time interval between locations in the trajectory sequence is in minutes, and both the time interval threshold θ _t and the geographic distance threshold θ _g are used to divide the space-time region; specifically, the time interval and geographic distance of the trajectory sequence are calculated and different The median value of is respectively set as the threshold for dividing spatio-temporal regions; therefore, each sequence gets a different number of spatio-temporal regions, which are considered as local features in the present invention.

S313: Use the sliding window w to divide each spatiotemporal region in the spatiotemporal region set into fine-grained subsequences. The sliding window is an existing known experimental strategy, and the specific expression is as follows:

Among them, w represents the size of the sliding window; this experimental strategy can help the model pay more attention to the connection between adjacent check-in locations in the spatio-temporal region, while ignoring irrelevant locations with little influence; for each user, there are For the corresponding spatiotemporal region set, a sliding window is used to generate a subsequence set for each spatiotemporal region of the user, and the number of subsequence sets is determined by the size of the sliding window.

计算表达式如下：S314: Calculating subsequence features of fine-grained subsequences

The calculation expression is as follows:

由于子序列不是GNN建模的固有图，因此需要构建一个图来捕获位置之间的连接，在子序列中的位置之间添加边，计算所有用户中提取的项目对的边数并建立自适应邻接矩阵，根据边数对邻接矩阵进行归一化，这个自适应邻接矩阵可以自动学习区域中子序列地点间的依赖关系。Among them, W ₁ represents the learnable parameters, M _a ∈ R ^{w × d} represents the adaptive adjacency matrix, d represents the feature dimension, tanh represents the activation function, which is used to make the value of the matrix range from -1 to 1; The subsequence embedding within a region is expressed as

Since subsequences are not inherent graphs modeled by GNNs, a graph needs to be built to capture connections between locations, add edges between locations in a subsequence, count the number of edges for item pairs extracted in all users, and build an adaptive Adjacency matrix, which normalizes the adjacency matrix according to the number of edges, this adaptive adjacency matrix can automatically learn the dependencies between subsequence locations in the region.

的注意力分数矩阵，计算表达式如下：S315: Calculate

The attention score matrix of , the calculation expression is as follows:

Among them, At ^l represents the attention score matrix, W ₂ , W ₃ , b ₁ and b ₂ all represent learnable parameters, and Softmax represents the activation function; here, two layers of GNN are used to aggregate information, which is for good mining The subsequence in the spatio-temporal region has its own characteristics and the information of its surrounding locations; the attention score matrix can assign attention weights to each location, and extract the locations that the user pays more attention to in the local area.

计算得到该用户在时空区域上的局部特征向量表示Loc_u；S316: using At ^l and

Calculate and obtain the local feature vector representation Loc _u of the user in the spatio-temporal region;

S317: Traverse all user historical POI trajectory sequences in the training set, and calculate the local feature vector representation of each user in the spatio-temporal region.

S320: Taking the user's historical POI trajectory sequence as input, and using the global feature module to calculate the user's global feature vector representation Glo _u in the time unit of weeks, the expression is as follows:

Glo _u =LN(A ^h +Dropout(PFFN(A ^h ))); (2)

Among them, A ^h represents the output of the feature after multi-head attention, LN( ) represents the layer normalization in the neural network, Dropout( ) represents the method used to prevent the model from overfitting, PFFN( ) represents the forward feed-forward network;

The specific steps of calculating the user's global feature vector representation Glo _u in the spatio-temporal region in the S320 are as follows:

S321: Randomly select a user's historical POI trajectory sequence, and fuse the time information in the user's historical POI trajectory sequence, and the calculation expression is as follows:

表示完成时间信息融合后的用户历史POI轨迹序列，W₄表示可学习参数，

表示轨迹序列与特殊时间周期的拼接向量；in,

Indicates the user's historical POI trajectory sequence after time information fusion is completed, W ₄ indicates the learnable parameters,

A concatenated vector representing a sequence of trajectories and a special time period;

The specific steps for performing time information fusion in S321 are as follows:

用户历史POI轨迹序列中的特殊时间模式，

其中，

表示轨迹序列中的地点表示向量，t_i表示特殊时间周期的表示向量，u表示用户，week表示以周为单位的特殊时间周期；S321-1: The historical trajectory sequence of each user is

Special temporal patterns in the sequence of user historical POI trajectories,

in,

Represents the location representation vector in the trajectory sequence, t _i represents the representation vector of a special time period, u represents the user, and week represents the special time period in weeks;

S321-2: According to the time unit conversion, convert the POI check-in time into a special time period in weeks, and obtain the embedded matrix of the special time period

Calculate the embedding matrix E (S _u ) of the trajectory sequence according to the word2vec word embedding method, and the word2vec word embedding method is prior art;

和E(S_u)在特征维度上进行拼接，得到拼接矩阵

具体表达式如下：S321-3: will

and E(S _u ) are spliced on the feature dimension to get the splicing matrix

The specific expression is as follows:

Among them, con( ) represents the splicing function;

进行处理，得到完成时间信息融合后的用户历史POI轨迹序列

S321-4: Use the activation function pair

Process to obtain the user's historical POI trajectory sequence after time information fusion

的非侵入式的自注意力

计算表达式如下：S322: Calculate

non-intrusive self-attention

The calculation expression is as follows:

和S_u映射得到的可学习矩阵，K^T表示K的转置矩阵，σ表示可学习的参数；Among them, N means that there are N layers of multi-head attention network layers, Attention(·) means the attention function, Q, K, and V are respectively from

The learnable matrix obtained by mapping with S _u , K ^T represents the transpose matrix of K, and σ represents the learnable parameters;

S323: Calculate N-layer multi-head attention, the calculation expression is as follows:

是第y层多头注意力网络层，A^h是多头注意力层的输出，GELU表示高斯误差线性单元，W₅、W₆、b₅、和b₆表示可学习参数；in,

is the multi-head attention network layer of the y-th layer, A ^h is the output of the multi-head attention layer, GELU represents the Gaussian error linear unit, W ₅ , W ₆ , b ₅ , and b ₆ represent the learnable parameters;

S324: Perform layer normalization processing and dropout function processing on the output of each sub-layer, and obtain the global feature vector representation Glo _u of the user in the spatio-temporal region;

S325: Traverse all user historical POI trajectory sequences in the training set, and calculate the global feature vector representation of each user in the spatio-temporal region.

S400: Set the initial weight coefficient α, use the feature fusion module to combine Loc _u and Glo _u , that is, aggregate the user's interests, and fuse the local features with the global features through weighted summation to obtain the user's total preference feature C _u , the specific expression is as follows:

C _u =αLoc _u +(1-α)Glo _u ; (3)

Among them, α is the representation weight coefficient, α ∈ [0,1]; combining local and global features can more effectively capture the user's behavior preference at the spatio-temporal level, instead of only considering the trajectory sequence; combining these two feature representations It can complete the position prediction work very well. The combination here is obtained by fusing local features with global features through weighted summation. The value of α is from {0.0,0.1,0.2,...,0.8,0.9,1.0 } in order to conduct parameter experiments to calculate the _Cu value, and the corresponding α value when the _Cu value is the largest is taken as the optimal α value.

S500: Using the total preference feature C _u to calculate the predicted POI of the user, the specific steps are as follows:

S510: Calculate the predicted POI representation vector P _u of the user, the expression is as follows:

表示轨迹序列中的地点表示特征，d表示特征维度，t表示时间，Co表示用户时空区域中地点间的特征关系，Co的计算表达式如下：in,

Indicates that the location in the trajectory sequence represents the feature, d represents the feature dimension, t represents the time, Co represents the feature relationship between the locations in the user's space-time area, and the calculation expression of Co is as follows:

表示时空区域的子序列表示向量，W_r是可学习的参数；in,

The subsequence representation vector representing the space-time region, W _r is a learnable parameter;

S520: Map Pu to the location label of the _POI through index mapping to obtain the predicted POI of the user; the index mapping is that a POI location label corresponds to a POI representation vector, and the two are in a one-to-one correspondence;

S600: Using the user's predicted POI representation vector to calculate the objective function K of the LGSA model, the specific expression is as follows:

K＝argmin∑ _{(u,pos,neg)∈O} -log(σ(P _u,pos -P _u,neg )); (6)

Among them, P _u,pos represents the distance between the real POI and the predicted POI representation vector, P _u,neg represents the distance between the POI in the non-current user trajectory sequence and the predicted POI representation vector, u represents the user label, pos represents the real POI label, neg Represents the POI label in the non-current user trajectory sequence, σ represents the sigmoid function, and the objective function is to express the distance between the predicted POI vector and the real value vector, and the distance between the negative sampling vector and the negative sampling vector. The POI value of the non-current user, that is, the POI that the current user has not checked in; the objective function argmin uses the Bayesian personalized sorting objective function;

S700: Using the objective function K as a loss function to train the LGSA model, and using the gradient descent method to reversely update the parameters of the LGSA model;

S800: Traverse all the historical POI trajectory sequences of users in the training set, repeat steps S300-S700 to train the model, preset the maximum number of iterations for training, stop training when the training reaches the maximum number of iterations, and obtain a trained LGSA model;

S900: Taking the historical POI trajectory sequence of the user to be predicted as the input of the trained LGSA, and outputting a prediction result of the next POI of the user to be predicted.

Experimental verification

1. Dataset

The experimental evaluation of the present invention is carried out on the two public LBSN data sets listed below, which belong to known public data sets and have relatively high credibility. The details of the dataset are shown in Table 1.

Foursquare Dataset: Foursquare is a location-based social networking site where users share their locations by checking in. The dataset includes long-term (approximately 10 months) check-in data of New York City and Tokyo collected from Foursquare from April 12, 2012 to February 16, 2013. Users who visited less than 10 POIs and POIs with fewer than 10 visitors, after preprocessing, the New York City dataset contains 134,691 check-ins and 4,513 POIs from 1,078 users, while the Tokyo dataset contains 401,857 check-ins and 6,952 POIs from 2,266 users.

Brightkite dataset: Brightkite used to be a location-based social network service provider, where users shared their locations by checking in. The data set contains the check-in records generated on Brightkite from April 2008 to October 2010. In this experiment, select the top 10,000 users, users who have visited less than 10 POIs or more than 1,000 POIs, and the number of visitors POIs with less than 10 people are removed, and after preprocessing, the Brightkite dataset contains 884,373 check-ins and 19,034 POIs from 5153 users.

In this experiment, the first 80% of each user's sequence is used as the training set, and the latter 20% is used as the test set. The user's historical trajectories are sorted in chronological order. See Table 1 for more detailed information about the dataset.

Table 1 Dataset Statistics

data set NYC TKY Bright kite user 1,078 2,266 5,153 POIs 4,513 6,952 19,034 Number of check-in POIs 134,691 401,857 884,373 The average number of check-ins per user 124.95 177.34 171.62 The average number of check-ins per POI 29.85 57.80 46.46 Sparsity 97.235% 97.451% 99.098%

2. Baseline

In this experiment, we mainly focus on the characteristics of the trajectory sequence and the spatiotemporal influence between POIs to predict the next POI. Therefore, the LGSA proposed by the present invention is compared with the following baseline model.

TOP: This method records the popularity of POIs in the dataset and recommends the most popular POIs to users.

ST-RNN: This method establishes a specific time transfer matrix at different time intervals, and a specific distance transfer matrix at different geographical distances for the prediction of the next location.

HGN: This method combines hierarchical gating networks with Bayesian personalized ranking to leverage users' long-term and short-term interests for sequential recommendations.

MA-GNN: This method uses a memory-augmented graph neural network to capture users' long-term and short-term interests for sequential recommendation.

STAN: This method uses a spatio-temporal attention network to aggregate all relevant visits from a user's trajectory and recall the most plausible candidates from weighted representations for place recommendation.

3. Evaluation indicators

According to the previous settings, this experiment uses four evaluation indicators for the next POI prediction task. These evaluation indicators are Precision@10, Recall@10, Normalized Discounted Cumulative Gain (NDCG@10) and Average Precision (MAP@10) , where 10 is the number of POIs in the ranking list; when the correct POI is in the top 10 POIs, the scores of Precision@10 and Recall@10 are high, and NDCG@10 is used as the evaluation index of the ranking result, which is used to evaluate the ranking Accuracy, MAP scores the quality of the entire ranking set, and the larger the value of these four evaluation indicators, the better the performance.

4. Experimental details

For the spatio-temporal threshold, the experiments were carried out at the threshold value of one-quarter digit, three-quarter digit, median and average value of the three-digit number, and the median of the optimal result was taken as the threshold value. Geographic distance and time interval exceed the median at the same time, and a region is divided. The size of the sliding window in the local area is selected from {4, 6, 8, 10} to conduct experiments respectively. When L=6, the results on the three data sets are the best.

In the LGSA model, the Adam optimizer is used to train the model, the learning rate is 0.001, and the regularization parameter is set to 0.001; the hyperparameters are adjusted by grid search on the validation set; the embedding size is set to 100. The batch size of NYC and Brightkite is set to 2048, and the batch size of TKY is set to 4096. For the spatio-temporal threshold, use the average value of the time interval and space interval, 25%, 50% and 75% as the threshold of the experiment, and take the best 50% as the threshold; when the geographical distance and time interval exceed the user's historical POI trajectory Spatio-temporal regions are divided when the corresponding median in the series. The sliding window size w in local features is selected from {4, 6, 8, 10} for experiments. When w is 6, the results on the three data sets are the best, the length of the trajectory sequence is set to 100, and the local weight coefficient α is selected from {0.2, 0.4, 0.6, 0.8, 1.0} for experiments, When α is 0.4, the result is optimal.

5. Results

The results are shown in Figures 3, 4, and 5. In the baseline, top has the lowest effect on the four evaluation indicators, indicating that it is not enough to simply obtain the popularity of POI and make predictions; both HGN and MA-GNN It focuses on the long-term and short-term preferences of users, and the effect is obviously better than TOP, but it lacks the consideration of spatio-temporal features, so the effect is lower than LGSP; ST-RNN and STAN both mine the spatio-temporal features of the trajectory to obtain the context features of the trajectory, but ignore The sequence characteristics of the trajectory and the periodicity of the sequence, so the result is lower than LGSP.

The LGSP model is significantly better than the baseline in the index of NDCG@10, which shows that the model has a good ranking effect on the prediction results; secondly, it is in the index of recall@10, which shows that the model can recall the item set to be predicted well; But in terms of the precision@10 indicator, the LGSP model has little advantage, only a little better than the baseline. On the TKY dataset, LGSP is significantly better than baseline in terms of pre@10, NDCG@10, and MAP evaluation indicators, which shows that the LGSP model has improved prediction accuracy and sorting effect, but has little improvement in recall; in Brightkite data The effect on the set is not much improved, and it is only a little better than the baseline.

6. Parameter experiment

In order to study the impact of different settings on key hyperparameters, this experiment evaluates the LGSP model by changing the batch size of related itemsets and the size of the spatiotemporal threshold:

First, change the batch size. Doubled from 256 to 4096, as shown in Figure 6. It shows that on the New York City and Brightkite datasets, higher performance is obtained when the batch size is 2048, while on the TKY dataset, the batch size is 4096. As the batch size value increases, the performance of the model gradually increases, because when the batch size value is too small, the training data is difficult to converge, resulting in underfitting.

Second, aim at the spatio-temporal threshold. In this experiment, the thresholds of the mean, lower quartile, median and upper quartile were tested. The results shown in Figure 7 show that on the three data sets, when the spatiotemporal threshold is the median, A higher performance is obtained; this shows that the median-based spatio-temporal threshold can well divide the user's personalized spatio-temporal region; the average ignores the distribution of time intervals and geographical distances, and the existence of outliers tends to expand or shrink the results ; neither the lower quartile nor the upper quartile can accurately capture the spatio-temporal region of user activity.

7. Ablation Experiment

The ablation research of the LGSA model was carried out on the NYC, TKY and Brightkite datasets, including four aspects: only considering local features (LF), only considering global features (GF), combining local and global features (LF+GF), and Fusion time segment of local and global features (LF+GF+time). The experimental results are shown in Table 2.

Table 2 Results of ablation experiments on NYC, TKY and Brightkite datasets

It can be seen from the experimental results that when there are only global features, the performance on the TKY dataset is the worst, which shows that the spatio-temporal region of users in the TKY dataset is a very important feature, and users on this platform tend to be in the spatio-temporal region Activities; from the three data sets, it can be seen that predicting the next POI should not only consider the sequentiality of the user's historical POI trajectory globally, that is, the dynamic preference that changes over time, but also consider the user's personalized spatio-temporal region locally and mine the gap between POIs. Dependence; From the experimental results, it can be seen that the performance of the time period feature does not improve the performance of the model very much, because the time period is considered from a global perspective, and the proportion in the model structure is not large.

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements without departing from the spirit and scope of the technical solution of the present invention shall be covered by the claims of the present invention.

Claims (4)

1. A POI prediction method based on space-time perception and in combination with local and global preferences, characterized in that: comprising the steps: S100: Select the public check-in POIs data set as the training set O, the training set O includes the user's historical POI trajectory sequence, and the historical POI trajectory sequence of each user is expressed as

Indicates the point of interest where the user is at time t _s ; The user's historical POI trajectory is arranged in chronological order, and the user's historical POI trajectory sequence includes a user label, a time label, a location label, and the latitude and longitude corresponding to the location label; S200: Construct a POI prediction model LGSA, where the LGSA includes a local feature module, a global feature module, and a feature fusion module; S300: Randomly select a user's historical POI trajectory sequence from the training set O, and calculate the local feature vector representation and global feature vector representation of the user's historical POI trajectory sequence: S310: Taking the user's historical POI track sequence as input, and using the local feature module to calculate the local feature vector representation Loc _u of the user in the space-time region, the expression is as follows:

Among them, T represents the matrix transpose, avg( ) represents the average function, tanh( ) represents the activation function, At ^l represents the attention score matrix,

Represents subsequence features of fine-grained subsequences; S320: Taking the user's historical POI trajectory sequence as input, and using the global feature module to calculate the user's global feature vector representation Glo _u in the time unit of weeks, the expression is as follows: Glo _u =LN(A ^h +Dorpout(PFFN(A ^h ))); (2) Among them, A ^h represents the output of the feature after multi-head attention, LN( ) represents the layer normalization in the neural network, Dropout( ) represents the method used to prevent the model from overfitting, PFFN( ) represents the forward feed-forward network; S400: Set the initial weight coefficient α, use the feature fusion module to combine Loc _u and Glo _u , and fuse the local features and global features through weighted summation to obtain the user's total preference feature C _u , the specific expression is as follows: C _u =αLoc _u +(1-α)Glo _u ; (3) Among them, α is the weight coefficient, α∈[0,1]; S500: Using the total preference feature C _u to calculate the predicted POI of the user, the specific steps are as follows: S510: Calculate the predicted POI representation vector P _u of the user, the expression is as follows:

in,

Indicates that the location in the trajectory sequence represents the feature, d represents the feature dimension, t represents the time, Co represents the feature relationship between the locations in the user's space-time area, and the calculation expression of Co is as follows:

in,

The subsequence representation vector representing the space-time region, W _r is a learnable parameter; S520: Map Pu to the location label of the _POI through index mapping to obtain the predicted POI of the user; the index mapping is that a POI location label corresponds to a POI representation vector, and the two are in a one-to-one correspondence; S600: Using the user's predicted POI representation vector to calculate the objective function K of the LGSA model, the specific expression is as follows: K＝argmin∑ _{(u,pos,neg)∈O} -log(σ(P _u,pos -P _u,neg )); (6) Among them, P _u,pos represents the distance between the real POI and the predicted POI representation vector, P _u,neg represents the distance between the POI in the non-current user trajectory sequence and the predicted POI representation vector, u represents the user label, pos represents the real POI label, neg Represents the POI label in the non-current user trajectory sequence, and σ represents the sigmoid function; S700: Using the objective function K as a loss function to train the LGSA model, and using the gradient descent method to reversely update the parameters of the LGSA model; S800: Traverse all the historical POI trajectory sequences of users in the training set, repeat steps S300-S700 to train the model, preset the maximum number of iterations for training, stop training when the training reaches the maximum number of iterations, and obtain a trained LGSA model; S900: Taking the historical POI trajectory sequence of the user to be predicted as the input of the trained LGSA, and outputting a prediction result of the next POI of the user to be predicted. 2. A kind of POI prediction method based on spatio-temporal perception and combining local and global preferences as claimed in claim 1, characterized in that: in said S310, the local eigenvector of each user on the spatio-temporal region is calculated to represent the specific value of Loc _u Proceed as follows: S311: Randomly select a user historical POI trajectory sequence, and perform time interval processing on the user historical POI trajectory sequence, the specific expression is as follows:

in,

Indicates the time interval between the user visiting location i and location j, t _i indicates the timestamp at location i, and t _j indicates the timestamp at location j; Perform geographical distance processing on the user's historical POI trajectory sequence, the specific expression is as follows:

in,

Indicates the geographical distance between visiting location i and location j, r indicates the radius, lon _i and lat _i indicate the longitude and latitude of GPS _i of location i, lon _j and lat _j indicate the longitude and latitude of GPS _j of location j, Haversine( ) means geographical distance function; S312: Divide the user's historical POI trajectory sequence according to the time interval and geographical distance to obtain the space-time region set Reg _u , the specific expression is as follows: Reg _u = {reg ₁ ,reg ₂ ,...,reg _x }; (9) Among them, reg _x represents the xth space-time region, and x represents the number of space-time regions; S313: Use the sliding window w to divide each spatio-temporal area in the spatio-temporal area set into fine-grained subsequences, the specific expression is as follows:

Among them, w represents the size of the sliding window; S314: Calculating subsequence features of fine-grained subsequences

The calculation expression is as follows:

Among them, W ₁ represents the learnable parameters, M _a ∈ R ^{w × d} represents the adaptive adjacency matrix, d represents the feature dimension, tanh represents the activation function, and the subsequence embedding in the region is expressed as

S315: Calculate

The attention score matrix of , the calculation expression is as follows:

Among them, At ^l represents the attention score matrix, W ₂ , W ₃ , b ₁ and b ₂ all represent learnable parameters, and Softmax represents the activation function; S316: using At ^l and

Calculate and obtain the local feature vector representation Loc _u of the user in the spatio-temporal region; S317: Traverse all user historical POI trajectory sequences in the training set, and calculate the local feature vector representation of each user in the spatio-temporal region. 3. A kind of POI prediction method based on spatio-temporal perception combined with local and global preferences as claimed in claim 2, characterized in that: in said S320, the specific steps of calculating the user's global feature vector representation Glo _u on the spatio-temporal region are as follows : S321: Randomly select a user's historical POI trajectory sequence, and fuse the time information in the user's historical POI trajectory sequence, and the calculation expression is as follows:

in,

Indicates the user's historical POI trajectory sequence after time information fusion is completed, W ₄ indicates the learnable parameters,

A concatenated vector representing a sequence of trajectories and a special time period; S322: Calculate

non-intrusive self-attention

The calculation expression is as follows:

Among them, N means that there are N layers of multi-head attention network layers, Attention(·) means the attention function, Q, K, and V are respectively from

The learnable matrix obtained by mapping with S _u , K ^T represents the transpose matrix of K, and σ represents the learnable parameters; S323: Calculate N-layer multi-head attention, the calculation expression is as follows:

in,

is the multi-head attention network layer of the y-th layer, A ^h is the output of the multi-head attention layer, GELU represents the Gaussian error linear unit, W ₅ , W ₆ , b ₅ , and b ₆ represent the learnable parameters; S324: Perform layer normalization processing and dropout function processing on the output of each sub-layer, and obtain the global feature vector representation Glo _u of the user in the spatio-temporal region; S325: Traverse all user historical POI trajectory sequences in the training set, and calculate the global feature vector representation of each user in the spatio-temporal region. 4. A kind of POI prediction method based on spatio-temporal perception combined with local and global preferences as claimed in claim 3, characterized in that: the specific steps of performing temporal information fusion in the described S321 are as follows: S321-1: The historical trajectory sequence of each user is

Special temporal patterns in the sequence of user historical POI trajectories,

in,

Represents the location representation vector in the trajectory sequence, t _i represents the representation vector of a special time period, u represents the user, and week represents the special time period in weeks; S321-2: According to the time unit conversion, convert the POI check-in time into a special time period in weeks, and obtain the embedded matrix of the special time period

Calculate the embedding matrix E(S _u ) of the trajectory sequence according to the word2vec word embedding method; S321-3: will

and E(S _u ) are spliced on the feature dimension to get the splicing matrix

The specific expression is as follows:

Among them, con( ) represents the splicing function; S321-4: Use the activation function pair

Process to obtain the user's historical POI trajectory sequence after time information fusion

Images