Next Point of Interest Recommendation Method, Apparatus, Device, and Storage Medium

Abstract

A method, device, equipment and storage medium for recommending the next point of interest, including obtaining a target point-of-interest embedding matrix corresponding to historical points of interest in a historical check-in sequence from a point-of-interest embedding matrix generated by a decorrelated adaptive simple graph learning network; determining a target point-of-interest vector based on the target point-of-interest embedding matrix and performing positional encoding on the target point-of-interest vector to obtain a user access sequence; performing spatio-temporal interval calculation and embedding operation on the historical check-in sequence to generate a time and space embedding matrix; performing attention processing on the user access sequence, the time and space embedding matrix based on a self-attention module and then outputting the next point of interest; wherein, performing an embedding operation on all points of interest to obtain an initial embedding matrix and calculating the cosine similarity thereof to construct an adaptive simple graph, and layer-by-layer denoising the adaptive simple graph through an orthogonal regularized graph signal denoising method during the adaptive simple graph propagation learning process to generate a decorrelated adaptive simple graph learning network.

Description

Technical Field

[0001] The present application relates to the technical field of point-of-interest recommendation, and particularly relates to a method, device, equipment and storage medium for recommending the next point-of-interest. Background Art

[0002] With the popularization of mobile devices, users often provide rich point-of-interest (POI) access information (such as commercial buildings and taverns) on location-based social networks (LSBNs). In order to provide users with better recommendation services and improve the user experience, it is very important to study POI recommendation. Among them, as a type of POI recommendation, the next POI recommendation is based on the user's historical check-in data and the user's current status, and recommends more personalized POIs for users by fully capturing the dynamic changes of the user's interest preferences. The characteristics of the next POI recommendation task are usually affected by time factors and geographical factors, and there is also a potential relationship between the two. For example, users have less interest in POIs that are farther away and take more time to reach. However, although there is a potential relationship between the distance of the POI and the time required to reach it, the two cannot be confused.

[0003] In the related art, the recommendation of the next POI is achieved by using a graph representation learning method, which uses a graph neural network (GNN) to provide powerful expressive ability for the embedding vectors of POIs, enabling downstream sequence tasks to better capture the intricate potential dependencies between POIs.

[0004] Although the graph representation learning method provides good-quality POI embeddings, the pre-defined graph structure it constructs inevitably introduces too much prior knowledge, so that the structure of the graph (such as whether an edge exists and the weight of the edge) is affected by humans. Therefore, once there are some errors in the designer or some transition relationships between POIs cannot be recorded accidentally, the resulting graph cannot well represent the transition relationships and potential relationships between POIs. In addition, as a type of graph neural network, the graph representation learning method naturally cannot avoid the two major problems existing in the graph itself: over-smoothing and over-correlation. It can be seen that when implementing the next POI recommendation, how to reduce the dependence on prior knowledge and alleviate the problems of over-smoothing and over-correlation is an urgent problem to be solved currently. Summary of the Invention

[0005] The present application provides a method, an apparatus, a device, and a storage medium for recommending the next point of interest, which can solve the technical problems of high dependence on prior knowledge, over-smoothing, and over-correlation caused by implementing POI recommendation based on a predefined graph structure in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a method for recommending the next point of interest. The method for recommending the next point of interest includes:

[0007] Obtain all target point-of-interest embedding matrices corresponding to historical points of interest from a preset embedding matrix library. The point-of-interest embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network. The historical points of interest are the points of interest included in the historical check-in sequence corresponding to the target user;

[0008] Determine all target point-of-interest vectors based on all the target point-of-interest embedding matrices, and perform position encoding processing on all the target point-of-interest vectors to obtain a user access sequence;

[0009] Perform spatio-temporal interval calculation and embedding operation on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix;

[0010] Perform attention processing on the user access sequence, the time embedding matrix, and the space embedding matrix based on a preset self-attention module to output the next point of interest corresponding to the target user;

[0011] Among them, the construction method of the decorrelation adaptive simple graph learning network is: after performing embedding operation on all points of interest, obtain an initial embedding matrix, calculate the cosine similarity of the initial embedding matrix to construct an adaptive simple graph, and control the adaptive simple graph to perform propagation learning. During the propagation learning process of the adaptive simple graph, layer-by-layer noise reduction is performed on the adaptive simple graph through an orthogonal regularized graph signal denoising method to generate a decorrelation adaptive simple graph learning network.

[0012] In combination with the first aspect, in an implementation manner, the self-attention module includes an additional residual mechanism, and controls the output of the self-attention module based on the additional residual mechanism as:

[0013]

[0014] In the formula, represents a feed-forward neural network, represents layer normalization, represents an attention mechanism, represents the input of the self-attention module.

[0015] In combination with the first aspect, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention. The local attention calculates local attention coefficients through a local mask to re-weight the recently visited historical interest points.

[0016] In combination with the first aspect, in one embodiment, the local attention coefficient is calculated by the following formula:

[0017]

[0018] where represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the size of the local field of view.

[0019] In combination with the first aspect, in one embodiment, the self-attention module includes an additional loss function, and the additional loss function is used to reduce the content similarity learned by the multi-head attention mechanism in the self-attention module. The additional loss function is:

[0020]

[0021] where represents the attention loss of the -th layer attention module, represents the length of the historical check-in sequence, represents the number of attention heads, represents the -th attention head in the -th layer attention module, represents the -th attention head in the -th layer attention module.

[0022] In combination with the first aspect, in one embodiment, the noise reduction formula corresponding to the orthogonal regularized graph signal denoising method is:

[0023]

[0024] where represents the embedding matrix output by the graph neural network, represents the denoised embedding matrix, represents the embedding matrix input to the graph neural network, and represent hyperparameters, represents the normalized Laplacian matrix, represents the identity matrix.

[0025] In combination with the first aspect, in one embodiment, the propagation rules of each layer in the decorrelated adaptive simple graph are as follows:

[0026]

[0027] In the formula, represents the output embedding matrix after propagation through the th layer, represents the output embedding matrix after propagation through the - 1st layer, represents a hyperparameter, represents the adjacency matrix with self-connections after Laplacian normalization.

[0028] In a second aspect, an embodiment of the present application provides a next point of interest recommendation device, and the next point of interest recommendation device includes:

[0029] A first processing module, which is used to obtain all target interest point embedding matrices corresponding to historical interest points from a preset embedding matrix library, and the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelated adaptive simple graph learning network, and the historical interest points are the interest points included in the historical check-in sequence corresponding to the target user;

[0030] A second processing module, which is used to determine all target interest point vectors based on all target interest point embedding matrices, and perform position encoding processing on all target interest point vectors to obtain a user access sequence;

[0031] A third processing module, which is used to perform spatio-temporal interval calculation and embedding operation on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix;

[0032] An interest point recommendation module, which is used to perform attention processing on the user access sequence, the time embedding matrix, and the space embedding matrix based on a preset self-attention module to output the next interest point corresponding to the target user;

[0033] A model construction module, which is used to perform embedding operation on all interest points to obtain an initial embedding matrix, perform cosine similarity calculation on the initial embedding matrix to construct an adaptive simple graph, and control the adaptive simple graph to perform propagation learning, and during the propagation learning process of the adaptive simple graph, layer-by-layer noise reduction is performed on the adaptive simple graph through an orthogonal regularized graph signal denoising method to generate a decorrelated adaptive simple graph learning network.

[0034] In combination with the second aspect, in one embodiment, the self-attention module includes an additional residual mechanism, and controls the output of the self-attention module based on the additional residual mechanism as:

[0035]

[0036] Wherein, represents a feedforward neural network, represents layer normalization, represents an attention mechanism, represents the input of the self-attention module.

[0037] Combined with the second aspect, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention. The local attention calculates local attention coefficients through a local mask to re-weight recently visited historical interest points.

[0038] Combined with the second aspect, in one embodiment, the local attention coefficient is calculated by the following formula:

[0039]

[0040] Wherein, represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the size of the local field of view.

[0041] Combined with the second aspect, in one embodiment, the self-attention module includes an additional loss function. The additional loss function is used to reduce the content similarity learned by the multi-head attention mechanism in the self-attention module. The additional loss function is:

[0042]

[0043] Wherein, represents the attention loss of the -th layer attention module, represents the length of the historical check-in sequence, represents the number of attention heads, represents the -th attention head in the -th layer attention module, represents the -th attention head in the -th layer attention module.

[0044] Combined with the second aspect, in one embodiment, the noise reduction formula corresponding to the orthogonal regularized graph signal noise reduction method is:

[0045]

[0046] In the formula, represents the embedding matrix output by the graph neural network, represents the denoised embedding matrix, represents the embedding matrix input to the graph neural network, and represent hyperparameters, represents the normalized Laplacian matrix, represents the identity matrix.

[0047] Combined with the second aspect, in one implementation, the propagation rules of each layer in the decorrelated adaptive simple graph are:

[0048]

[0049] In the formula, represents the embedding matrix output after propagation through the th layer, represents the embedding matrix output after propagation through the - 1st layer, represents a hyperparameter, represents the adjacency matrix with self-connections after Laplacian normalization.

[0050] In the third aspect, an embodiment of the present application provides a next point of interest recommendation device, which includes a processor, a memory, and a next point of interest recommendation program stored on the memory and executable by the processor. When the next point of interest recommendation program is executed by the processor, the steps of the foregoing next point of interest recommendation method are implemented.

[0051] In the fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a next point of interest recommendation program is stored. When the next point of interest recommendation program is executed by a processor, the steps of the foregoing next point of interest recommendation method are implemented.

[0052] The beneficial effects brought by the technical solution provided by the embodiment of the present application include:

[0053] After performing embedding operations on all points of interest, an initial embedding matrix is obtained. Then, cosine similarity is selected to calculate the similarity of the initial embedding matrix and construct an adaptive adjacency matrix to build an adaptive simple graph. During the propagation learning process of the adaptive simple graph, an orthogonal regularized graph signal denoising method is used to perform layer-by-layer denoising on the adaptive simple graph to modify the forward propagation method of the adaptive simple graph, so as to reduce the influence of prior experience by only caring about the relationships between the dimensions of the POI embedding itself, and learn expressive POI embeddings. Subsequently, the over-smoothing and over-correlation problems are alleviated to a certain extent, thereby constructing a de-correlated adaptive simple graph learning network with low dependence on prior knowledge and alleviated over-smoothing and over-correlation problems. Therefore, when it is necessary to recommend the next POI for the target user, all target interest point embedding matrices corresponding to all historical interest points in the historical check-in sequence can be obtained from all the expressive interest point embedding matrices generated by the de-correlated adaptive simple graph learning network. Then, all target interest point vectors are determined based on all the target interest point embedding matrices, and position encoding processing is performed on all the target interest point vectors to obtain the user access sequence. Then, spatio-temporal interval calculation and embedding operations are performed on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix. Finally, through the self-attention module, attention processing is performed on the user access sequence, the time embedding matrix, and the space embedding matrix, and the next interest point of the target user can be accurately predicted. It can be seen that through this application, the dependence of POI recommendation on prior knowledge can be effectively reduced, and the over-smoothing and over-correlation problems can be alleviated. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic flowchart of an embodiment of the method for recommending the next interest point in this application;

[0055] Figure 2 It is a schematic diagram of the overall architecture of DASGRAN involved in the solution of this embodiment of the application;

[0056] Figure 3 It is a schematic diagram of the architecture of a traditional attention module;

[0057] Figure 4 It is a schematic diagram of the architecture of the extended self-attention module involved in the solution of this embodiment of the application;

[0058] Figure 5 It is a schematic diagram of the joint global and local attention aggregation involved in the solution of this embodiment of the application;

[0059] Figure 6 It is a schematic diagram of the hardware structure of the device for recommending the next interest point involved in the solution of this embodiment of the application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] To enable those skilled in the art to better understand the solution of this application, the following will clearly and completely describe the technical solution in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts belong to the scope of protection of this application.

[0061] First, some technical terms in this application are explained to facilitate understanding by those skilled in the art.

[0062] Definition 1 (Check-in sequence): Let represent all users, represent all POIs, where each POI is associated with a unique latitude and longitude geographical coordinate ; it should be noted that the check-in record of each user can be represented as a triple which represents that user visits the th POI at time . Then the check-in sequence of the user can be represented as , where represents the total number of check-ins of user .

[0063] Definition 2 (Next POI recommendation): Given a user and their check-in sequence from time , the goal of the next POI recommendation task is to generate a list of the top k POIs that the user is most likely to visit at the next time based on their historical trajectory.

[0064] Simple Graph Convolutional Network (SGCN): For an undirected graph , where and represent the vertex set, edge set, and adjacency matrix respectively, the number of rows of the adjacency matrix A is equal to the number of columns and is equal to the number N of vertices contained in the vertex set V of the graph, that is, the adjacency matrix A is an N×N matrix; based on this, the backward propagation rule of the l th layer of SGCN is:

[0065]

[0066] In the formula, denotes an activation function, denotes the adjacency matrix with self-connections added and adding self-connections means adding an edge connecting each node to itself, denotes the identity matrix, is a diagonal matrix, and its diagonal elements are determined by the sum of the corresponding row elements, then the value of the i-th row and i-th column of is equal to the sum of the values of the i-th row and j-th column of , i.e., is a learnable matrix.

[0067] For , Laplacian normalization is performed to prevent problems such as gradient explosion during the graph convolution process, i.e., let , denotes the adjacency matrix with self-connections added after Laplacian normalization; among them, SGCN linearizes the propagation rule of traditional GCN (Graph Convolutional Network):

[0068]

[0069] Different from traditional GCN, SGCN directly constructs the output , rather than defining the propagation process of each layer. That is, when corresponding SGCN to GCN, the propagation process of SGCN can be regarded as:

[0070]

[0071] It should be noted that is not the final output. When the final result of SGCN is needed, an unbiased linear transformation needs to be performed on ; in other words, when outputting the -th layer, it is multiplied by in Equation (2), that is, perform , and in deep learning, the linear transformation is usually expressed as , is a learnable weight matrix, is the bias term, so is called an unbiased linear transformation.

[0072] To make the objectives, technical solutions, and advantages of this application clearer, the following will further describe the embodiments of this application in detail with reference to the accompanying drawings.

[0073] In a first aspect, an embodiment of this application provides a method for recommending the next point of interest.

[0074] In one embodiment, referring to Figure 1 , Figure 1 is a schematic flowchart of an embodiment of the method for recommending the next point of interest in the present application. As Figure 1 shown, the method for recommending the next point of interest includes:

[0075] Step S10: Obtain all target point-of-interest embedding matrices corresponding to the historical points of interest from a preset embedding matrix library. The point-of-interest embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network. The historical points of interest are the points of interest included in the historical check-in sequence corresponding to the target user. Among them, the construction method of the decorrelation adaptive simple graph learning network is: after performing an embedding operation on all points of interest to obtain an initial embedding matrix, calculating the cosine similarity of the initial embedding matrix to construct an adaptive simple graph, and controlling the adaptive simple graph to perform propagation learning. And during the propagation learning process of the adaptive simple graph, layer-by-layer noise reduction is performed on the adaptive simple graph through an orthogonal regularization graph signal denoising method to generate a decorrelation adaptive simple graph learning network.

[0076] Exemplarily, it should be understood that predefined graph structures will inevitably introduce too much prior knowledge, while the adaptive graph structure inherits the key neighbor aggregation idea and nonlinear transformation of GNN to hand over the generation method of the adjacency matrix to the POI embedding, that is, to measure the distance between POIs through a certain distance metric method to form the edge weights. Therefore, the adaptive graph method can be used to avoid introducing too many prior factors and generate a changing graph structure through the learning process of the model to improve the quality of POI embedding.

[0077] It should be noted that although the adaptive graph method has produced good results, when the adaptive graph lacks guidance, the learned POI embeddings are often unsatisfactory. Therefore, the quality of the learned POI embeddings can be improved by introducing a KL divergence metric method during the graph learning process. However, this distribution approximation method that approximates the learned distribution to the distribution derived from prior knowledge introduces additional domain knowledge (i.e., prior knowledge), so that the adaptive graph method deviates from its original intention (i.e., reducing prior knowledge), and what it needs to compare is the relationship between two POI matrices. Assuming the number of POIs is , then the space complexity of the two matrices that need to be compared is , which results in the inability of this method to be calculated on a large dataset.

[0078] In this embodiment, in order to reduce the influence of human factors during the graph learning process, a decorrelation adaptive graph method will be proposed by exploring graph structure learning to reduce the dependence of POI recommendation on prior knowledge. Specifically, refer to Figure 2As shown, before constructing the graph structure, it is necessary to first perform an embedding operation on all POIs to achieve initialization embedding and then obtain the initial embedding matrix , where represents the embedding dimension to facilitate the subsequent learning of the relationships between POIs; immediately afterwards, the graph structure is constructed. However, the primary issue is which distance metric method to choose for constructing the adjacency matrix. Among them, in this embodiment, cosine similarity can be preferably used as the metric method. However, the weight matrix is discarded during the distance measurement , that is, the distance between POI embeddings is directly calculated to reduce the number of parameters; among them, each item of the adjacency matrix

[0079]

[0080] In the formula, represents calculating the cosine similarity between two vectors; it should be noted here that , it is unreasonable to use negative numbers as weights, and the traditional method usually selects a certain value as the threshold to filter out unreasonable values. For example, setting the threshold to 0 is equivalent to using the Relu function on each value, but this will cause some elements of the adjacency matrix to not be able to backpropagate, thus restricting the gradient magnitude of the POI embedding; while in this embodiment, the sigmoid function is used to map it to , strictly speaking, it is mapped to . Since the input range avoids the saturation interval of the sigmoid, when the backpropagation gradient flows through the adjacency matrix, it is not easy to have the problem of gradient explosion or gradient disappearance. Then the adjacency matrix at this time can be expressed as:

[0081]

[0082] In the formula, represents the element-wise sigmoid calculation, represents the initial embedding matrix the cosine similarity between rows (including between a row and itself) in . It should be noted that to better explain , can be in the form of representing the first , and representing the second . Then means that the matrices and are row-normalized to obtain and , and then matrix multiplication is performed ; It can be seen that according to the above adjacency matrix, an Adaptive Simple Graph (ASGCN) as shown in Figure 2 can be constructed.

[0083] However, just constructing the ASGCN as shown in Figure 2 is not enough because it loses the non-linear activation, and there is no reliable method to guide its learning process. It will also face the problems of over-smoothing and over-correlation, resulting in low-quality POI embeddings. Referring to Figure 2 as shown, in this embodiment, each layer of the ASGCN is regarded as an independent graph denoising process, and its forward propagation method is modified through a decorrelation method to learn high-quality POI embeddings without introducing too much prior experience. That is, the quality of the learned POI embeddings is improved by integrating the gradient of the orthogonal regularized graph signal denoising method into the graph propagation process to achieve decorrelation; it should be understood that by forcing the columns of the graph node embedding matrix to be orthogonal through the orthogonal regularized graph signal denoising method, the difference between columns is increased, that is, the similarity of each dimension in the vector of the graph node itself becomes lower, so as to obtain more information, so that the learned POI embeddings can express richer information; it can be seen that this embodiment reduces the influence of prior experience by only caring about the relationship between the dimensions of the POI embeddings themselves, and alleviates the problems of over-smoothing and over-correlation to a certain extent.

[0084] In addition, to avoid gradient explosion, this embodiment will perform column normalization on the embeddings after propagation:

[0085]

[0086] In the formula, represents the embedding matrix output after propagation through the layer; therefore, when the graph has layers, the approach of the decorrelated adaptive simple graph is: first perform times ( ( represents the number of layers of this graph method)) of propagation with column normalization on the input initial embedding matrix , and then perform a linear transformation without a bias term, thus completing the construction of the decorrelated adaptive simple graph learning network.

[0087] In this embodiment, a de-correlated adaptive simple graph learning network is constructed to generate an expressive POI embedding matrix corresponding to each POI, so as to form an embedding matrix library. Therefore, when making the next POI recommendation for the target user, the target POI embedding matrix corresponding to each historical POI included in the historical check-in sequence of the target user can be directly obtained from the embedding matrix library. Among them, the historical check-in sequence refers to the sequence formed by the target user's check-ins in the past period of time, which includes all the points of interest visited by the target user in the past period of time.

[0088] Further, in one embodiment, the noise reduction formula corresponding to the orthogonal regularization graph signal noise reduction method is:

[0089]

[0090] In the formula, represents the embedding matrix output by the graph neural network, represents the denoised embedding matrix, represents the embedding matrix input to the graph neural network, and represent hyperparameters, represents the normalized Laplacian matrix, represents the identity matrix.

[0091] Among them, the propagation rules of each layer in the de-correlated adaptive simple graph are:

[0092]

[0093] In the formula, represents the embedding matrix output after propagation through the th layer, represents the embedding matrix output after propagation through the -1th layer, represents a hyperparameter, represents the adjacency matrix with self-connection after Laplacian normalization.

[0094] Exemplarily, it should be understood that the orthogonal regularization graph signal noise reduction formula is as follows:

[0095]

[0096] In the formula, represents the embedding matrix output by the graph neural network, represents the denoised embedding matrix, represents the embedding matrix input to the graph neural network, and both represent hyperparameters, denotes the normalized Laplacian matrix, and ; among them, when using formula (7) to denoise SGCN layer by layer, it is necessary to calculate its gradient, and the specific calculation formula is as follows:

[0097]

[0098] Therefore, the initial embedding matrix of the POI at the layer has the following propagation rule:

[0099]

[0100] In the formula, denotes the embedding matrix output after propagation through the layer, denotes the embedding matrix output after propagation through the -1 layer, denotes a hyperparameter, which is used to control the modification rate of the original output size, denotes the adjacency matrix with self-connection added after Laplacian normalization; since in this embodiment, SGCN is regarded as single-step graph denoising layer by layer, so let , then the modified propagation rule is:

[0101]

[0102] To more clearly describe the graph propagation process integrating the orthogonal regularized graph signal denoising method, the input can be denoted as , then the calculation process of the decorrelated adaptive simple graph is as follows:

[0103]

[0104]

[0105] Through times of the above process, a bias-free linear transformation is performed on :

[0106]

[0107] Then the processed POI embedding matrix can be obtained, that is, the learned POI embedding matrix with rich expressiveness.

[0108] Step S20: Determine all target interest point vectors based on all target interest point embedding matrices, and perform position encoding processing on all target interest point vectors to obtain the user access sequence.

[0109] Exemplarily, in this embodiment, after obtaining the expressive POI embedding matrix, the check-in sequence encoding will be performed based on it. It should be understood that the next POI recommendation task infers the most likely location that the user will visit at the next moment based on the user's historical trajectory. Therefore, first, the access footprints of the target user need to be processed into a fixed length n; if the length is greater than n, it will be truncated, and only the most recent n pieces of data will be retained to form the historical check-in sequence ; if the length is less than n, it will be filled to the left until the length is n to form the historical check-in sequence ; however, for a more general expression, this embodiment will use to to represent the n access footprints of the filled or truncated user , that is, the historical check-in sequence is .

[0110] After obtaining the target POI embedding matrix of each historical POI in the historical check-in sequence , all POIs are extracted according to the access time sequence to form the target interest point vector of the sequence , where . .

[0111] Since the self-attention mechanism is insensitive to the order of positions, this embodiment uses the absolute position encoding method tAPE to perceive the position information:

[0112]

[0113] In the formula, represents the 2Pth position in the embedding dimension, represents the (2P + 1)th position in the embedding dimension, that is, different encoding methods for odd and even positions; takes values in the interval , is the sequence length, is the embedding dimension, represents corresponding to ; in this embodiment, by considering the input embedding dimension and length, the lower-dimensional embeddings can also correctly express the distance relationship during calculation. Refer to Figure 2 as shown, the process of adding position encoding is as follows:

[0114]

[0115]

[0116] After the above calculations, the user access sequence with positional encoding can be obtained .

[0117] Step S30: Perform spatio-temporal interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix.

[0118] Exemplarily, referring to Figure 2 shown, in this embodiment, a time matrix and a space matrix will be constructed for the target user respectively, which are used to record the time interval and space interval between any two check-ins, that is, for calculate the spatial distance of the n check-in records in and the time interval for the user to visit them, and then obtain a time matrix and a space matrix for the historical check-in sequence of the target user; where the calculation methods of the elements in the time matrix and the space matrix are as follows:

[0119]

[0120]

[0121] In the formula, represents the minimum value of the time interval of the user ; represents the haversine formula, which is used to calculate the distance between the coordinates of the two; respectively represent the thresholds of the time interval and the space distance, both of which are hyperparameters; after obtaining the above time and space relationships, two groups of embedding matrices are obtained through embedding operations, namely the time embedding matrix and the space embedding matrix .

[0122] Step S40: Perform attention processing on the user access sequence, the time embedding matrix, and the space embedding matrix based on a preset self-attention module to output the next point of interest corresponding to the target user.

[0123] Exemplarily, in this embodiment, the user access sequence is combined with the time embedding matrix and the space embedding matrix The data is sent to the self-attention module for attention aggregation and output, and the prediction of the POI that the target user will visit at the next moment of each current moment can be output, that is, the list of the top k POIs that the target user is most likely to visit at the next moment can be output; it should be noted that the specific value of k can be determined according to actual needs and is not limited here. It can be seen that this embodiment uses a simpler idea to perform graph structure learning, which can not only effectively reduce the dependence of POI recommendation on prior knowledge, but also alleviate the problems of over-smoothing and over-correlation.

[0124] Furthermore, in one embodiment, the self-attention module includes an additional residual mechanism, and the output of the self-attention module is controlled based on the additional residual mechanism. for:

[0125]

[0126] In the formula, represents a feed-forward neural network, Representation layer normalization, represents the attention mechanism, Represents the input of the self-attention module.

[0127] For example, see Figure 3 As shown, although the traditional attention module expands the attention mechanism in terms of spatiotemporal information, it does not fully consider the gradient relationship between the adaptive graph and the attention module. It is an implementation form of an Encoder similar to the Pre-LN Transformer, where Attention represents the spatiotemporal interval-aware self-attention mechanism. The original structure is simplified in the figure (that is, the specific information aggregation process is not drawn), FNN represents feedforward neural network, LayerNorm represents layer normalization, Dropout represents dropout, Q, K, and V represent query, key, and value, respectively; for the sake of simplicity in description, subsequent embodiments will use "Attn" to represent "Attention" and "LN" to represent "LayerNorm". Figure 3 For the attention module in , assuming the input is , then the output of this module It can be expressed as:

[0128]

[0129] The above formula is The gradient of is:

[0130]

[0131] It is worth noting that Figure 3The shorter residual mechanism inside the attention module cannot give a large gradient to the input; in this embodiment, the attention module will be extended, that is, an additional residual mechanism will be added to the attention module (i.e., Figure 4 the part shown by the dotted line in Figure 4 ), to obtain an extended self-attention module as shown in . Then the output of the extended self-attention module

[0132]

[0133] At this time, the gradient of is:

[0134]

[0135] It can be seen that when the extended self-attention module provided in this embodiment performs backpropagation, a larger gradient value is transmitted to the adaptive graph part through the self-attention layer, and can learn more effective POI embeddings for the adaptive graph, so that a better graph structure can be constructed in the next forward propagation.

[0136] Therefore, stacked into a matrix in the order of user access time The processing process of sending it into the extended self-attention module is as follows:

[0137]

[0138]

[0139]

[0140] In the formula, is the output of Attention in Figure 4 , is the output of FNN in Figure 4 ; represents layer normalization, which performs row normalization on the last dimension of the input, represents dropout; represents comparing each value of matrix C with 0 and keeping the larger one; represents the learnable weight matrix in the FNN block, represents the bias term in the FNN.

[0141] It should be noted that the extended self-attention module has a total of blocks. This embodiment only gives the calculation process of the first block. For the calculation process of each subsequent block, only the output of its previous block needs to be used as for calculation; for the output Perform another layer normalization operation:

[0142]

[0143] In the formula, represents the result output by the extended self-attention module for the historical check-in sequence information of the user .

[0144] Furthermore, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention. The local attention calculates the local attention coefficient through a local mask to re-weight the historical points of interest visited recently. Among them, the local attention coefficient The calculation formula is:

[0145]

[0146] In the formula, represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the size of the local field of view.

[0147] Exemplarily, it should be understood that when implementing the attention mechanism traditionally, the POI information is often viewed from a global perspective, so that the current POI sees information that is too far back. Although the Softmax operation can greatly reduce the influence of unimportant items on the current decision, the scaling factor reduces this sharpening phenomenon; to alleviate this problem, as shown in Figure 5 , this embodiment proposes a global and local combined attention method to focus on recent visits through a local mask, so as to achieve the effect of re-weighting recent visits.

[0148] Among them, when calculating the global attention coefficient in this embodiment, the following calculation formula is used:

[0149]

[0150]

[0151] In the formula, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the global attention coefficient corresponding to the element in the i-th row and j-th column of the global attention coefficient matrix; based on this, global weight calculation and weighted summation are performed to achieve global aggregation. It should be noted that Equation (21) only gives the specific solution method for each element in the global attention coefficient matrix.

[0152] However, when calculating the local attention coefficient, the number of elements for softmax calculation is restricted:

[0153]

[0154] In the formula, represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the size of the local visual field; based on this, local weight calculation and weighted summation are performed to achieve local aggregation. It should be noted that Equation (23) only gives the specific solution method for each element in the local attention coefficient matrix.

[0155] Finally, adding the global aggregation result and the local aggregation result can output the overall aggregation result:

[0156]

[0157] Among them, , that is is for each row in. It should be understood that in specific implementation, the attention coefficient matrix calculated by Equation (21) is set as , then when performing attention aggregation, the global attention uses a lower triangular mask matrix to ensure its causality, that is , where " " represents the Hadamard product (i.e., element-wise multiplication); while for local attention, a k-diagonal matrix (compared to a tridiagonal matrix) is designed, here , and then through to achieve the visual field limitation of local attention, that is .

[0158] It can be seen that the idea of this embodiment is to achieve the effect of re-aggregating recent accesses by restricting the current visual field, without using additional attention methods or other sequence modeling methods to model recent preferences, thereby reducing the number of parameters for separately modeling user long-term and short-term dependencies in the traditional idea. This embodiment emphasizes recent accesses during the aggregation process by forcing the current attention to be re-weighted, making the current POI refocus on the recently visited POI, thereby achieving a better recommendation effect.

[0159] Furthermore, in one embodiment, the self-attention module includes an additional loss function, and the additional loss function is used to reduce the content similarity learned by the multi-head attention mechanism in the self-attention module. The additional loss function is:

[0160]

[0161] In the formula, represents the attention loss of the -th layer attention module, represents the length of the historical check-in sequence, represents the number of attention heads, represents the -th attention head in the -th layer attention module, represents the -th attention head in the -th layer attention module.

[0162] Exemplarily, it should be understood that when directly using multi-head attention in the self-attention module, it is inevitable that similar content will be learned between heads, that is, the content learned by multi-head attention is of low rank. In order to enable multi-head attention to learn richer content, guidance needs to be provided for the learning of multi-head attention; therefore, in this embodiment, multi-head attention regularization will be performed, that is, the similarity of the content learned by the multi-head attention mechanism in the self-attention module will be reduced through an additional loss function, that is, the cosine similarity is used to calculate the distance between heads in each layer attention module:

[0163]

[0164] In the formula, represents the attention loss of the -th layer attention module, represents the length of the historical check-in sequence, represents the number of attention heads, represents the -th attention head in the -th layer attention module, represents the -th attention head in the -th layer attention module.

[0165] Finally, the losses of each layer are added together as the overall loss of the self-attention module:

[0166]

[0167] In the formula, represents the number of layers of the attention module.

[0168] It should be understood that the extended self-attention module and the decorrelated adaptive simple graph learning network in this embodiment constitute a decorrelated adaptive simple graph representation-enhanced attention network (DASGRAN); therefore, after obtaining the overall loss of the self-attention module, this loss is added to the overall loss of the DASGRAN model for joint training. Among them, the calculation method of the overall loss of the above attention module is for the case of a single sample. In specific implementation, the mini-batch gradient descent method is usually used for training. Therefore, in order to balance the contributions of each sample when the batch size is b, after obtaining the sum of the attention losses of different layers in batches, it is divided by b and then added to the final loss for joint training.

[0169] The following embodiments will illustrate the prediction and optimization of the DASGRAN model. First, for the prediction part, it can be understood that the output obtained after being processed by the extended self-attention module is , which is arranged row by row to obtain . This sequence fully captures the user's long-term dependencies and the tendency of short-term interest changes. Therefore, the interest score of the current user for each POI at time t + 1 can be calculated in the following way:

[0170]

[0171] In the formula, represents the i-th row of , which is the POI embedding learned through the decorrelated adaptive simple graph network, and represents the interest state of user at time t; in the recommendation evaluation stage, the POIs that the current user may visit are sorted according to the interest scores from high to low to obtain a top-k list.

[0172] Secondly, for the model optimization part, this embodiment will use the cross-entropy loss function to measure the distance between the predicted POI and the target POI:

[0173]

[0174] In the formula, is the true positive label, and its value is 1 when the user visits location at time I, otherwise it is 0. Therefore, after adding the regularization term of the multi-head attention, the final loss function is:

[0175]

[0176] In the formula, is a hyperparameter, is the L2 regularization term of all parameters of the model.

[0177] In summary, this embodiment proposes to use a decorrelated adaptive simple graph neural network to learn POIs, which can not only learn richer representations for POIs, but also make the vector representations more expressive with the various dimensions of the differentiated POI representations. Moreover, compared with the adaptive graph, it depends less on prior experience and can alleviate the over-smoothing and over-correlation problems to a certain extent. At the same time, this embodiment also proposes the DASGRAN model, which adds additional residual connections to the extended self-attention network, explains the gradient relationship between the self-attention network and the adaptive graph network, and enables the adaptive graph network to train POI embeddings more effectively. In addition, this embodiment also designs a global and local joint attention calculation method, which can re-weight locally, thereby alleviating the problem of being distracted by distant access when only using global attention and improving the performance of the self-attention network without introducing too many parameters. Finally, an additional loss function is added to the multi-head attention mechanism to improve the learning quality of the multi-head attention by increasing the difference between the multi-head attentions.

[0178] In a second aspect, the embodiments of the present application also provide a next POI recommendation device.

[0179] In one embodiment, the next POI recommendation device includes:

[0180] A first processing module, which is used to obtain all target POI embedding matrices corresponding to historical POIs from a preset embedding matrix library, and the POI embedding matrices in the embedding matrix library are generated based on a preset decorrelated adaptive simple graph learning network, and the historical POIs are the POIs included in the historical check-in sequence corresponding to the target user;

[0181] A second processing module, which is used to determine all target POI vectors based on all target POI embedding matrices and perform position encoding processing on all target POI vectors to obtain a user access sequence;

[0182] A third processing module, which is used to perform spatio-temporal interval calculation and embedding operation on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix;

[0183] A POI recommendation module, which is used to perform attention processing on the user access sequence, the time embedding matrix, and the space embedding matrix based on a preset self-attention module to output the next POI corresponding to the target user;

[0184] The model construction module is used to perform an embedding operation on all points of interest to obtain an initial embedding matrix, calculate the cosine similarity of the initial embedding matrix to construct an adaptive simple graph, and control the adaptive simple graph to perform propagation learning. During the propagation learning process of the adaptive simple graph, the adaptive simple graph is denoised layer by layer through an orthogonal regularized graph signal denoising method to generate a decorrelated adaptive simple graph learning network.

[0185] Further, in one embodiment, the self-attention module includes an additional residual mechanism to control the output of the self-attention module based on the additional residual mechanism. It is:

[0186]

[0187] In the formula, represents a feed-forward neural network, represents layer normalization, represents an attention mechanism, represents the input of the self-attention module.

[0188] Further, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention. The local attention calculates the local attention coefficient through a local mask to re-weight the historical points of interest visited recently.

[0189] Further, in one embodiment, the local attention coefficient The calculation formula is:

[0190]

[0191] In the formula, represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the size of the local field of view.

[0192] Further, in one embodiment, the self-attention module includes an additional loss function. The additional loss function is used to reduce the content similarity learned by the multi-head attention mechanism in the self-attention module. The additional loss function is:

[0193]

[0194] In the formula, represents the attention loss of the layer attention module, represents the length of the historical check-in sequence, Indicates the th attention head in the layer attention module, Indicates the th attention head in the layer attention module.

[0195] Furthermore, in one embodiment, the noise reduction formula corresponding to the orthogonal regularization graph signal noise reduction method is:

[0196]

[0197] In the formula, represents the embedding matrix output by the graph neural network, represents the denoised embedding matrix, represents the embedding matrix input to the graph neural network, and represent hyperparameters, represents the normalized Laplacian matrix, represents the identity matrix.

[0198] Furthermore, in one embodiment, the propagation rules of each layer in the decorrelated adaptive simple graph are:

[0199]

[0200] In the formula, represents the embedding matrix output after propagation through the th layer, represents the embedding matrix output after propagation through the - 1st layer, represents a hyperparameter, represents the adjacency matrix with self-connections after Laplacian normalization.

[0201] Among them, the function implementation of each module in the above next point of interest recommendation device corresponds to each step in the above next point of interest recommendation method embodiment, and its function and implementation process will not be elaborated here one by one.

[0202] In a third aspect, an embodiment of the present application provides a next point of interest recommendation device, and the next point of interest recommendation device can be a device with data processing functions such as a personal computer (PC), a laptop computer, a server, etc.

[0203] Referring to Figure 6 , Figure 6 is the hardware structure schematic diagram of the next point of interest recommendation device involved in the embodiment solution of the present application. In the embodiment of the present application, the next point of interest recommendation device may include a processor, a memory, a communication interface, and a communication bus.

[0204] Among them, the communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0205] The communication interface includes interfaces such as input / output (I / O) interfaces, physical interfaces, and logical interfaces for interconnecting components inside the next point-of-interest recommendation device, as well as interfaces for interconnecting the next point-of-interest recommendation device with other devices (such as other computing devices or user devices). The physical interface can be an Ethernet interface, a fiber optic interface, an ATM interface, etc.; the user device can be a display, a keyboard, etc.

[0206] The memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical memory, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0207] The processor can be a general-purpose processor, and the general-purpose processor can call the next point-of-interest recommendation program stored in the memory and execute the next point-of-interest recommendation method provided by the embodiments of the present application. For example, the general-purpose processor can be a central processing unit (CPU). Among them, the method executed when the next point-of-interest recommendation program is called can refer to the various embodiments of the next point-of-interest recommendation method of the present application and will not be elaborated here.

[0208] Those skilled in the art can understand that Figure 6 the hardware structure shown in [[ ]] does not constitute a limitation to the present application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0209] Fourthly, the embodiments of the present application also provide a computer-readable storage medium.

[0210] The next point-of-interest recommendation program is stored on the readable storage medium of the present application. When the next point-of-interest recommendation program is executed by a processor, the steps of the next point-of-interest recommendation method as described above are implemented.

[0211] Among them, the method implemented when the next point of interest recommendation program is executed can refer to the various embodiments of the next point of interest recommendation method of this application, which will not be elaborated here.

[0212] The terms "including" and "having" in the specification, claims and above-mentioned drawings of this application, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally further include steps or units not listed, or may optionally further include other steps or units inherent to these processes, methods, products or devices. The descriptions of terms such as "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit that "first", "second" and "third" are different types.

[0213] In the description of the embodiments of this application, words such as "exemplary", "for example" or "for illustration" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary", "for example" or "for illustration" in the embodiments of this application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary", "for example" or "for illustration" is intended to present relevant concepts in a specific manner.

[0214] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B; "and / or" in the text is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "a plurality of" means two or more than two.

[0215] In some processes described in the embodiments of this application, a plurality of operations or steps appear in a specific order. However, it should be understood that these operations or steps may not be executed in the order in which they appear in the embodiments of this application or may be executed in parallel. The serial numbers of the operations are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed in sequence or in parallel, and these operations or steps may be combined.

[0216] Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above and includes several instructions for causing a terminal device to execute the methods described in the embodiments of the present application.

[0217] The above are only the preferred embodiments of the present application, and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

Claims

1. A method for recommending the next point of interest, characterized in that The next point of interest recommendation method includes: Obtain all target point of interest embedding matrices corresponding to historical points of interest from a preset embedding matrix library. The point of interest embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network. The historical points of interest are the points of interest included in the historical check-in sequence corresponding to the target user; Determine all target point of interest vectors based on all the target point of interest embedding matrices, and perform positional encoding processing on all the target point of interest vectors to obtain a user access sequence; Perform spatio-temporal interval calculation and embedding operation on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; Perform attention processing on the user access sequence, the time embedding matrix, and the space embedding matrix based on a preset self-attention module to output the next point of interest corresponding to the target user; Among them, the construction method of the decorrelation adaptive simple graph learning network is: perform embedding operation on all points of interest to obtain an initial embedding matrix, calculate the cosine similarity of the initial embedding matrix to construct an adaptive simple graph, and control the adaptive simple graph to perform propagation learning. And during the propagation learning process of the adaptive simple graph, layer-by-layer noise reduction is performed on the adaptive simple graph through an orthogonal regularized graph signal noise reduction method to generate a decorrelation adaptive simple graph learning network; The self-attention module includes an additional residual mechanism, and controls the output of the self-attention module based on the additional residual mechanism It is as follows: wherein, represents a feedforward neural network, represents layer normalization, represents an attention mechanism, represents the input of the self-attention module; The attention mechanism is a mechanism that combines local attention and global attention. The local attention calculates local attention coefficients through a local mask to re-weight the historical points of interest visited recently; The local attention coefficient is calculated by the following formula: In the formula, represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the size of the local field of view.

2. The next point of interest recommendation method according to claim 1, wherein The self-attention module includes an additional loss function, and the additional loss function is used to reduce the content similarity learned by the multi-head attention mechanism in the self-attention module. The additional loss function is: In the formula, represents the attention loss of the -th layer attention module, represents the length of the historical check-in sequence, represents the number of attention heads, represents the -th attention head in the -th layer attention module, represents the -th attention head in the -th layer attention module.

3. The next point of interest recommendation method according to claim 1, wherein The noise reduction formula corresponding to the orthogonal regularized graph signal noise reduction method is: In the formula, represents the embedding matrix output by the graph neural network, represents the denoised embedding matrix, represents the embedding matrix input to the graph neural network, and represent hyperparameters, represents the normalized Laplacian matrix, represents the identity matrix.

4. The next point of interest recommendation method according to claim 3, characterized in that The propagation rules of each layer in the decorrelation adaptive simple graph are: wherein, represents the embedding matrix output after propagation through the th layer, represents the embedding matrix output after propagation through the - 1st layer, represents a hyperparameter, represents the adjacency matrix with self - connection added for Laplacian normalization.

5. A next point of interest recommendation device, characterized in that, The next point of interest recommendation device includes: A first processing module, which is used to obtain all target point of interest embedding matrices corresponding to historical points of interest from a preset embedding matrix library. The point of interest embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network. The historical points of interest are the points of interest included in the historical check-in sequence corresponding to the target user; A second processing module, which is used to determine all target point of interest vectors based on all the target point of interest embedding matrices, and perform positional encoding processing on all the target point of interest vectors to obtain a user access sequence; A third processing module, which is used to perform spatio-temporal interval calculation and embedding operation on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; A point of interest recommendation module, which is used to perform attention processing on the user access sequence, the time embedding matrix, and the space embedding matrix based on a preset self-attention module to output the next point of interest corresponding to the target user; A model construction module, which is used to perform an embedding operation on all points of interest to obtain an initial embedding matrix, calculate the cosine similarity of the initial embedding matrix to construct an adaptive simple graph, and control the adaptive simple graph to perform propagation learning. During the propagation learning process of the adaptive simple graph, layer-by-layer noise reduction is performed on the adaptive simple graph through an orthogonal regularized graph signal noise reduction method to generate a decorrelated adaptive simple graph learning network; Among them, the self-attention module includes an additional residual mechanism, and the output of the self-attention module is controlled based on the additional residual mechanism That is: In the formula, represents a feedforward neural network, represents layer normalization, represents an attention mechanism, represents the input of the self-attention module; The attention mechanism is a mechanism that combines local attention and global attention. The local attention calculates local attention coefficients through a local mask to re-weight the historical points of interest visited recently; The local attention coefficient is calculated by the following formula: In the formula, represents the local attention coefficient corresponding to the element in the i-th row and j-th column of the local attention coefficient matrix, represents the attention score corresponding to the element in the i-th row and j-th column of the attention score matrix, represents the size of the local field of view.

6. A next point of interest recommendation device, characterized in that, The next point of interest recommendation device includes a processor, a memory, and a next point of interest recommendation program stored on the memory and executable by the processor. When the next point of interest recommendation program is executed by the processor, the steps of the next point of interest recommendation method according to any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium, characterized in that, A next point of interest recommendation program is stored on the computer-readable storage medium. When the next point of interest recommendation program is executed by a processor, the steps of the next point of interest recommendation method according to any one of claims 1 to 4 are implemented.

Images