Next interest point recommendation method and device, equipment and storage medium

Abstract

The invention discloses a next interest point recommendation method and device, equipment and a storage medium. The method comprises the following steps: acquiring a target interest point embedding matrix corresponding to historical interest points in a historical sign-in sequence from an interest point embedding matrix generated by a decorrelation adaptive simple graph learning network; determining a target interest point vector based on the target interest point embedding matrix, and performing position coding on the target interest point vector to obtain a user access sequence; performing space-time interval calculation and embedding operation on the historical sign-in sequence to generate a time and space embedding matrix; carrying out attention processing on the user access sequence and the time and space embedding matrix based on a self-attention module, and then outputting a next interest point; wherein embedding operation is carried out on all interest points to obtain an initial embedding matrix, cosine similarity calculation is carried out on the initial embedding matrix to construct a self-adaptive simple graph, and noise reduction is carried out on the self-adaptive simple graph layer by layer through an orthogonal regularization graph signal noise reduction method in the self-adaptive simple graph propagation learning process to generate a decorrelation self-adaptive simple graph learning network.

Description

Technical Field

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

[0002] With the popularity of mobile devices, users often provide rich POI (Point-of-Interest) access information (such as shopping malls and pubs) on location-based social networks (LSBNs); in order to provide users with better recommendation services and improve user experience, it is very important to study POI recommendation. Among them, the next POI recommendation, as a type of POI recommendation, is based on the user's historical check-in data and the user's current status, and recommends more personalized POIs to users by fully capturing the trend of user 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 a weaker 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 it takes to reach it, the two cannot be confused.

[0003] In related technologies, the recommendation of the next POI is achieved through a graph-based representation learning method, which uses a graph neural network (GNN) to provide powerful expression capabilities for the POI embedding vector, enabling downstream sequence tasks to better capture the intricate potential dependencies between POIs.

[0004] Although the graph representation learning method provides good quality POI embedding, the predefined graph structure it constructs will inevitably introduce 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 human factors; therefore, once the designer makes some mistakes or the transition relationship between some POIs cannot be recorded due to accidents, the resulting graph will not be able to well represent the transition relationship between POIs and their potential relationship; in addition, as a type of graph neural network, the graph representation learning method naturally cannot avoid the two major problems of 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 are issues that need to be solved urgently. Summary of the invention

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

[0006] In a first aspect, an embodiment of the present application provides a next point of interest recommendation method, the next point of interest recommendation method comprising: Obtaining all target interest point embedding matrices corresponding to historical interest points from a preset embedding matrix library, wherein the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network, and the historical interest points are interest points included in a historical check-in sequence corresponding to the target user; 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; Perform time-space interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; Based on the preset self-attention module, the user access sequence, time embedding matrix and spatial embedding matrix are processed 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: after embedding all interest points, an initial embedding matrix is ​​obtained, the cosine similarity of the initial embedding matrix is ​​calculated to construct an adaptive simple graph, and the adaptive simple graph is controlled to perform propagation learning, and in 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 decorrelation adaptive simple graph learning network.

[0007] In combination with the first aspect, 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:

[0008] In the formula, represents a feed-forward neural network, Representation layer normalization, represents the attention mechanism, Represents the input of the self-attention module.

[0009] In combination with the first aspect, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention, and the local attention implements local attention coefficient calculation through local masking to re-weight historical points of interest visited recently.

[0010] In combination with the first aspect, in one implementation, the local attention coefficient The calculation formula is:

[0011] 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, Indicates the size of the local field of view.

[0012] In combination with the first aspect, in one implementation, 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, and the additional loss function is:

[0013] In the formula, Indicates The attention loss of the layer attention module, Indicates the length of the historical sign-in sequence, represents the number of attention heads, Indicates The first layer of the attention module A head of attention, Indicates The first layer of the attention module An attention head.

[0014] In combination with the first aspect, in one implementation, the denoising formula corresponding to the orthogonal regularized graph signal denoising method is:

[0015] In the formula, represents the embedding matrix of the graph neural network output, represents the embedding matrix after denoising, represents the embedding matrix of the input graph neural network, and represents the hyperparameter, represents the normalized Laplacian matrix, Represents the identity matrix.

[0016] In combination with the first aspect, in one implementation, the propagation rules of each layer in the decorrelation adaptive simple graph are:

[0017] In the formula, Indicates the The embedding matrix output after layer propagation, Indicates the - The embedding matrix output after 1 layer of propagation, represents the hyperparameter, Represents the Laplace-normalized adjacency matrix with added self-connection.

[0018] In a second aspect, an embodiment of the present application provides a next point of interest recommendation device, the next point of interest recommendation device comprising: 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, wherein the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network, and the historical interest points are interest points included in a historical check-in sequence corresponding to the target user; A second processing module 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; A third processing module is used to perform time-space interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; The POI recommendation module is used to perform attention processing on the user access sequence, the temporal embedding matrix and the spatial embedding matrix based on the preset self-attention module to output the next POI corresponding to the target user; A model building module is used to obtain an initial embedding matrix after embedding all interest points, 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. In 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.

[0019] In combination with the second aspect, 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:

[0020] In the formula, represents a feed-forward neural network, Representation layer normalization, represents the attention mechanism, Represents the input of the self-attention module.

[0021] In combination with the second aspect, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention, and the local attention implements local attention coefficient calculation through local masking to re-weight historical points of interest visited recently.

[0022] In combination with the second aspect, in one implementation, the local attention coefficient The calculation formula is:

[0023] 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, Indicates the size of the local field of view.

[0024] In combination with the second aspect, in one implementation, 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, and the additional loss function is:

[0025] In the formula, Indicates The attention loss of the layer attention module, Indicates the length of the historical sign-in sequence, represents the number of attention heads, Indicates The first layer of the attention module A head of attention, Indicates The first layer of the attention module An attention head.

[0026] In combination with the second aspect, in one implementation, the denoising formula corresponding to the orthogonal regularized graph signal denoising method is:

[0027] In the formula, represents the embedding matrix of the graph neural network output, represents the embedding matrix after denoising, represents the embedding matrix of the input graph neural network, and represents the hyperparameter, represents the normalized Laplacian matrix, Represents the identity matrix.

[0028] In conjunction with the second aspect, in one implementation, the propagation rules of each layer in the decorrelation adaptive simple graph are:

[0029] In the formula, Indicates the The embedding matrix output after layer propagation, Indicates the - The embedding matrix output after 1 layer of propagation, represents the hyperparameter, Represents the Laplace-normalized adjacency matrix with added self-connection.

[0030] In a 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 in the memory and executable by the processor, wherein when the next point of interest recommendation program is executed by the processor, the steps of the next point of interest recommendation method as described above are implemented.

[0031] In a 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, wherein 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.

[0032] The beneficial effects brought by the technical solution provided by the embodiments of the present application include: The initial embedding matrix is ​​obtained by embedding all points of interest, and the cosine similarity is selected to calculate the similarity of the initial embedding matrix and then construct an adaptive adjacency matrix to construct an adaptive simple graph; in the propagation learning process of the adaptive simple graph, the adaptive simple graph is denoised layer by layer by an orthogonal regularized graph signal denoising method 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 relationship between the dimensions of the POI embedding itself, and learn expressive POI embedding, and then alleviate the over-smoothing and over-correlation problems to a certain extent, thereby constructing a decorrelated 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, the embedding matrix of all target interest points corresponding to all historical interest points in the historical check-in sequence can be obtained from the expressive embedding matrix of all interest points generated by the decorrelated adaptive simple graph learning network, and then all target interest point vectors are determined based on the embedding matrix of all target interest points, and all target interest point vectors are position-encoded to obtain the user access sequence; then the historical check-in sequence is subjected to time-space interval calculation and embedding operations to generate the time embedding matrix and the space embedding matrix; finally, the user access sequence, the time embedding matrix and the space embedding matrix are subjected to attention processing by the self-attention module, so as to accurately predict the next interest point of the target user. It can be seen that the present application can effectively reduce the dependence of POI recommendation on prior knowledge and alleviate the problems of over-smoothing and over-correlation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flowchart of an embodiment of a method for recommending points of interest in the present application; Figure 2 This is a schematic diagram of the overall architecture of DASGRAN involved in the embodiment of the present application; Figure 3 This is a schematic diagram of the architecture of the traditional attention module; Figure 4 This is a schematic diagram of the architecture of the extended self-attention module involved in the embodiment of the present application; Figure 5 A schematic diagram of the joint global and local attention aggregation involved in the embodiment of the present application; Figure 6 This is a schematic diagram of the hardware structure of the next point of interest recommendation device involved in the embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0035] First, some technical terms in the present application are explained to facilitate those skilled in the art to understand the present application.

[0036] Definition 1 (Sign-in sequence): Let Represents all users, Represents all POIs, where each POI Each location has a unique latitude and longitude coordinate. It should be noted that each user's sign-in record It can be represented as a triple , which means that the user At the moment Visit POIs, the user's check-in sequence can be expressed as ,in, Indicates user The total number of sign-ins.

[0037] Definition 2 (Next POI Recommendation): Given a user And its from time arrive Sign-in sequence , then the goal of the next POI recommendation task is to generate the next moment based on its historical trajectory The list of top k POIs that the user is most likely to visit.

[0038] Simple Graph Convolutional Network (SGCN): For an undirected graph ,in 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, which is equal to the number of vertices N contained in the vertex set V of the graph. That is, the adjacency matrix A is an N×N matrix. The SGCN based on this l The backward propagation rule of the layer is:

[0039] In the formula, represents the activation function, represents the adjacency matrix added with self-connection and , adding self-connection means adding an edge connecting each node to itself. represents the identity matrix, is a diagonal matrix whose diagonal elements are given by The sum of the corresponding row elements, then The value of the i-th row and i-th column is equal to The sum of the values ​​in the i-th row and j-th column of , is a learnable matrix.

[0040] right Laplace normalization is performed to prevent problems such as gradient explosion during graph convolution, that is, , Represents the Laplace-normalized adjacency matrix with self-connection added; SGCN linearizes the traditional GCN (Graph Convolutional Network) propagation rule:

[0041] Unlike traditional GCN, SGCN directly constructs the output , rather than defining the propagation process of each layer, that is, when SGCN corresponds to GCN, the propagation process of SGCN can be regarded as:

[0042] It is worth noting that It is not the final output. When the final result of SGCN is needed, Do a linear transformation without bias; in other words, at the output When the layer is Perform matrix multiplication, that is, , while in deep learning, linear transformation is usually expressed as , is a learnable weight matrix, is a bias term, so It is called a linear transformation without bias.

[0043] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0044] In a first aspect, an embodiment of the present application provides a method for recommending a next point of interest.

[0045] In one embodiment, referring to Figure 1 , Figure 1This is a flow chart of an embodiment of the method for recommending points of interest in this application. Figure 1 As shown, the next POI recommendation method includes: Step S10: Obtain all target interest point embedding matrices corresponding to historical interest points from a preset embedding matrix library, wherein the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network, and the historical interest points are interest points contained in a historical check-in sequence corresponding to a target user; wherein the decorrelation adaptive simple graph learning network is constructed by: performing an embedding operation on all interest points to obtain an initial embedding matrix, performing cosine similarity calculation on the initial embedding matrix to construct an adaptive simple graph, and controlling the adaptive simple graph to perform propagation learning, and in the propagation learning process of the adaptive simple graph, performing layer-by-layer denoising on the adaptive simple graph by an orthogonal regularized graph signal denoising method to generate a decorrelation adaptive simple graph learning network.

[0046] Exemplarily, it should be understood that the predefined graph structure will inevitably introduce too much prior knowledge, while the adaptive graph structure inherits the key neighbor aggregation idea and nonlinear transformation of GNN, so as to hand over the generation method of the adjacency matrix to POI embedding, that is, to measure the distance between POIs by some distance metric, thereby forming the weight of the edge; 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.

[0047] It should be noted that although the adaptive graph method has produced good results, the POI embedding learned by the adaptive graph is often unsatisfactory when there is no guidance. Therefore, the quality of the learned POI embedding can be improved by introducing the KL divergence measurement method in the process of graph learning; 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 is far away from its original intention (i.e., reducing prior knowledge), and what it wants to compare is the relationship between the two POI matrices. Assuming that the number of POIs is , then the space complexity of the two matrices that need to be compared is , which makes this method unable to be calculated on larger data sets.

[0048] In this embodiment, in order to reduce the influence of human factors in the process of graph learning, a decorrelated adaptive graph method is proposed by exploring graph structure learning, so as to reduce the dependence of POI recommendation on prior knowledge by constructing a decorrelated adaptive simple graph learning network. Figure 2 As shown, before constructing the graph structure, all POIs need to be Perform embedding operation to initialize embedding and obtain the initial embedding matrix , Represents the embedding dimension to facilitate the subsequent learning of the relationship between POIs; then the graph structure is constructed, but the primary issue is which distance measurement method to choose to construct the adjacency matrix. Among them, this embodiment can preferably use cosine similarity as a measurement method, but the weight matrix is ​​discarded in the distance measurement. , that is, directly calculate the distance between POI embeddings to reduce the number of parameters; where each item of the adjacency matrix The calculation is as follows:

[0049] In the formula, Indicates the calculation of the cosine similarity between two vectors; it should be noted here that , negative numbers are unreasonable as weights, and traditional methods usually select a certain value as a threshold to filter 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 be unable to backpropagate, thereby limiting the gradient size of POI embedding; this embodiment uses the sigmoid function to map it to , strictly speaking, it is mapped to ,because The input range avoids the saturation interval of sigmoid. When the back-propagation gradient flows through the adjacency matrix, it is less likely to have gradient explosion or gradient vanishing problems. The adjacency matrix at this time can be expressed as:

[0050] In the formula, Represents element-by-element sigmoid calculation, represents the initial embedding matrix The cosine similarity between rows (including between rows and themselves) in , you can The form is ,Right now Represents the first ,and Represents the second ,So Indicates that the matrix and Perform row normalization to obtain and , and then perform matrix multiplication ; It can be seen that according to the above adjacency matrix, we can construct Figure 2The Adaptive Simple Graph Convolutional Network (ASGCN) shown in Figure 2.

[0051] However, only constructs such Figure 2 The ASGCN shown is not enough because it loses the nonlinear activation and there is no reliable method to guide its learning process. It also faces the problem of over-smoothing and over-correlation, so that it learns low-quality POI embedding. Figure 2 As shown, this embodiment regards each layer of ASGCN as a separate graph denoising process, and modifies its forward propagation method through a decorrelation method, so as to learn high-quality POI embedding without introducing too much prior experience, that is, by integrating the gradient of the orthogonal regularized graph signal denoising method into the graph propagation process to improve the quality of the learned POI embedding, so as to achieve decorrelation; it should be understood that the orthogonal regularized graph signal denoising method forces the columns of the graph node embedding matrix to be orthogonal, so that the difference between the columns increases, that is, the similarity of each dimension in the graph node vector itself becomes lower, so as to obtain more information, so that the learned POI embedding 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 embedding itself, and alleviates the problems of over-smoothing and over-correlation to a certain extent.

[0052] In addition, to avoid gradient explosion, this embodiment will perform column normalization on the embedding after propagation:

[0053] In the formula, Indicates the The embedding matrix output after layer propagation; therefore, when the graph has The decorrelated adaptive simple graph is done by first embedding the input initial matrix Do Second-rate( Represents the number of layers of the graph method) with column normalization propagation, and then does a linear transformation without bias terms to complete the construction of the decorrelated adaptive simple graph learning network.

[0054] In this embodiment, the constructed decorrelation adaptive simple graph learning network is used to generate a corresponding expressive POI embedding matrix for each POI to form an embedding matrix library; therefore, when recommending the next POI to 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. The historical check-in sequence refers to the sequence formed by the check-in of the target user in the past period of time, which includes all the points of interest visited by the target user in the past period of time.

[0055] Furthermore, in one embodiment, the denoising formula corresponding to the orthogonal regularized graph signal denoising method is:

[0056] In the formula, represents the embedding matrix of the graph neural network output, represents the embedding matrix after denoising, represents the embedding matrix of the input graph neural network, and represents the hyperparameter, represents the normalized Laplacian matrix, Represents the identity matrix.

[0057] Among them, the propagation rules of each layer in the decorrelation adaptive simple graph are:

[0058] In the formula, Indicates the The embedding matrix output after layer propagation, Indicates the -1 layer of propagation output embedding matrix, represents the hyperparameter, Represents the Laplace-normalized adjacency matrix with added self-connection.

[0059] Exemplarily, it should be understood that the orthogonal regularized graph signal denoising formula is as follows:

[0060] In the formula, represents the embedding matrix of the graph neural network output, represents the embedding matrix after denoising, represents the embedding matrix of the input graph neural network, and All represent hyperparameters, represents the normalized Laplacian matrix, and ; Among them, when using formula (7) to reduce noise layer by layer for SGCN, it is necessary to calculate its The specific calculation formula of the gradient is as follows:

[0061] So the initial embedding matrix of POI In the The propagation rule of the layer is:

[0062] In the formula, Indicates the The embedding matrix output after layer propagation, Indicates the -1 layer of propagation output embedding matrix, represents a hyperparameter that controls the rate of modification of the original output size. represents the Laplace normalized adjacency matrix with self-connection added; since this embodiment regards SGCN as a single-step graph denoising layer by layer, let , then the modified propagation rule is:

[0063] In order to more clearly describe the graph propagation process of the regularized graph signal denoising method integrated with orthogonality, the input Recorded as , then the calculation process of the decorrelation adaptive simple graph is as follows:

[0064]

[0065] pass After the above process, Perform an unbiased linear transformation:

[0066] Then we can get the processed POI embedding matrix , that is, the learned expressive POI embedding matrix.

[0067] 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 a user access sequence.

[0068] For example, in this embodiment, after obtaining an 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 is based on the user's historical trajectory to infer the location that the user is most likely to visit at the next moment, so firstly, the target user's visit footprint needs to be processed into a fixed length n; if the length is greater than n, it is truncated and only the most recent n data are retained to form a historical check-in sequence. If the length is less than n, then fill it to the left until the length is n to form a historical sign-in sequence. ; However, for a more general expression, this embodiment will use arrive To indicate a padded or truncated user The n visit footprints of .

[0069] Getting the historical sign-in sequence The target POI embedding matrix of each historical POI in Then extract all POIs according to the access time sequence to form a sequence The target interest point vector ,in .

[0070] 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:

[0071] In the formula, Representation Embedding The 2Pth position in the dimension, Representation Embedding The 2P+1th position in the dimension, that is, the odd and even positions are encoded differently; The value of is in the interval Inside, is the sequence length, is the embedding dimension, Representation and Correspondingly, this embodiment takes the input embedding dimension and length into consideration, so that the lower-dimensional embedding can also correctly express the distance relationship during calculation. Figure 2 As shown, the process of adding position encoding is as follows:

[0072]

[0073] After the above calculation, the user access sequence with position coding can be obtained .

[0074] Step S30: performing time-space interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix.

[0075] For example, see Figure 2 As shown, this embodiment will construct a time matrix for each target user. and the space matrix , used to record the time interval and space interval between any two check-ins, that is, n check-in records in the calculation of their spatial distance and user Visit their time intervals and get a historical check-in sequence for the target user Time Matrix and a The space matrix ; Among them, the time matrix and the space matrix The elements in are calculated as follows:

[0076]

[0077] In the formula, Indicates user Minimum value of the time interval; Represents the haversine formula, which is used to calculate the distance between two coordinates; Representing the thresholds of time interval and spatial distance respectively, both are hyperparameters; after obtaining the above time and space relationship, two sets of embedding matrices are obtained through embedding operation, namely, time embedding matrix and the spatial embedding matrix .

[0078] Step S40: Perform attention processing on the user access sequence, the time embedding matrix and the spatial embedding matrix based on the preset self-attention module to output the next point of interest corresponding to the target user.

[0079] Exemplarily, in this embodiment, the user access sequence Time embedding matrix with target users and the spatial 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.

[0080] 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:

[0081] In the formula, represents a feed-forward neural network, Representation layer normalization, represents the attention mechanism, Represents the input of the self-attention module.

[0082] 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:

[0083] The above formula is The gradient of is:

[0084] It is worth noting that Figure 3The shorter residual mechanism inside the attention module cannot give a larger gradient to the input; in this embodiment, the attention module is extended, that is, an additional residual mechanism is added to the attention module (i.e. Figure 4 The dotted line in the figure above shows the part of the image (shown by the dotted line in the figure) to obtain Figure 4 The extended self-attention module shown in Figure 2 has the output of for:

[0085] At this time The gradient of is:

[0086] It can be seen that during the back propagation of the extended self-attention module provided in this embodiment, a larger gradient value is transmitted to the adaptive graph part through the self-attention layer, which can learn more effective POI embedding for the adaptive graph, so that a better graph structure can be constructed in the next round of forward propagation.

[0087] therefore, Stacked into a matrix in the order of user access time The processing fed into the extended self-attention module is as follows:

[0088]

[0089]

[0090] In the formula, for Figure 4 The output of Attention in for Figure 4 The output of FNN in ; Representation layer normalization, row normalization of the last dimension of the input, Indicates the discard method; It means that each value of the matrix C is compared with 0, and the larger one is retained; represents the learnable weight matrix in the FNN block, Represents the bias term in FNN.

[0091] It is worth noting that the extended self-attention module has a total of This example only shows the calculation process of the first block. The calculation process of each subsequent block only needs to calculate the output of the previous block. As Just do the calculation; for the output given in the last block Perform another layer normalization operation:

[0092] In the formula, Indicates user The historical check-in sequence information is output by the extended self-attention module.

[0093] Furthermore, in one embodiment, the attention mechanism is a mechanism combining local attention and global attention, wherein the local attention implements local attention coefficient calculation through local masking to re-weight the historical points of interest visited recently; wherein the local attention coefficient The calculation formula is:

[0094] 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, Indicates the size of the local field of view.

[0095] For example, it should be understood that the traditional attention mechanism is often used to view POI information from a global perspective, so that the current POI sees too much information from the past. Although the Softmax operation can greatly reduce the impact of non-important items on the current decision, the scaling factor This sharpening phenomenon is reduced; to alleviate this problem, see Figure 5 As shown, this embodiment proposes a global and local combined attention method to focus on recent visits through local masks, thereby achieving the effect of re-weighting recent visits.

[0096] In this embodiment, the following calculation formula is used to calculate the global attention coefficient:

[0097]

[0098] 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, a global weight calculation and weighted summation are performed to achieve global aggregation. It should be noted that formula (21) only gives the specific solution method for each element in the global attention coefficient matrix.

[0099] However, when calculating the local attention coefficient, the number of elements calculated by softmax is limited:

[0100] 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 field of view; based on this, local weight calculation and weighted summation are performed to achieve local aggregation. It should be noted that formula (23) only gives the specific solution method for each element in the local attention coefficient matrix.

[0101] Finally, add the global aggregation results and the local aggregation results to output the overall aggregation results:

[0102] in, ,Right now for It should be understood that in the specific implementation, the attention coefficient matrix calculated by formula (21) is set to , then when performing attention aggregation, the global attention is obtained through a lower triangular mask matrix To ensure causality, ,in" " represents the Hadamard product (i.e., bitwise multiplication of elements); while local attention designs a k-diagonal matrix (as opposed to a tridiagonal matrix) , here , and then by To achieve the visual field limitation of local attention, that is .

[0103] It can be seen that the idea of ​​this embodiment is to achieve the re-aggregation effect of recent visits by limiting the current field of view, without using additional attention methods or other sequence modeling methods to model recent preferences, thereby reducing the number of parameters for modeling user long-term and short-term dependency parameters separately in traditional ideas. This embodiment emphasizes recent visits in the aggregation process by forcing the current attention to be re-weighted, so that the current POI refocuses on the recently visited POI, thereby achieving a better recommendation effect.

[0104] Furthermore, in one embodiment, the self-attention module includes an additional loss function, and the additional loss function is used to reduce the similarity of the content learned by the multi-head attention mechanism in the self-attention module, and the additional loss function is:

[0105] In the formula, Indicates The attention loss of the layer attention module, Indicates the length of the historical sign-in sequence, represents the number of attention heads, Indicates The first layer of the attention module A head of attention, Indicates The first layer of the attention module An attention head.

[0106] Exemplarily, it should be understood that when multi-head attention is used directly in the self-attention module, it is inevitable that the heads will learn similar content, that is, the content learned by the multi-head attention is low-rank. In order to enable the multi-head attention to learn richer content, it is necessary to provide guidance for the learning of the multi-head attention; therefore, this embodiment will perform multi-head attention regularization, that is, reduce the similarity of the content learned by the multi-head attention mechanism in the self-attention module through an additional loss function, that is, use cosine similarity in each layer of the attention module to calculate the distance between heads:

[0107] In the formula, Indicates The attention loss of the layer attention module, Indicates the length of the historical sign-in sequence, represents the number of attention heads, Indicates The first layer of the attention module A head of attention, Indicates The first layer of the attention module An attention head.

[0108] Finally, the loss of each layer is added as the overall loss of the self-attention module:

[0109] In the formula, Indicates the number of layers of the attention module.

[0110] 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, the loss is added to the overall loss of the DASGRAN model for joint training. Among them, the overall loss calculation method of the above-mentioned attention module is for the case of a single sample. In the specific implementation, the small batch gradient descent method is usually used for training. Therefore, in order to balance the contribution of each sample when the batch size is b, after obtaining the sum of the batched attention losses of different layers, it is divided by b and then added to the final loss joint training.

[0111] The following examples will illustrate the prediction and optimization of the DASGRAN model. First, for the prediction part, it can be understood that the output obtained after processing by the extended self-attention module is , arranging them in rows to obtain , this sequence fully captures the user's long-term dependence and short-term interest change tendency, so the current user has a certain preference for each POI at time t + 1. Interest score It can be calculated as follows:

[0112] In the formula, express The i-th row of is the POI embedding learned by the decorrelated adaptive simple graph network, Indicates user The interest status at time t; in the recommendation evaluation stage, the POIs that the current user may visit are sorted from high to low according to the interest score to obtain a top-k list.

[0113] 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:

[0114] In the formula, is a true example label, when the user visits the location at time I When , its value is 1, otherwise it is 0. Therefore, after adding the regularization term of multi-head attention, the final loss function is:

[0115] In the formula, is a hyperparameter, is the L2 regularization term for all parameters of the model.

[0116] In summary, this embodiment proposes to use a decorrelated adaptive simple graph neural network to learn POI, which can not only learn richer representations for POI, but also differentiate the dimensions of POI representation to make the vector representation more expressive. Compared with the adaptive graph, it relies less on prior experience and can alleviate the problems of over-smoothing and over-correlation to a certain extent. At the same time, this embodiment also proposes a 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, so that the adaptive graph network can train POI embedding 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 distracting attention caused by distant visits when only global attention is used, 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 multi-head attention by increasing the differences between multi-head attention.

[0117] In a second aspect, an embodiment of the present application also provides a device for recommending a next point of interest.

[0118] In one embodiment, the next point of interest recommendation device includes: 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, wherein the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network, and the historical interest points are interest points included in a historical check-in sequence corresponding to the target user; A second processing module 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; A third processing module is used to perform time-space interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; The POI recommendation module is used to perform attention processing on the user access sequence, the temporal embedding matrix and the spatial embedding matrix based on the preset self-attention module to output the next POI corresponding to the target user; A model building module is used to obtain an initial embedding matrix after embedding all interest points, 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. In 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.

[0119] 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:

[0120] In the formula, represents a feed-forward neural network, Representation layer normalization, represents the attention mechanism, Represents the input of the self-attention module.

[0121] Furthermore, in one embodiment, the attention mechanism is a mechanism that combines local attention and global attention, and the local attention implements local attention coefficient calculation through local masking to re-weight the historical points of interest visited recently.

[0122] Furthermore, in one embodiment, the local attention coefficient The calculation formula is:

[0123] 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, Indicates the size of the local field of view.

[0124] Furthermore, in one embodiment, the self-attention module includes an additional loss function, and the additional loss function is used to reduce the similarity of the content learned by the multi-head attention mechanism in the self-attention module, and the additional loss function is:

[0125] In the formula, Indicates The attention loss of the layer attention module, Indicates the length of the historical sign-in sequence, represents the number of attention heads, Indicates The first layer of the attention module A head of attention, Indicates The first layer of the attention module An attention head.

[0126] Furthermore, in one embodiment, the denoising formula corresponding to the orthogonal regularized graph signal denoising method is:

[0127] In the formula, represents the embedding matrix of the graph neural network output, represents the embedding matrix after denoising, represents the embedding matrix of the input graph neural network, and represents the hyperparameter, represents the normalized Laplacian matrix, Represents the identity matrix.

[0128] Furthermore, in one embodiment, the propagation rules of each layer in the decorrelation adaptive simple graph are:

[0129] In the formula, Indicates the The embedding matrix output after layer propagation, Indicates the - The embedding matrix output after 1 layer of propagation, represents the hyperparameter, Represents the Laplace-normalized adjacency matrix with added self-connection.

[0130] Among them, the functional implementation of each module in the above-mentioned next interest point recommendation device corresponds to each step in the above-mentioned next interest point recommendation method embodiment, and its functions and implementation processes are no longer repeated here.

[0131] In a third aspect, an embodiment of the present application provides a next point of interest recommendation device, which may be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.

[0132] Reference Figure 6 , Figure 6 Schematic diagram of the hardware structure of the next point of interest recommendation device involved in the embodiment 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.

[0133] The communication bus may be of any type and is used to interconnect the processor, the memory, and the communication interface.

[0134] The communication interface includes input / output (I / O) interface, physical interface and logical interface, etc., which are used to realize the interconnection of devices inside the next point of interest recommendation device, and the interface used to realize the interconnection between the next point of interest recommendation device and 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 (Display), a keyboard (Keyboard), etc.

[0135] 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 storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0136] The processor may be a general-purpose processor, and the general-purpose processor may call the next point of interest recommendation program stored in the memory and execute the next point of interest recommendation method provided in the embodiment of the present application. For example, the general-purpose processor may be a central processing unit (CPU). The method executed when the next point of interest recommendation program is called may refer to the various embodiments of the next point of interest recommendation method of the present application, and will not be repeated here.

[0137] Those skilled in the art will understand that Figure 6 The hardware structure shown in the figure does not constitute a limitation on the present application, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0138] In a fourth aspect, an embodiment of the present application also provides a computer-readable storage medium.

[0139] The readable storage medium of the present application stores a next point of interest recommendation program, wherein 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.

[0140] 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 the present application, and will not be repeated here.

[0141] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. 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 optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit "first", "second" and "third" to different types.

[0142] In the description of the embodiments of the present application, "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a specific way.

[0143] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; the “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.

[0144] In some processes described in the embodiments of the present application, multiple operations or steps that appear in a specific order are included, but it should be understood that these operations or steps may not be executed in the order in which they appear in the embodiments of the present application or in parallel, and the sequence number of the operation is only used to distinguish the different operations, and the sequence number itself does 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.

[0145] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course 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 is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD) as described above, and includes a number of instructions for a terminal device to execute the methods described in each embodiment of the present application.

[0146] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for recommending a next point of interest, characterized in that: The next point of interest recommendation method comprises: Obtaining all target interest point embedding matrices corresponding to historical interest points from a preset embedding matrix library, wherein the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network, and the historical interest points are interest points included in a historical check-in sequence corresponding to the target user; 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; Perform time-space interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; Based on the preset self-attention module, the user access sequence, time embedding matrix and spatial embedding matrix are processed 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: after embedding all interest points, an initial embedding matrix is ​​obtained, the cosine similarity of the initial embedding matrix is ​​calculated to construct an adaptive simple graph, and the adaptive simple graph is controlled to perform propagation learning, and in 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 decorrelation adaptive simple graph learning network.

2. The next point of interest recommendation method according to claim 1, characterized in that: The self-attention module includes an additional residual mechanism based on which the output of the self-attention module is controlled for: In the formula, represents a feed-forward neural network, Representation layer normalization, represents the attention mechanism, Represents the input of the self-attention module.

3. The next point of interest recommendation method according to claim 2, characterized in that: 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.

4. The next point of interest recommendation method according to claim 3, characterized in that: The local attention coefficient The calculation formula is: 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, Indicates the size of the local field of view.

5. The next point of interest recommendation method according to claim 2, characterized in that: The self-attention module includes an additional loss function, which 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, Indicates The attention loss of the layer attention module, Indicates the length of the historical sign-in sequence, represents the number of attention heads, Indicates The first layer of the attention module A head of attention, Indicates The first layer of the attention module An attention head.

6. The next point of interest recommendation method according to claim 1, characterized in that: The denoising formula corresponding to the orthogonal regularized graph signal denoising method is: In the formula, represents the embedding matrix of the graph neural network output, represents the embedding matrix after denoising, represents the embedding matrix of the input graph neural network, and represents the hyperparameter, represents the normalized Laplacian matrix, Represents the identity matrix.

7. The next point of interest recommendation method according to claim 6, characterized in that: The propagation rules of each layer in the decorrelation adaptive simple graph are: In the formula, Indicates the The embedding matrix output after layer propagation, Indicates the - The embedding matrix output after 1 layer of propagation, represents the hyperparameter, Represents the Laplace-normalized adjacency matrix with added self-connection.

8. A device for recommending a next point of interest, characterized in that: The next point of interest recommendation device comprises: 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, wherein the interest point embedding matrices in the embedding matrix library are generated based on a preset decorrelation adaptive simple graph learning network, and the historical interest points are interest points included in a historical check-in sequence corresponding to the target user; A second processing module 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; A third processing module is used to perform time-space interval calculation and embedding operations on the historical check-in sequence to generate a time embedding matrix and a space embedding matrix; The POI recommendation module is used to perform attention processing on the user access sequence, the temporal embedding matrix and the spatial embedding matrix based on the preset self-attention module to output the next POI corresponding to the target user; A model building module is used to obtain an initial embedding matrix after embedding all interest points, 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. In 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.

9. A device for recommending a next point of interest, characterized in that: The next point of interest recommendation device includes a processor, a memory, and a next point of interest recommendation program stored in the memory and executable by the processor, wherein when the next point of interest recommendation program is executed by the processor, the steps of the next point of interest recommendation method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a next point of interest recommendation program, wherein 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 7 are implemented.

Images