A context-aware next point of interest recommendation method based on graph neural networks

Abstract

This application provides a context-aware next point of interest recommendation method based on graph neural networks, relating to the field of point of interest recommendation. The method includes: acquiring and preprocessing a user's check-in records to obtain a valid check-in dataset; constructing a multi-relational cross-sequence graph and a graph neural network; using the graph neural network to learn item representations based on the multi-relational cross-sequence graph to obtain a personalized embedding vector for each item in the check-in sequence in the valid check-in dataset; calculating the sequence representation of the check-in sequence using the personalized embedding vector for each item in the check-in sequence; constructing soft labels for the check-in sequence; predicting the probability distribution of items to be interacted with at the next timestamp in the current check-in sequence using the soft labels, the sequence representation of the check-in sequence, and the personalized embedding vector; and generating k points of interest that the user may visit at the next timestamp based on the probability distribution, thus completing the recommendation of the user's next point of interest.

Description

Technical Field

[0001] This application relates to the field of interest point recommendation, and more particularly to a context-aware next interest point recommendation method based on graph neural networks. Background Technology

[0002] With the rapid development of information technology, especially the widespread adoption of mobile internet and smartphones, people are increasingly reliant on digital platforms for social interaction and the recording and sharing of personal activities. Social media and location-based services (such as Foursquare, Gowalla, and Yelp) have become an indispensable part of modern life. These platforms not only allow users to easily discover new things around them but also enable them to share their experiences with friends and communities. Against this backdrop, next point of interest recommendation technology has become an important research area. The goal of this technology is to accurately predict the locations users are likely to visit next, in order to provide more personalized recommendation services.

[0003] Currently, scholars both domestically and internationally have conducted much valuable research on point-of-interest (POI) recommendation. Early research methods only utilized intra-sequence context, which can negatively impact performance when the check-in sequence is very sparse. Some existing methods leverage rich contextual information within the check-in sequence, typically including intra-sequence and cross-sequence information. The former can be further divided into item-level context and sequence-level context. Item-level context describes adjacent items in the check-in sequence, while sequence-level context refers to the complete sequence information. Similarly, the latter includes collaborative information from sequences with similar items and sequence behavior patterns. Recently, some methods have utilized intra-sequence and cross-sequence context to incorporate contextual information from the sequence to enrich the representation of a given drawing. In particular, some methods propose constructing a global graph based on intra-sequence and cross-sequence item-level context information to link items from various sequences, and then learning item / sequence embeddings based on this graph.

[0004] Although graph neural networks are very effective in modeling and computing the similarity between points of interest, there are three main limitations to using them in point of interest recommendation:

[0005] 1. When constructing the global graph of POI, it is impossible to distinguish the POI-POI edge type.

[0006] 2. Previous studies have learned a fixed embedding vector for each point of interest. However, since user interests may vary across different sequences, it is necessary to learn the embeddings of interest points to reflect changes in user interests and to personalize the data by referencing different sequences.

[0007] 3. The one-hot encoded vector of the target interest point is typically used as the hard label to be predicted, which may not reflect true user preferences. However, the true distribution of user preferences is usually unknown because only a limited number of check-in points are exposed to users. Simply treating the one-hot encoded vector of the target interest point as the true distribution can introduce bias and lead to overfitting.

[0008] Existing patent CN115964748A relates to a user interest point recommendation method, which uses soft tags to achieve the technical concept of interest point recommendation, but the application scenarios are different and it cannot solve the three limitations proposed in this application. Summary of the Invention

[0009] The purpose of this invention is to address the problem of low recommendation accuracy in existing point-of-interest (POI) recommendation methods by providing a context-aware next POI recommendation method based on graph neural networks.

[0010] The above-mentioned objective of this application is achieved through the following technical solution:

[0011] S1: Obtain user check-in records and preprocess them to obtain a valid check-in dataset; the check-in records include: user identification information, type of point of interest, geographical information of point of interest, and check-in time information;

[0012] S2: Construct a multi-relationship cross-sequence graph and a graph neural network;

[0013] S3: Based on the multi-relational cross-sequence graph, a graph neural network is used to learn the item representation and obtain the check-in sequences from the valid check-in dataset. Each item's personalized embedding vector;

[0014] S4: By check-in sequence For each item's personalized embedding vector, calculate the check-in sequence. Sequence representation ;

[0015] S5: Construct the check-in sequence Soft tags;

[0016] S6: Using soft tags and check-in sequences Using sequence representations and personalized embedding vectors, predict the current check-in sequence. The probability distribution of items that will interact with the next timestamp in the sequence;

[0017] S7: Based on the probability distribution, generate the list of places the user might visit at the next timestamp. Based on the user's points of interest, the system recommends the user's next point of interest.

[0018] Optionally, step S1 includes:

[0019] The preprocessing includes: removing check-in records with incomplete information; filtering out check-in records of users whose historical check-in count is less than a preset number; and deleting check-in records involving fewer than a preset number of visits.

[0020] Optionally, step S2 includes:

[0021] From the valid check-in dataset, obtain the user's check-in sequence; treat a user's visit to a point of interest as an item, and each visit is considered an independent item; a user's check-in sequence includes a preset number of items, and each item in a check-in sequence is an item; each user includes a preset number of check-in sequences;

[0022] Based on all check-in sequences The set of neighbor items is used to establish edge connections, defined as follows:

[0023]

[0024] in, express In sequence The set of neighboring items in; The range length of the left and right neighbor sets is a hyperparameter. Indicates in sequence Any item in the list; Represents a sequence middle Neighboring items within the specified range that meet the criteria; These represent the positions of the items in the corresponding sequences; Indicates in sequence Zhongyu Same items; Representation and sequence Other check-in sequences that are not the same; U represents all check-in sequences;

[0025] based on The set of neighboring items, links to the same item from different sequences, i.e., linked sequences. and sequence The same items in the middle;

[0026] Construct a multi-relation cross-sequence graph based on project-level context information and types of interest points. ,in Indicates inclusion The set of nodes for all items in the set. This represents the set of points of interest in all check-in sequences; Represents the edge set in a multi-relation cross sequence graph; using To represent edge types, edge types capture the type and type Contextual relationships between projects; types express Project category; type express Project categories; It represents any node in the set of all items in a multi-relational cross sequence graph; express Nodes in the neighbor item set; item-level context information includes: geographic information and time information;

[0027] For points of interest and points of interest The weights of the edges between them are as follows:

[0028]

[0029] in, It is a frequency counting function for all check-in sequences; Indicates edge connection; This represents hyperparameters.

[0030] Optionally, step S3 includes:

[0031] S31: Use an embedded lookup table to initialize items. Embedded vector ;

[0032] Using an attention mechanism based on a graph neural network, projects are generated. The general embedding vector:

[0033] Based on multi-relational cross sequence graphs, obtain item-level context information within and between sequences;

[0034] Due to each project Neighboring items have different levels of importance; this introduces project-level attention and learning items. Embedded vector representation;

[0035] In the graph convolutional neural network, the first Within the layer, items are obtained through domain aggregation. In the first Typical embedding vectors in a layer:

[0036]

[0037] in, Represents the self-attention score; Indicates trainable parameters;

[0038] Attention Score The calculation method is as follows:

[0039]

[0040] in, It is the LeakyReLU function. and These are trainable parameters, and This indicates the cascading operator; This represents the general embedding vector of layer k-1; express and Edge weights; Indicates the embedding of edge relationships; express and Edge weights; express and Edge relationship embedding; Representation Item The intersection of its neighboring item sets;

[0041] S32: Passing Item The sequence information, combined with the general embedding vector of the item, generates the item. Personalized embedding vectors;

[0042] Given a sign-in sequence and a project , Including not The project in, and All items reflect the user's interest in the current check-in sequence;

[0043] For the project Introduce a check-in sequence Perform personalized embedding vectors The details are as follows:

[0044] Add a link to the sequence Virtual nodes for all projects Virtual nodes Embedded vector Used to capture check-in sequences The sequence information of all items in the list;

[0045] Application gating mechanism to integrate and Generate personalized embedding vectors :

[0046]

[0047] in, This indicates that the k-th layer refers to the project. The sign-in sequence Generate personalized embedding vectors; Indicates project The general embedding vector at the k-th layer; Represents a virtual node at level k-1. Generate a sequence of check-ins to capture. The embedding vector of sequence information for all items in the sequence;

[0048] Gated score The calculation method is as follows:

[0049]

[0050] in, These are learnable parameters. It is the dimensionality coefficient;

[0051] The first Virtual nodes in layers Embedded vector Updated to:

[0052]

[0053] Among them, weight The calculation formula is: , These are learnable parameters.

[0054] Optionally, step S4 includes:

[0055] Determine the item's position and sequence length within the check-in sequence;

[0056] For check-in records, use a shared location embedded lookup table. , The first in OK Indicates length is The sequence of the first Embedded vectors at inverted positions;

[0057] Introduce a shared sequence length embedding lookup table , The Middle OK Indicates the length of the corresponding sequence Embedding; limiting the maximum length of the sequence to ;

[0058] For length of The first in the check-in sequence For each item, the item's location information and sequence length are unified to the item's general embedding vector. In the middle, the updated item is obtained. Embedded vector :

[0059]

[0060] in, Representation Item Reverse position The embedding vector;

[0061] For all projects, define a shared project type embedding lookup table, where each row represents the embedding vector of a project's type:

[0062]

[0063] in, Indicates project Type embedding;

[0064] Using attention mechanism to check-in sequence The information from all items in the middle is merged, and it has the following characteristics:

[0065]

[0066] in, This indicates the use of attention mechanisms to fuse sequences. The embedded vector representation of information for all items in the dataset; Indicates attention weight;

[0067] Attention weights are calculated using two layers of multilayer perceptrons. :

[0068]

[0069] here, Indicates the updated item The embedding vector; Indicates the sign-in sequence The embedding vector of the last term in ; It is a virtual node Embedding; This represents the embedding vector of a project's category attributes;

[0070] Will and Combined, in the check-in sequence Capture user interests and check-in sequences The sequence representation is expressed by the following formula:

[0071]

[0072] in, These are weight parameters.

[0073] Optionally, step S5 includes:

[0074] For a given check-in sequence From having Searching for sequences in a fixed-size pool of the nearest sequences. Most similar A sequence;

[0075] Using the SimHash method, the sign-in sequence is... Multiplying the sequence representation by a hash function, the given sequence The embeddings of the candidate sequences are projected onto the binary sequence, and the hash function is set to a fixed random projection matrix. ,in Similar The embedding vectors of all candidate sequences yield the same hash output;

[0076] Output similar Hamming distance between the embedding vectors of candidate sequences, from Retrieved from candidate sequences Most similar front 1 candidate sequence;

[0077] By the most similar The check-in sequence is constructed by weighting the one-hot encodings of the candidate sequences. Soft tags.

[0078] Optionally, step S6 includes:

[0079] Soft tags, check-in sequence Sequence representations and personalized embedding vectors are input to the prediction layer of a graph neural network, and inner product sums are applied. The function outputs the current sign-in sequence. The probability distribution of items that will interact in the next timestamp:

[0080]

[0081] in, Indicates the item in the next timestamp. The probability of interaction.

[0082] Optionally, the total loss function of the graph neural network consists of two components: hard labels based on the check-in sequence. Cross-entropy loss and soft labeling based on check-in sequences KL divergence loss KLD:

[0083]

[0084] Here, λ is a trade-off parameter used to control the relative weight of these two losses. Indicates the predicted value. This represents the actual value.

[0085] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform a context-aware next point of interest recommendation method based on a graph neural network.

[0086] A computer-readable storage medium storing instructions that, when executed, perform a context-aware next point of interest recommendation method based on a graph neural network.

[0087] The beneficial effects of the technical solution provided in this application are:

[0088] 1. Construction of a multi-relationship cross-sequence graph. Based on all check-in sequences... The method uses a set of neighboring items to establish these edge connections, linking items from different sequences, thus leveraging global-level item transition relationships. It further utilizes item types to differentiate between different item transitions. This addresses the problem that most previous graph-based methods only model item transition patterns as graphs within a single check-in sequence, neglecting global-level relationships between items across different sequences. It fully utilizes contextual information within sequences, especially cross-sequence contextual information, achieving state-of-the-art performance on three real-world datasets, demonstrating practical feasibility and application potential.

[0089] 2. Construction of Personalized Item Embedding Vectors Based on Check-in Sequences. In addition to learning general item embedding vectors using graph neural networks based on multi-relational cross-sequence graphs, this method utilizes a gating mechanism to generate personalized item representations to encode variations in user interests. The item embedding vectors are combined with the item's position information in the check-in sequence and the length of the check-in sequence to learn a sequence representation of the check-in sequence. This improves the accuracy of recommendation performance.

[0090] 3. Generating Soft Labels for User Preferences. To mitigate the bias introduced by hard labels (i.e., one-hot encoded vectors) of target items in the sequence, this method employs internal and external sequence-level context and proposes a label collaboration strategy. This strategy generates soft labels for the user preferences to be predicted by leveraging collaborative filtering with the most similar historical sequences to the current sequence. This addresses the problem of bias and overfitting that can arise from treating the one-hot encoded vectors of target interest points as the true distribution in current techniques. Attached Figure Description

[0091] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0092] Figure 1 This is a step diagram of an embodiment of this application;

[0093] Figure 2 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0094] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0095] Embodiments of this application provide a context-aware next point of interest recommendation method based on graph neural networks.

[0096] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a context-aware next interest point recommendation method based on graph neural networks in an embodiment of this application, including:

[0097] S1: Obtain user check-in records and preprocess them to obtain a valid check-in dataset; the check-in records include: user identification information, type of point of interest, geographical information of point of interest, and check-in time information;

[0098] S2: Construct a multi-relationship cross-sequence graph and a graph neural network;

[0099] S3: Based on the multi-relational cross-sequence graph, a graph neural network is used to learn the item representation and obtain the check-in sequences from the valid check-in dataset. Each item's personalized embedding vector;

[0100] S4: By check-in sequence For each item's personalized embedding vector, calculate the check-in sequence. Sequence representation ;

[0101] S5: Construct the check-in sequence Soft tags;

[0102] Specifically, most existing technologies use the one-hot encoded vector of the target item as a hard label for user preferences, but this may not reflect the user's true preferences. Furthermore, user preferences are influenced by different time periods and contexts, which may deviate from historical data over time. Therefore, to address this issue, this method proposes a label collaboration strategy utilizing sequence-level context, aiming to explicitly utilize the target item from the historical sequence most similar to the current sequence as collaborative label information.

[0103] S6: Using soft tags and check-in sequences Using sequence representations and personalized embedding vectors, predict the current check-in sequence. The probability distribution of items that will interact with the next timestamp in the sequence;

[0104] S7: Based on the probability distribution, generate the list of places the user might visit at the next timestamp. Based on the user's points of interest, the system recommends the user's next point of interest.

[0105] Specifically, compared with existing methods, this patented method proposes a novel context-aware point-of-interest (POI) recommendation method that fully utilizes the contextual information in the check-in sequence. Firstly, after obtaining the check-in dataset, rigorous data cleaning is performed, retaining only core fields such as user ID, POI ID, check-in timestamp, and latitude / longitude, while removing users with fewer than 10 check-ins and POIs with a monthly average access frequency of less than 10. Secondly, a multi-relationship cross-sequence graph is constructed, connecting items based on intra-sequence and cross-sequence item-level context, where the edge relationships are defined based on item category. Then, based on this multi-relationship cross-sequence graph, a graph neural network is used to learn a general item embedding vector. Furthermore, to encode changes in user interests, this method also learns personalized item representations with a gating mechanism. Next, the item embedding vector is combined with the item's position embedding in the check-in sequence and the length of the check-in sequence to learn the embedding vector representation of the check-in sequence. Additionally, to mitigate the bias caused by the one-hot encoded vector (i.e., hard label) of the target item in the sequence, this method employs internal and external sequence-level context and proposes a label collaboration strategy. This strategy utilizes collaborative filtering with historical sequences most similar to the current sequence to generate soft labels for user preferences to be predicted. Finally, these encoded vectors are fed into a decoder component, which uses a predictive model to infer the potential probability of a user visiting each point of interest based on historical check-in patterns, thereby generating a recommendation list of the most likely next points of interest.

[0106] Step S1 includes:

[0107] The preprocessing includes: removing check-in records with incomplete information; filtering out check-in records of users whose historical check-in count is less than a preset number; and deleting check-in records involving fewer than a preset number of visits.

[0108] Step S2 includes:

[0109] Specifically, most graph-based methods only model the item transition patterns within a single check-in sequence as a graph, ignoring the global-level relationships between items in different sequences. Therefore, this method establishes connections between different sequences to further utilize global-level item transition relationships.

[0110] From the valid check-in dataset, obtain the user's check-in sequence; treat a user's visit to a point of interest as an item, and each visit is considered an independent item; a user's check-in sequence includes a preset number of items, and each item in a check-in sequence is an item; each user includes a preset number of check-in sequences;

[0111] Based on all check-in sequences The set of neighbor items is used to establish edge connections, defined as follows:

[0112]

[0113] in, express In sequence The set of neighboring items in; The range length of the left and right neighbor sets is a hyperparameter. Indicates in sequence Any item in the list; Represents a sequence middle Neighboring items within the specified range that meet the criteria; These represent the positions of the items in the corresponding sequences; Indicates in sequence Zhongyu Same items; Representation and sequence Other check-in sequences that are not the same; U represents all check-in sequences;

[0114] based on The set of neighboring items, links to the same item from different sequences, i.e., linked sequences. and sequence The same items in the middle;

[0115] Construct a multi-relation cross-sequence graph based on project-level context information and types of interest points. ,in Indicates inclusion The set of nodes for all items in the set. This represents the set of points of interest in all check-in sequences; Represents the edge set in a multi-relation cross sequence graph; using To represent edge types, edge types capture the type and type Contextual relationships between projects; types express Project category; type express Project categories; It represents any node in the set of all items in a multi-relational cross sequence graph; express Nodes in the neighbor item set; item-level context information includes: geographic information and time information;

[0116] Specifically, item-level contextual information refers to additional information related to user check-in behavior, which helps to better understand user behavior and preferences. In the context of point-of-interest recommendation, item-level contextual information includes geographical and temporal information.

[0117] Specifically, by considering the type of project, the system can better understand user behavior patterns and make more accurate point-of-interest recommendations; these types include restaurants, supermarkets, attractions, hotels, stations, etc. Users repeatedly visit points of interest in different categories. and This method cannot simply assume that these two points of interest are related, as this could also indicate a shift in user interests. Furthermore, if a user repeatedly visits points of interest of the same category... and If so, these two points of interest are more likely to be highly correlated.

[0118] For points of interest and points of interest The weights of the edges between them are as follows:

[0119]

[0120] in, It is a frequency counting function for all check-in sequences; Indicates edge connection; This represents hyperparameters.

[0121] Specifically, to mitigate the dominance effect of frequently occurring projects, this method introduces a hyperparameter. Its default value is set to 0.75. To speed up model efficiency, for each item in the graph... This method only retains the top-weighted relation in each relation. 100 neighbors. To further simplify the graph, this method only retains the most common neighbors in the graph. Each context relationship. By using different representations for the node edge connections in the multi-relation cross sequence graph, when... At that time, a unified setting (That is, the categories of points of interest did not change before and after the visit); when At that time, a unified setting (That is, the categories of points of interest visited before and after have changed).

[0122] Step S3 includes:

[0123] S31: Use an embedded lookup table to initialize items. Embedded vector ;

[0124] Using an attention mechanism based on a graph neural network, projects are generated. The general embedding vector:

[0125] Based on multi-relational cross sequence graphs, obtain item-level context information within and between sequences;

[0126] Due to each project Neighboring items have different levels of importance; this introduces project-level attention and learning items. Embedded vector representation;

[0127] Specifically, for each item, there are various contextual relationships, so when calculating the attention score, it is necessary to distinguish the marginal relationships.

[0128] In the graph convolutional neural network, the first Within the layer, items are obtained through domain aggregation. In the first Typical embedding vectors in a layer:

[0129]

[0130] in, Represents the self-attention score; Indicates trainable parameters;

[0131] Attention Score The calculation method is as follows:

[0132]

[0133] in, It is the LeakyReLU function. and These are trainable parameters, and This indicates the cascading operator; This represents the general embedding vector of layer k-1; express and Edge weights; Indicates the embedding of edge relationships; express and Edge weights; express and Edge relationship embedding; Representation Item The intersection of its neighboring item sets;

[0134] S32: Passing Item The sequence information, combined with the general embedding vector of the item, generates the item. Personalized embedding vectors;

[0135] Specifically, Use project-level context information to reflect the project The general embedding vector is used, but an item is usually contained in various sequences, so the representation of an item can be further enriched by the sequence information.

[0136] Given a sign-in sequence and a project , Including not The project in, and All items reflect the user's interest in the current check-in sequence;

[0137] For the project Introduce a check-in sequence Perform personalized embedding vectors The details are as follows:

[0138] Add a link to the sequence Virtual nodes for all projects Virtual nodes Embedded vector Used to capture check-in sequences The sequence information of all items in the list;

[0139] Application gating mechanism to integrate and Generate personalized embedding vectors :

[0140]

[0141] in, This indicates that the k-th layer refers to the project. The sign-in sequence Generate personalized embedding vectors; Indicates project The general embedding vector at the k-th layer; Represents a virtual node at level k-1. Generate a sequence of check-ins to capture. The embedding vector of sequence information for all items in the sequence;

[0142] Gated score The calculation method is as follows:

[0143]

[0144] in, These are learnable parameters. It is the dimensionality coefficient;

[0145] Specifically, in this way, this method can provide items Based on the current check-in sequence Generate personalized embedding vectors .when When smaller, Approaching the project The general embedding vector representation; otherwise, This will better represent the current sequence. The information in the middle.

[0146] The first Virtual nodes in layers Embedded vector Updated to:

[0147]

[0148] Among them, weight The calculation formula is: , These are learnable parameters.

[0149] Step S4 includes:

[0150] Specifically, for a given check-in sequence Although virtual nodes The embedding includes It contains information about all items, but it ignores the item's position in the check-in sequence, so it cannot be simply used as a representation of the check-in sequence. In the previous step, for each item... This method has referenced the sequence. Their general embedding vectors were calculated respectively. and personalized embedding vectors For the sake of simplicity, this method will [describe the project]. The embedding vector is further represented as Next, we describe how to compute the sequence representation based on the item embedding vector representation. In order to utilize the item sequence in the check-in record, in addition to the item embedding, this method further incorporates the item's position information in the check-in sequence and the sequence length.

[0151] Determine the item's position and sequence length within the check-in sequence;

[0152] For check-in records, use a shared location embedded lookup table. , The first in OK Indicates length is The sequence of the first Embedded vectors at inverted positions;

[0153] Specifically, this method selects the reverse order for location information because the most recent item may be more useful for predicting the next item in the sequence.

[0154] Introduce a shared sequence length embedding lookup table , The Middle OK Indicates the length of the corresponding sequence Embedding; limiting the maximum length of the sequence to ;

[0155] For length of The first in the check-in sequence For each item, the item's location information and sequence length are unified to the item's general embedding vector. In the middle, the updated item is obtained. Embedded vector :

[0156]

[0157] in, Representation Item Reverse position The embedding vector;

[0158] For all projects, define a shared project type embedding lookup table, where each row represents the embedding vector of a project's type:

[0159]

[0160] in, Indicates project Type embedding;

[0161] Using attention mechanism to check-in sequence The information from all items in the middle is merged, and it has the following characteristics:

[0162]

[0163] in, This indicates the use of attention mechanisms to fuse sequences. The embedded vector representation of information for all items in the dataset; Indicates attention weight;

[0164] Attention weights are calculated using two layers of multilayer perceptrons. :

[0165]

[0166] here, Indicates the updated item The embedding vector; Indicates the sign-in sequence The embedding vector of the last term in ; It is a virtual node Embedding; This represents the embedding vector of a project's category attributes;

[0167] Specifically, this method also uses a check-in sequence. The embedding vector of the last term in Because it is likely to be highly correlated with the prediction of the next item.

[0168] Will and Combined, in the check-in sequence Capture user interests and check-in sequences The sequence representation is expressed by the following formula:

[0169]

[0170] in, These are weight parameters.

[0171] Step S5 includes:

[0172] Specifically, given a check-in sequence The objective of this application is to be the first to obtain from those with Searching for sequences in a fixed-size pool of the nearest sequences. Most similar The more sequences retrieved by this method, the more accurate the estimation of user preferences becomes, but the greater the computational cost. The SimHash method is used to accelerate model efficiency. The SimHash function takes the sequence representation as input and generates its binary sequence, where each entry is either 0 or 1. Previous research has shown that the output of SimHash satisfies the locality-sensitive property, meaning that if the input vectors are similar to each other, the outputs will also be similar.

[0173] For a given check-in sequence From having Searching for sequences in a fixed-size pool of the nearest sequences. Most similar A sequence;

[0174] Using the SimHash method, the sign-in sequence is... Multiplying the sequence representation by a hash function, the given sequence The embeddings of the candidate sequences are projected onto the binary sequence, and the hash function is set to a fixed random projection matrix. ,in Similar The embedding vectors of all candidate sequences yield the same hash output;

[0175] Output similar Hamming distance between the embedding vectors of candidate sequences, from Retrieved from candidate sequences Most similar front 1 candidate sequence;

[0176] Specifically, after searching After finding the most similar sequences, the next step is to construct sequences. Soft tags. (Compared to using only...) In comparison, these A sequence can help provide more than just using More comprehensive user interest prediction.

[0177] By the most similar The check-in sequence is constructed by weighting the one-hot encodings of the candidate sequences. Soft tags.

[0178] Step S6 includes:

[0179] Specifically, the prediction layer of the graph neural network outputs the probability distribution of the user's next point of interest in the current sequence. Due to the long-tail distribution problem in recommendation data, this method normalizes the personalized embedding vectors of items and check-in sequences in each layer. Sequence representation .

[0180] Soft tags, check-in sequence Sequence representations and personalized embedding vectors are input to the prediction layer of a graph neural network, and inner product sums are applied. The function outputs the current sign-in sequence. The probability distribution of items that will interact in the next timestamp:

[0181]

[0182] in, Indicates the item in the next timestamp. The probability of interaction.

[0183] The total loss function of the graph neural network consists of two components: hard labels based on the check-in sequence. Cross-entropy loss and soft labeling based on check-in sequences KL divergence loss KLD:

[0184]

[0185] Here, λ is a trade-off parameter used to control the relative weight of these two losses. Indicates the predicted value. This represents the actual value.

[0186] This application also discloses an electronic device. (See reference...) Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0187] The communication bus 502 is used to enable communication between these components.

[0188] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0189] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0190] This application also discloses a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the above-described context-aware next interest point recommendation method based on a graph neural network.

[0191] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will readily conceive of those skilled in the art upon consideration of the specification and the disclosure of practical truths.

[0192] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A context-aware next point of interest recommendation method based on graph neural networks, characterized in that, The method includes the following steps: S1: Obtain user check-in records and preprocess them to obtain a valid check-in dataset; the check-in records include: user identification information, type of point of interest, geographical information of point of interest, and check-in time information; S2: Construct a multi-relationship cross-sequence graph and a graph neural network; Step S2 includes: From the valid check-in dataset, obtain the user's check-in sequence; treat a user's visit to a point of interest as an item, and each visit is considered an independent item; a user's check-in sequence includes a preset number of items, and each item in a check-in sequence is an item; each user includes a preset number of check-in sequences; Based on all the check-in sequences The neighbor item set, edge connection is established, defined as follows: in, express In sequence The set of neighboring items in; The range length of the left and right neighbor sets is a hyperparameter. Indicates in sequence Any item in the list; Represents a sequence middle Neighboring items within the specified range that meet the criteria; These represent the positions of the items in the corresponding sequences; Indicates in sequence Zhongyu Same items; Representation and sequence Other check-in sequences that are not the same; U represents all check-in sequences; based on The set of neighboring items, links to the same item from different sequences, i.e., linked sequences. and sequence The same items in the middle; Construct a multi-relation cross-sequence graph based on project-level context information and interest type. ,in Indicates inclusion The set of nodes for all items in the set. This represents the set of points of interest in all check-in sequences; Represents the edge set in a multi-relation cross sequence graph; using To represent edge types, edge types capture the type and type Contextual relationships between projects; types express Project category; type express Project categories; It represents any node in the set of all items in a multi-relational cross sequence graph; express Nodes in the neighbor item set; item-level context information includes: geographic information and time information; For points of interest and points of interest The weights of the edges between them are as follows: in, It is a frequency counting function for all check-in sequences; Indicates edge connection; Indicates hyperparameters; S3: Based on the multi-relational cross-sequence graph, a graph neural network is used to learn the item representation and obtain the check-in sequences from the valid check-in dataset. Each item's personalized embedding vector; S4: By check-in sequence For each item's personalized embedding vector, calculate the check-in sequence. Sequence representation ; S5: Construct the check-in sequence Soft tags; S6: Using soft tags and check-in sequences Using sequence representations and personalized embedding vectors, predict the current check-in sequence. The probability distribution of items that will interact with the next timestamp in the sequence; S7: Based on the probability distribution, generate k points of interest that the user may visit at the next timestamp, and complete the recommendation of the user's next point of interest.

2. The context-aware next point of interest recommendation method based on graph neural networks as described in claim 1, characterized in that, Step S1 includes: The preprocessing includes: removing check-in records with incomplete information; filtering out check-in records of users whose historical check-in count is less than a preset number; and deleting check-in records involving fewer than a preset number of visits.

3. The context-aware next interest point recommendation method based on graph neural networks as described in claim 1, characterized in that, Step S3 includes: S31: Use an embedded lookup table to initialize items. Embedded vector ; Using an attention mechanism based on a graph neural network, projects are generated. The general embedding vector: Based on the multi-relational cross sequence graph, obtain item-level context information within and between sequences; Due to each project Neighboring items have different levels of importance; this introduces project-level attention and learning items. Embedded vector representation; In the graph convolutional neural network, the first Within the layer, items are obtained through domain aggregation. In the first Typical embedding vectors in a layer: in, Represents the self-attention score; Indicates trainable parameters; Attention Score The calculation method is as follows: in, It is the LeakyReLU function. and These are trainable parameters, and Indicates the cascading operator; This represents the general embedding vector of layer k-1; express and Edge weights; Indicates the embedding of edge relationships; express and Edge weights; express and Edge relationship embedding; Representation Item The intersection of its neighboring sets; S32: Passing item The sequence information, combined with the general embedding vector of the item, generates the item. Personalized embedding vectors; Given a sign-in sequence and a project , Including not The project in, and All items reflect the user's interest in the current check-in sequence; For the project Introduce a check-in sequence Perform personalized embedding vectors The details are as follows: Add a link to the sequence Virtual nodes for all projects Virtual nodes Embedded vector Used to capture check-in sequences The sequence information of all items in the list; Application gating mechanism to integrate and Generate personalized embedding vectors : in, This indicates that the k-th layer refers to the project. The sign-in sequence Generate personalized embedding vectors; Indicates project The general embedding vector at the k-th layer; Represents a virtual node at level k-1. Generate a sequence of check-ins to capture. The embedding vector of sequence information for all items in the sequence; Gated score The calculation method is as follows: in, These are learnable parameters. It is the dimensionality coefficient; The first Virtual nodes in layers Embedded vector Updated to: Among them, weight The calculation formula is: , These are learnable parameters.

4. The context-aware next point of interest recommendation method based on graph neural networks as described in claim 3, characterized in that, Step S4 includes: Determine the item's position and sequence length within the check-in sequence; For check-in records, use a shared location embedded lookup table. , The first in OK Indicates length is The sequence of the first Embedded vectors at inverted positions; Introduce a shared sequence length embedding lookup table , The Middle OK Indicates the length of the corresponding sequence Embedding; limiting the maximum length of the sequence to ; For length of The first in the check-in sequence For each item, the item's location information and sequence length are unified to the item's general embedding vector. In the middle, the updated item is obtained. Embedded vector : in, Representation Item Reverse position The embedding vector; For all projects, define a shared project type embedding lookup table, where each row represents the embedding vector of a project's type: in, Indicates project Type embedding; Using attention mechanism to check-in sequence The information from all items in the middle is merged, and it has the following characteristics: in, This indicates the use of attention mechanisms to fuse sequences. The embedded vector representation of information for all items in the dataset; Indicates attention weight; Attention weights are calculated using two layers of multilayer perceptrons. : here, Indicates the updated item The embedding vector; Indicates the sign-in sequence The embedding vector of the last term in ; It is a virtual node Embedding; This represents the embedding vector of a project's category attribute; Will and Combined, in the check-in sequence Capture user interests and check-in sequences The sequence representation is expressed by the following formula: in, These are weight parameters.

5. The context-aware next point of interest recommendation method based on graph neural networks as described in claim 4, characterized in that, Step S5 includes: For a given check-in sequence From having Searching for sequences in a fixed-size pool of the nearest sequences. Most similar A sequence; Using the SimHash method, the sign-in sequence is... Multiplying the sequence representation by a hash function, the given sequence The embeddings of the candidate sequences are projected onto the binary sequence, and the hash function is set to a fixed random projection matrix. ,in Similar The embedding vectors of all candidate sequences yield the same hash output; Similar outputs Hamming distance between the embedding vectors of candidate sequences, from Retrieved from candidate sequences Most similar front 1 candidate sequence; By the most similar The check-in sequence is constructed by weighting the one-hot encodings of the candidate sequences. Soft tags.

6. The context-aware next point of interest recommendation method based on graph neural networks as described in claim 5, characterized in that, Step S6 includes: Soft tags, check-in sequence The sequence representation and personalized embedding vectors are input to the prediction layer of the graph neural network, and the inner product and sum are applied. The function outputs the current sign-in sequence. The probability distribution of items that will interact in the next timestamp: in, Indicates the item in the next timestamp. The probability of interaction.

7. The context-aware next point of interest recommendation method based on graph neural networks as described in claim 1, characterized in that, The total loss function of the graph neural network consists of two components: hard labels based on the check-in sequence. Cross-entropy loss and soft labeling based on check-in sequences KL divergence loss KLD: Here, λ is a trade-off parameter used to control the relative weight of these two losses. Indicates the predicted value. This represents the actual value.

8. An electronic device, characterized in that, The device includes a processor (501), a memory (505), a user interface (503), and a network interface (504). The memory (505) is used to store instructions. The user interface (503) and the network interface (504) are used to communicate with other devices. The processor (501) is used to execute the instructions stored in the memory (505) to cause the electronic device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the method as described in any one of claims 1-7.

Images