Recommendation methods for cross-regional points of interest based on user preferences and personalized preference shifts

Abstract

This invention discloses a method for recommending points of interest (POIs) in different locations based on user preferences and personalized preference transfer, belonging to the field of terminal location-based recommendation. The method includes: constructing a heterogeneous hypergraph for five different types of nodes, and obtaining user preference representations through training the hypergraph; constructing a POI-category graph, and learning POI representations through a continuous skipping word model; constructing an attention network with POI representations as input to obtain user-transferable features; constructing a parameter learning network using a multilayer perceptron and user-transferable features as input, and constructing a transfer network with user preference representations as input and the output of the parameter learning network as parameters to achieve personalized user preference transfer; constructing a geographic map between POIs based on latitude and longitude, and learning the embedding representations of different POIs through a convolutional network; calculating the score for each POI by combining the user's transferred preferences with the embedding representations of different POIs, thus completing the final recommendation.

Description

Technical Field

[0001] This invention relates to the field of terminal location recommendation, and in particular to a method for recommending points of interest in different locations based on user preferences and personalized preference shifts. Background Technology

[0002] The widespread use of location-aware devices such as smartphones allows users to freely share their check-in activities through various location-based social networks (such as Foursquare and Yelp). The vast amount of user-contributed check-in data makes it possible to develop effective Point-of-Interest (POI) recommendation models. This not only guides users to explore more interesting points of interest but also helps location service providers with targeted advertising. However, traditional POI recommendation models may not provide adequate service when users leave their place of residence and arrive at a new destination, as these models primarily focus on recommending POIs within a specific area.

[0003] Compared to traditional POI recommendations, recommendations for cross-regional POIs cannot directly address user location preferences because there is a difference between user's residential location preferences and cross-regional preferences—a phenomenon known as preference transfer. While some research has attempted to address this interest transfer issue—for example, some studies have proposed a probabilistic generative model, TS-LDA, which learns individual region-dependent preferences based on the category text information of each user's check-in points of interest in each region. Another approach is the TRAINOR cross-regional POI recommendation model, which constructs a multilayer perceptron as a nonlinear mapping to facilitate user preference transfer. However, none of these methods consider personalized preference transfer. Different users have different residential location preferences, leading to different preferences after transfer, thus resulting in personalized preference transfer. Furthermore, some studies have utilized the similarity of user activity trajectories to achieve cross-regional region recommendations rather than POI recommendations. One study proposed a recommendation method for cross-regional users that considers user preferences, social influence, and geographical proximity. However, these models only retain check-ins based on the user's place of residence or treat all check-in records as check-ins based on the user's place of residence when processing user check-in data. This leads to information loss and erroneous information, and fails to effectively uncover user preferences. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for recommending points of interest in different locations based on user preferences and personalized preference shifts, which solves the problems of incomplete information, information errors, and personalized preference shifts.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for recommending points of interest in different locations based on user preferences and personalized preference shifts includes the following steps:

[0007] S1. Construct a heterogeneous hypergraph based on the user's check-in records, with nodes of five different types: user, POI, region, location, and category. Then, obtain the representation of user preferences through the model training method.

[0008] S2. Construct an interest point-category graph and learn the representation of interest points through continuous word skipping; construct an attention network to obtain user-transferable features by taking the representation of POI as input.

[0009] S3. A parameter learning network is constructed using a multilayer perceptron and user-transferable features are used as input. Similarly, a transfer network is constructed using a multilayer perceptron and user preference representation is used as input, along with the output of the parameter learning network as parameters, thereby realizing personalized preference transfer for users.

[0010] S4. For the representation of POIs in different locations, in order to reflect the influence of geographical location between POIs, a geographical map between POIs is constructed based on latitude and longitude, and the embedding representation is learned through a convolutional network.

[0011] S5. Calculate the user's preferences after the transfer and the embedded representation of the POI in the different location to obtain the score of each POI, thereby completing the final recommendation.

[0012] A further improvement to the technical solution of this invention lies in: in S1, the user's preference representation includes:

[0013] Based on the user's check-in records, users, POIs, regions, locations, and categories are treated as different types of nodes. These five different types of nodes form a heterogeneous hyperedge. The hyperedge constructed in this way will make different POIs checked in by the same user distinguishable in terms of region, and POIs in different regions will be connected through the user node.

[0014] Based on heterogeneous hypergraph Define a size of Correlation matrix ,in , For users, For POI, For the region, For location, As a category, and express Otherwise, it is 0; for vertices in a heterogeneous hypergraph The degree of the vertex is determined by Definition; notation It is a diagonal matrix containing vertex degrees; then a heterogeneous hypergraph. adjacency matrix Defined as ,in yes The transpose of ; the values ​​in the adjacency matrix represent the number of times two nodes co-occur; adjacency matrix The Rows represent vertices The neighborhood structure; the adjacency matrix As input, an autoencoder is used as the model to preserve the neighborhood structure; the encoder and decoder are shown in the following formulas:

[0015]

[0016]

[0017] in, For the Sigmoid function, It is the th adjacency matrix OK, and It is a weight matrix. and It is the bias vector;

[0018] The goal of an autoencoder is to minimize the reconstruction error between the input and output. The reconstruction process ensures that the embedding representations of nodes with similar neighborhoods are also similar, thus preserving second-order similarity. Since the adjacency matrix of a heterogeneous hypergraph is very sparse, only the non-zero elements in the adjacency matrix are reconstructed to speed up model training. The reconstruction error is shown below:

[0019]

[0020] Where sign is the sign function;

[0021] A multilayer perceptron is used, and modeling is performed in a nonlinear manner.

[0022] Five nodes The embedded representation is taken as input, where the superscript , , , , The nodes represent five types: user, POI, region, location, and category, respectively. The embedding representations of these five nodes are connected and mapped using a non-linear method, resulting in the following joint representation:

[0023]

[0024] in, For the Sigmoid function, It is a weight matrix. It is the bias vector;

[0025] In obtaining Then, a non-linear layer is used to map them to a probability space to obtain similarity:

[0026]

[0027] in, For the Sigmoid function, It is a weight matrix. It is the bias vector;

[0028] For representing user preferences, an attention network is used to aggregate the representations; the trained heterogeneous hypergraph network, where heterogeneous hyperedges are represented by the nodes they contain, is shown in the following formula:

[0029]

[0030] in, It is the embedding representation of a heterogeneous hyperedge; it is obtained by averaging the embedding representations of the contained nodes.

[0031] Each heterogeneous hyperedge in the heterogeneous hypergraph can be viewed as a user's check-in; therefore, a user's preference can be obtained by aggregating the hyperedges corresponding to all of the user's check-ins. Since the contribution of each check-in to the user's preference is different, an attention network is used for implementation. The user's preference representation is shown below:

[0032]

[0033] in, Indicates user preferences, express The attention score is obtained by designing an attention network; its formula is as follows:

[0034]

[0035]

[0036] in, and These are the weight parameters of the attention network. It is a bias vector. It is the Sigmoid function.

[0037] A further improvement to the technical solution of the present invention is that S2 specifically includes the following steps:

[0038] S21. Construct an interest point-category graph and achieve representation learning of interest points through continuous word skipping;

[0039] To better represent user-transferable features, a pre-training process is performed before model inference to obtain POI representations. An attention network is constructed to extract user-transferable features by taking the representations of POIs that the user has checked in as input. Therefore, the quality of the input representations of POIs that the user has checked in determines the quality of user-transferable features.

[0040] Because each POI is associated with multiple text descriptions, this information makes it relatively easy to construct an interest-category graph. Defined as ,in It is a set of POIs. It is a vocabulary list. It is a set of edges connecting words and POIs, and the representation of interest points is achieved through a continuous skipping word model;

[0041] Using POIs as the central words, the following formula is used to train the representation of POIs.

[0042]

[0043] in, It is a word. It's a point of interest. It is the Softmax function;

[0044] S22. Construct an attention network to obtain user-transferable features by taking the representation of POI as input;

[0045] User-transferable features are derived from each Point of Interest (POI) in the user's historical check-in record. Therefore, the POI representation obtained based on the continuous skipping model is used to extract user-transferable features through an attention network. First, each POI in the user's check-in history contributes differently to interest transfer because the attention mechanism compresses multiple POIs into a single representation, with different parts having different contributions. Therefore, a weighted sum is performed on each POI using the attention mechanism:

[0046]

[0047] in, Indicates user The representation of transferable features, Indicates POI The attention score is interpreted as The importance of user-transferable features; the attention score is obtained by designing an attention network; its formula is as follows:

[0048]

[0049]

[0050] in, and These are the weight parameters of the attention network. It is a bias vector. It is the Sigmoid function; the attention network obtains user-transferable features as input to the learning network to generate the weight parameters of the personalized transfer network.

[0051] A further improvement to the technical solution of this invention lies in the following: In S3, there is a specific relationship between the user's destination preference and the user's transferable features; a parameter learning network is designed based on a multilayer perceptron, which takes the user's transferable features as input and then generates the weight parameters of the transfer network; the parameter learning network is represented as follows:

[0052]

[0053] in, and It is the weight matrix of the parameter learning network. and It is a bias vector. It is the Sigmoid function;

[0054] The input to the transfer network is the user preference representation obtained by S1. The weight matrix of the transfer network is not randomly initialized but is the output of the parameter learning network. Because the parameter network has an output for each user, each user has its own independent weight matrix in the transfer network, thus completing the personalized interest transfer of the user.

[0055] Similarly, the transfer network is completed through a multilayer perceptron, which will... Reconstruction Therefore, the formula for the transfer network is as follows:

[0056]

[0057] in, It is the output of the parameter learning network. It is a transfer network.

[0058] A further improvement to the technical solution of this invention lies in the characterization of the off-site POI in S4, including:

[0059] The geographic location of POIs in the target area helps in understanding users' cross-regional check-in behavior;

[0060] First, each POI in the target region is initialized based on the S2 method, denoted as... ,and in ,Notice Then, an undirected graph is constructed based on the geographical relationships between POIs. ,side The definition is as follows:

[0061]

[0062] in, Indicates POI and The distance is calculated based on latitude and longitude; an adjacency matrix is ​​constructed based on the distance between each pair of POIs. ;

[0063] To capture the spatial relationships between POIs, a convolutional network is used:

[0064]

[0065] in, Represents the weight matrix. It is a bias term; It is an updated target area POI embedding matrix that encodes the geographic impact of POIs.

[0066] A further improvement to the technical solution of the present invention is that, in S5, after the model parameters are optimized, the model is used to recommend POIs for the target area to the user;

[0067] For a user And his / her historical check-in records, generating preference representations after user interest shifts. Then use and The inner product yields the user's rating of the target area's Point of Interest (POI).

[0068]

[0069] Finally, based on the estimated scores, the top k target area points (POIs) were selected as the target for users. Recommended.

[0070] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0071] This invention presents a method for recommending points of interest (POIs) in different locations based on user preferences and personalized preference transfer. Specifically, it constructs a heterogeneous hyperedge based on the user's check-in records, and represents user preferences through training a heterogeneous hypergraph network. It then constructs an interest point-category graph to represent POIs, and an attention network is built to obtain user-transferable features using the POI representations as input. A parameter learning network takes the user-transferable features as input, and a transfer network takes the user's preference representations as input and the output of the feature extraction network as parameters to achieve personalized preference transfer. A POI geographic map is constructed based on the latitude and longitude distances between POIs in different regions, and a GCN is used to represent POIs in the target region. Finally, the user's transferred preferences are calculated with the embedded representations of the different POIs to obtain a score for each POI, thus completing the final recommendation. The method described in this invention can help users quickly explore and plan their travel destinations, and helps businesses understand the distribution of user behavior characteristics, analyze potential users, and obtain good economic and social benefits. Attached Figure Description

[0072] Figure 1 This is a flowchart of the method of the present invention;

[0073] Figure 2 This is a schematic diagram of the cross-regional point of interest recommendation method based on user preferences and personalized preference shifts in this invention;

[0074] Figure 3 This is the POI-category diagram in this invention. Detailed Implementation

[0075] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0076] like Figure 1 As shown, a method for recommending points of interest in different locations based on user preferences and personalized preference shifts specifically includes the following steps:

[0077] S1. Construct a heterogeneous hypergraph based on the user's check-in records, with nodes of five different types: user, POI, region, location, and category. Then, obtain the representation of user preferences through the model training method.

[0078] The preference representation of user S1 includes:

[0079] Before reaching their target area, users will check in not only in their place of residence but also in other areas. Previous studies have either deleted check-ins in other areas or treated them as check-in records in their place of residence. This approach can lead to a bias in the representation of user preferences, or in other words, information loss. Therefore, to address this issue, we propose constructing a heterogeneous hypergraph.

[0080] For the same user, the most appropriate way to distinguish his / her check-in records is by dividing them by the region where the check-in POI is located. POIs in the same region are spatially similar, which satisfies the first law of geography. Similarly, POIs in different regions may be related to each other because of their category information, but they may also be related because they were all checked in by the same user.

[0081] Based on user check-in records, the system is categorized by "user, POI, region, location, and category". Viewed as different types of nodes, five different types of nodes form a heterogeneous hyperedge. The hyperedge constructed in this way will make different POIs checked in by the same user in different regions, and at the same time, POIs in different regions will be connected through user nodes.

[0082] Based on heterogeneous hypergraph Define a size of Correlation matrix ,in , For users, For POI, For the region, For location, As a category, and express Otherwise, it is 0; for vertices in a heterogeneous hypergraph The degree of the vertex is determined by Definition; notation It is a diagonal matrix containing vertex degrees; then a heterogeneous hypergraph. adjacency matrix Defined as ,in yes The transpose of ; the values ​​in the adjacency matrix represent the number of times two nodes co-occur; adjacency matrix The Rows represent vertices The neighborhood structure; the adjacency matrix As input, an autoencoder is used as the model to preserve the neighborhood structure; the encoder and decoder are shown in the following formulas:

[0083]

[0084]

[0085] in, For the Sigmoid function, It is the th adjacency matrix OK, and It is a weight matrix. and It is the bias vector;

[0086] The goal of an autoencoder is to minimize the reconstruction error between the input and output. The reconstruction process ensures that the embedding representations of nodes with similar neighborhoods are also similar, thus preserving second-order similarity. Since the adjacency matrix of a heterogeneous hypergraph is very sparse, only the non-zero elements in the adjacency matrix are reconstructed to speed up model training. The reconstruction error is shown below:

[0087]

[0088] Where sign is the sign function;

[0089] A multilayer perceptron is used, and modeling is performed in a nonlinear manner.

[0090] Five nodes The embedded representation is taken as input, where the superscript , , , , The nodes represent five types: user, POI, region, location, and category, respectively. The embedding representations of these five nodes are connected and mapped using a non-linear method, resulting in the following joint representation:

[0091]

[0092] in, For the Sigmoid function, It is a weight matrix. It is the bias vector;

[0093] In obtaining Then, a non-linear layer is used to map them to a probability space to obtain similarity:

[0094]

[0095] in, For the Sigmoid function, It is a weight matrix. It is the bias vector.

[0096] For representing user preferences, an attention network is used to aggregate the representations; the trained heterogeneous hypergraph network, where heterogeneous hyperedges can be represented by the nodes they contain, is shown in the following formula:

[0097]

[0098] in, It is the embedding representation of a heterogeneous hyperedge. It is obtained by averaging the embedding representations of the contained nodes.

[0099] Each heterogeneous hyperedge in a heterogeneous hypergraph can be viewed as a user's check-in; therefore, a user's preference can be obtained by aggregating the hyperedges corresponding to all of the user's check-ins. Since each check-in contributes differently to the user's preference, an attention network is used to implement this. The user's preference representation is shown below:

[0100]

[0101] in, Indicates user preferences, express The attention score is obtained by designing an attention network. The formula is as follows:

[0102]

[0103]

[0104] in, and These are the weight parameters of the attention network. It is a bias vector. It is the Sigmoid function.

[0105] S2. Construct an interest point-category graph and learn the representation of interest points through Skip-gram; construct an attention network to obtain user-transferable features by taking the representation of POI as input.

[0106] S21. Construct an interest point-category graph as follows: Figure 3 As shown in the figure This represents a Point of Interest (POI). This represents words. Interest point representation learning is achieved through Skip-gram.

[0107] To better represent user-transferable features, a pre-training process is performed before model inference to obtain POI representations. An attention network is constructed to extract user-transferable features by taking the representations of POIs that the user has checked in as input. Therefore, the quality of the input representations of POIs that the user has checked in determines the quality of user-transferable features.

[0108] Because each POI is associated with multiple text descriptions, this information makes it relatively easy to construct an interest-category graph. Defined as ,in It is a set of POIs. It is a vocabulary list. It is a set of edges connecting words and POIs, and the interest points are represented by the Skip-gram model;

[0109] Specifically, using POI as the central keyword, the POI corresponds to The context words are shown in Table 1; the representation training of POIs is achieved through the following formula;

[0110]

[0111] in It is a word. It's a point of interest. It is the Softmax function.

[0112] Table 1. Correspondence between POIs and words

[0113]

[0114] S22. Construct an attention network to obtain user-transferable features by taking the representation of POI as input;

[0115] User transferable features are derived from each Point of Interest (POI) in the user's historical check-in record. Therefore, based on the POI representation obtained above, user transferable features are extracted using an attention network. First, each POI in the user's check-in history contributes differently to interest transfer because the attention mechanism can compress multiple POIs into a single representation, with different parts having different contributions. Therefore, a weighted sum is performed on each POI using the attention mechanism:

[0116]

[0117] in, Indicates user The representation of transferable features, Indicates POI The attention score can be interpreted as The importance of user-transferable features; the attention score is obtained by designing an attention network. Its formula is as follows:

[0118]

[0119]

[0120] in, and These are the weight parameters of the attention network. It is a bias vector. It's the Sigmoid function. The attention network takes user-transferable features as input to the learning network to generate the weight parameters for the personalized transfer network.

[0121] S3. A parameter learning network is constructed using MLP (Multilayer Perceptron) with user-transferable features as input. Similarly, a transfer network is constructed using MLP with user preference representation as input and the output of the parameter learning network as parameters, thereby realizing personalized preference transfer for users.

[0122] Different users have different preferences for destination regions; in other words, user preference transfer is personalized. Simply put, there is a specific relationship between a user's destination preference and their transferable features. A parameter learning network is designed based on a multilayer perceptron, which takes the user's transferable features as input and then generates the weight parameters for the transfer network. The parameter learning network is represented as follows:

[0123]

[0124] in, and It is the weight matrix of the parameter learning network. and It is a bias vector. It is the Sigmoid function;

[0125] The input to the transfer network is the user preference representation obtained by S1. The weight matrix of the transfer network is not randomly initialized but is the output of the parameter learning network. Because the parameter network has an output for each user, each user has its own independent weight matrix in the transfer network, thus enabling personalized interest transfer for users.

[0126] Similarly, the transfer network is completed through a multilayer perceptron, which will... Reconstruction Therefore, the formula for the transfer network is as follows:

[0127]

[0128] in, It is the output of the parameter learning network. It is a transfer network.

[0129] S4. For the representation of POIs in different locations, in order to reflect the influence of geographical location between POIs, a geographic map between POIs is constructed based on latitude and longitude, and the embedding representation is learned through GCN.

[0130] The characterization of the S4 off-site POI includes:

[0131] The geographic location of POIs in the target area helps in understanding users' cross-regional check-in behavior;

[0132] Specifically, firstly, each POI in the target region is initialized based on the S2 method, denoted as... ,and in ,Notice Then, an undirected graph is constructed based on the geographical relationships between POIs. ,side The definition is as follows:

[0133]

[0134] in, Indicates POI and The distance between POIs is calculated based on latitude and longitude; an adjacency matrix can be constructed based on the distance between each pair of POIs. ;

[0135] To capture the spatial relationships between POIs, the following methods are used: Figure 2 The convolutional network (GCN) shown:

[0136]

[0137] in, Represents the weight matrix. It is a bias term; It is an updated target area POI embedding matrix that encodes the geographic impact of POIs.

[0138] S5. Calculate the user's preferences after the transfer and the embedded representation of the POI in the different location to obtain the score of each POI, thereby completing the final recommendation.

[0139] After optimizing the model's parameters, this model can be used to recommend Points of Interest (POIs) to users within their target regions. Specifically, for a user... And his / her historical check-in records, generating preference representations after user interest shifts. Then use and The inner product yields the user's rating of the target area's Point of Interest (POI).

[0140]

[0141] Finally, based on the estimated scores, the top k target region POIs can be selected as the target for users. Recommended.

[0142] Example

[0143] For the same user, the most appropriate way to distinguish their check-in records is by the region where the POIs are located. POIs within the same region are spatially similar, satisfying the first law of geography. Similarly, POIs in different regions may be related due to their category information, but they may also be related because they were all checked in by the same user.

[0144] Based on user check-in records, the system is categorized by "user, POI, region, location, and category". Viewed as different types of nodes, five different types of nodes form a heterogeneous hyperedge. The hyperedge constructed in this way will make different POIs checked in by the same user in different regions, and at the same time, POIs in different regions will be connected through user nodes.

[0145] Learning Heterogeneous Hypergraphs: Heterogeneous hyperedges in heterogeneous hypergraphs are constructed based on user check-in records, thus reflecting the interaction relationships between different objects within the hyperedge. The heterogeneous hypergraph in this invention contains five different types of nodes. However, unlike homogeneous hypergraphs, two aspects are crucial in heterogeneous hypergraphs. First, the indivisibility of heterogeneous hypergraphs: Hyperedges in a network are typically indivisible. In this case, nodes within a hyperedge may have strong relationships, but their subsets may not. For example, in a cross-regional POI recommendation model based on "user, POI, region, location, category," the relationship between "user" and "category" is usually weak. This means that traditional hypergraph learning methods cannot be used solely to decompose hyperedges. Second, the structural preservation of heterogeneous hypergraphs: Local structure is preserved in network embeddings through observable relationships. However, due to the sparsity of the network, many existing relationships cannot be observed, and using only local structure to preserve the overall structure of the hypergraph is insufficient. Global structure, such as neighborhood structure, is affected by data sparsity. How to preserve both local and global structure in a hypergraph is an important question.

[0146] This invention constructs five different types of heterogeneous hypergraphs. Firstly, the heterogeneous hypergraph should satisfy first-order and second-order proximity, with first-order proximity primarily measuring the proximity between nodes. Meta-similarity. For any Nodes , … If this If there is a superedge between any two vertices, then these... The first-order proximity of vertices is defined as 1, but this does not mean that these vertices... For any given vertex, there exists a first-order proximity. Second-order proximity measures the similarity of the neighborhood structures of two nodes. , Defined as The neighborhood. If neighborhood and If their neighborhoods are similar, then Embedded representation and Their embedding representations are similar.

[0147] Based on heterogeneous hypergraph Define a size of Correlation matrix ,in , For users, For POI, For the region, For location, As a category, and express Otherwise, it is 0. For vertices in the hypergraph... The degree of the vertex is determined by Definition. Note. This is a diagonal matrix containing vertex degrees. Then, the heterogeneous hypergraph... adjacency matrix It can be defined as ,in yes The transpose of . The values ​​in the adjacency matrix represent the number of times two nodes co-occur. Adjacency matrix The Rows represent vertices The neighborhood structure. The adjacency matrix. An autoencoder is used as the input model to preserve the neighborhood structure. The encoder and decoder are shown in the following formulas:

[0148]

[0149]

[0150] in, For the Sigmoid function, It is the th adjacency matrix OK, and It is a weight matrix. and It is the bias vector.

[0151] The goal of an autoencoder is to minimize the reconstruction error between the input and output. The reconstruction process ensures that the embedding representations of nodes with similar neighborhoods are also similar, thus preserving second-order similarity. Since the adjacency matrix of a heterogeneous hypergraph is very sparse, only the non-zero elements of the adjacency matrix are reconstructed to speed up model training. The reconstruction error is shown below:

[0152]

[0153] Where sign is the sign function.

[0154] A multilayer perceptron is used, and modeling is performed in a non-linear manner. Five nodes are involved. The embedded representations are taken as input, concatenated, and passed through a nonlinear mapping. Therefore, their joint representation is as follows:

[0155]

[0156] in, For the Sigmoid function, It is a weight matrix. It is the bias vector.

[0157] In obtaining Then, a non-linear layer is used to map them to a probability space to obtain similarity:

[0158]

[0159] in, For the Sigmoid function, It is a weight matrix. It is the bias vector.

[0160] Preferably, the representation of user-transferable features includes:

[0161] To ensure better representation of user-transferable features, a pre-training process is performed before model inference to obtain representations of Points of Interest (POIs). An attention network is constructed using the representations of POIs that the user has checked in as input to extract user-transferable features; therefore, the quality of the input representations of POIs that the user has checked in determines the quality of the user-transferable features.

[0162] Because each POI is associated with multiple text descriptions, this information makes it relatively easy to construct an interest-category graph. Defined as ,in It is a set of POIs. It is a vocabulary list. It is a set of edges connecting words and POIs. Interest points are represented using Skip-gram.

[0163] Specifically, using POI as the central keyword, the POI corresponds to The context words are shown in Table 1. The representation training of POIs is achieved using the following formula.

[0164]

[0165] in, It is a word. It's a point of interest. It is the Softmax function.

[0166] User-transferable features are derived from each Point of Interest (POI) in the user's historical check-in record. Therefore, based on the POI representation obtained above, user-transferable features are extracted using an attention network. Firstly, each POI in the user's check-in history contributes differently to interest transfer because the attention mechanism can compress multiple POIs into a single representation, with different parts contributing differently. Therefore, a weighted sum is performed on each POI using the attention mechanism:

[0167]

[0168] in, Indicates user The representation of transferable features, Indicates POI The attention score can be interpreted as The importance of user-transferable features. Attention scores are obtained by designing an attention network. The formula is as follows:

[0169]

[0170]

[0171] in, and These are the weight parameters of the attention network. It is a bias vector. It's the Sigmoid function. The attention network takes user-transferable features as input to the learning network to generate the weight parameters for the personalized transfer network.

[0172] Preferably, based on the user preference representation obtained in S1 and the user transferable feature representation obtained in S2, the representation of personalized user preference transfer includes:

[0173] Different users have different preferences for destination regions. In other words, user preference transfer is personalized. Simply put, there is a specific relationship between a user's destination preference and their transferable features. A parameter learning network is designed based on a multilayer perceptron, which takes the user's transferable features as input and then generates the weight parameters for the transfer network. The parameter learning network is represented as follows:

[0174]

[0175] in, and It is the weight matrix of the parameter learning network. and It is a bias vector. It is the Sigmoid function.

[0176] The input to the transfer network is the user preference representation obtained by S1. The weight matrix of the transfer network is not randomly initialized but is the output of the parameter learning network. Because the parameter network has an output for each user, each user has its own independent weight matrix in the transfer network, thus enabling personalized interest transfer for users.

[0177] Similarly, the transfer network is completed using a multilayer perceptron. Reconstruction Therefore, the formula for the transfer network is as follows:

[0178]

[0179] in, It is the output of the parameter learning network. It is the bias vector of the transfer network. It's about user preferences. These are the user's preferences after they have switched.

[0180] Preferably, the characterization of off-site POIs includes:

[0181] The geographic location of POIs in the target area helps in understanding users' cross-regional check-in behavior.

[0182] Specifically, firstly, each POI in the target region is initialized based on the S2 method, denoted as... ,and in ,Notice Then, an undirected graph is constructed based on the geographical relationships between POIs. ,side The definition is as follows:

[0183]

[0184] in, Indicates POI and The distance between POIs is calculated based on latitude and longitude. An adjacency matrix can be constructed based on the distance between each pair of POIs. .

[0185] To capture the spatial relationships between POIs, a Generative Convolutional Network (GCN) as shown in the figure below is used:

[0186]

[0187] in, Represents the weight matrix. It is a bias term. It is an updated target area POI embedding matrix that encodes the geographic impact of POIs.

[0188] Preferably, the final recommendation is completed based on the user's post-transfer preference representation and the representation of the POI in a different location obtained in S3 and S4, including:

[0189] After optimizing the model's parameters, this model can be used to recommend Points of Interest (POIs) to users within their target regions. Specifically, for a user... And his / her historical check-in records, generating preference representations after user interest shifts. Then use and The inner product yields the user's rating of the target area's Point of Interest (POI).

[0190]

[0191] Finally, based on the estimated scores, the top k target region POIs can be selected as the target for users. Recommended.

Claims

1. A method for recommending a point of interest based on user preferences and personalized preference transfer, characterized in that: Includes the following steps: S1. Construct a heterogeneous hypergraph based on the user's check-in records, with nodes of five different types: user, POI, region, location, and category. Then, obtain the representation of user preferences through the model training method. S2. Construct an interest point-category graph and learn the representation of interest points through continuous word skipping; construct an attention network to obtain user-transferable features by taking the representation of POI as input. S3. A parameter learning network is constructed using a multilayer perceptron and user-transferable features are used as input. A transfer network is constructed using a multilayer perceptron and user preference representation is used as input, along with the output of the parameter learning network as parameters, thereby realizing personalized preference transfer for users. S4. For the representation of POIs in different locations, in order to reflect the influence of geographical location between POIs, a geographical map between POIs is constructed based on latitude and longitude, and the embedding representation is learned through a convolutional network. S5. Calculate the user's preferences after the transfer and the embedded representation of the POI in the different location to obtain the score of each POI, thereby completing the final recommendation. 2.The off-site point of interest recommendation method based on user preference and personalized preference transfer according to claim 1, characterized in that: In S1, the user's preference representation includes: Based on the user's check-in records, users, POIs, regions, locations, and categories are treated as different types of nodes. These five different types of nodes form a heterogeneous hyperedge. The hyperedge constructed in this way will make different POIs checked in by the same user distinguishable in terms of region, and POIs in different regions will be connected through the user node. Based on heterogeneous hypergraph Define a size of Correlation matrix ,in , For users, For POI, For the region, For location, As a category, and express Otherwise, it is 0; for vertices in a heterogeneous hypergraph The degree of the vertex is determined by Definition; notation It is a diagonal matrix containing vertex degrees; then a heterogeneous hypergraph. adjacency matrix Defined as ,in yes The transpose of ; the values ​​in the adjacency matrix represent the number of times two nodes co-occur; adjacency matrix The Rows represent vertices The neighborhood structure; the adjacency matrix As input, an autoencoder is used as the model to preserve the neighborhood structure; the encoder and decoder are shown in the following formulas: in, For the Sigmoid function, It is the th adjacency matrix OK, and It is a weight matrix. and It is the bias vector; The goal of an autoencoder is to minimize the reconstruction error between the input and output. The reconstruction process ensures that the embedding representations of nodes with similar neighborhoods are also similar, thus preserving second-order similarity. Since the adjacency matrix of a heterogeneous hypergraph is very sparse, only the non-zero elements in the adjacency matrix are reconstructed to speed up model training. The reconstruction error is shown below: Where sign is the sign function; A multilayer perceptron is used, and modeling is performed in a nonlinear manner. Five nodes The embedded representation is taken as input, where the superscript , , , , The nodes represent five types: user, POI, region, location, and category. The embedding representations of these five nodes are connected and mapped using a non-linear method, resulting in the following joint representation: in, For the Sigmoid function, It is a weight matrix. It is the bias vector; In obtaining Then, a non-linear layer is used to map them to a probability space to obtain similarity: in, For the Sigmoid function, It is a weight matrix. It is the bias vector; For representing user preferences, an attention network is used to aggregate the representations; the trained heterogeneous hypergraph network, where heterogeneous hyperedges are represented by the nodes they contain, is shown in the following formula: in, It is the embedding representation of a heterogeneous hyperedge; it is obtained by averaging the embedding representations of the contained nodes. Each heterogeneous hyperedge in the heterogeneous hypergraph can be viewed as a user's check-in; therefore, a user's preference can be obtained by aggregating the hyperedges corresponding to all of the user's check-ins. Since the contribution of each check-in to the user's preference is different, an attention network is used for implementation. The user's preference representation is shown below: in, Indicates user preferences, express The attention score is obtained by designing an attention network; its formula is as follows: in, and These are the weight parameters of the attention network. It is a bias vector. It is the Sigmoid function.

3. The method for recommending points of interest in different locations based on user preferences and personalized preference transfer as described in claim 1, characterized in that: S2 specifically includes the following steps: S21. Construct an interest point-category graph and achieve representation learning of interest points through continuous word skipping; To better represent user-transferable features, a pre-training process is performed before model inference to obtain POI representations. An attention network is constructed to extract user-transferable features by taking the representations of POIs that the user has checked in as input. Therefore, the quality of the input representations of POIs that the user has checked in determines the quality of user-transferable features. Because each POI is associated with multiple text descriptions, this information makes it relatively easy to construct an interest-category graph. Defined as ,in It is a set of POIs. It is a vocabulary list. It is a set of edges connecting words and POIs, and the representation of interest points is achieved through a continuous skipping word model; Using POIs as the central words, the following formula is used to train the representation of POIs. in, It is a word. It's a point of interest. It is the Softmax function; S22. Construct an attention network to obtain user-transferable features by taking the representation of POI as input; User transferable features are derived from each Point of Interest (POI) in the user's historical check-in record. Therefore, based on the representation of POIs trained using a continuous skipping model, user transferable features are extracted through an attention network. Firstly, each POI in the user's check-in history contributes differently to interest transfer because the attention mechanism compresses multiple POIs into a single representation, with different parts contributing differently. Therefore, a weighted sum is applied to each POI using the attention mechanism. in, Indicates user The representation of transferable features, Indicates POI The attention score is interpreted as The importance of user-transferable features; the attention score is obtained by designing an attention network; its formula is as follows: in, and These are the weight parameters of the attention network. It is a bias vector. It is the Sigmoid function; the attention network obtains user-transferable features as input to the learning network to generate the weight parameters of the personalized transfer network.

4. The method for recommending points of interest in different locations based on user preferences and personalized preference transfer according to claim 1, characterized in that: In S3, there is a specific relationship between users' destination preferences and their transferable features. A parameter learning network is designed based on a multilayer perceptron, which takes users' transferable features as input and then generates the weight parameters for the transfer network. The parameter learning network is represented as follows: in, and It is the weight matrix of the parameter learning network. and It is a bias vector. It is the Sigmoid function; The input to the transfer network is the user preference representation obtained by S1. The weight matrix of the transfer network is not randomly initialized but is the output of the parameter learning network. Because the parameter network has an output for each user, each user has its own independent weight matrix in the transfer network, thus completing the personalized interest transfer of the user. Similarly, the transfer network is completed through a multilayer perceptron, which will... Restructuring Therefore, the formula for the transfer network is as follows: in, It is the output of the parameter learning network. It is a transfer network.

5. The method for recommending points of interest in different locations based on user preferences and personalized preference transfer according to claim 1, characterized in that: The characterization of off-site POIs in S4 includes: The geographic location of POIs in the target area helps in understanding users' cross-regional check-in behavior; First, each POI in the target region is initialized based on the S2 method, denoted as... ,and in ,Notice Then, an undirected graph is constructed based on the geographical relationships between POIs. ,side The definition is as follows: in, Indicates POI and The distance is calculated based on latitude and longitude; an adjacency matrix is ​​constructed based on the distance between each pair of POIs. ; To capture the spatial relationships between POIs, a convolutional network is used: in, Represents the weight matrix. It is a bias term; It is an updated target area POI embedding matrix that encodes the geographic impact of POIs.

6. The method for recommending points of interest in different locations based on user preferences and personalized preference transfer according to claim 1, characterized in that: In S5, after the model parameters are optimized, the model is used to recommend POIs in the target area to users; For a user And his / her historical check-in records, generating preference representations after user interest shifts. Then use and The inner product yields the user's rating of the target area's Point of Interest (POI). Finally, based on the estimated scores, the top k target area points (POIs) were selected as the target for users. Recommended.

Images