A continuous point-of-interest recommendation method based on social spatio-temporal information and user preferences

Abstract

The application discloses a continuous interest point recommendation method and system based on social space-time information and user preferences. The method comprises the following steps: obtaining long-term preference features and short-term preference features of a user in a user check-in sequence; obtaining geographical space features of each location in the user check-in sequence; obtaining user social information features of each location in the user check-in sequence; adding and fusing the long-term preference features, the short-term preference features, the geographical space features and the user social information features into multi-semantics features, and inputting the multi-semantics features into a recommendation model to generate an interest point recommendation list meeting user preferences. The technical problems of check-in data sparsity, insufficient utilization of meta-information and inaccurate recommendation for cold users are solved, hierarchical feature extraction is realized for the user check-in sequence, the geographical space features are increased, the user social information features are enhanced, the behavior characteristics of the user are comprehensively learned, the next interest point is accurately recommended, and the use experience of the user is optimized.

Description

Technical Field

[0001] This invention belongs to the field of data mining and recommendation system technology, and more specifically, relates to a method and system for recommending continuous points of interest based on social spatiotemporal information and user preferences. Background Technology

[0002] The internet has developed for over half a century, reaching an unprecedented scale. This massive amount of data increases the time users spend acquiring information of interest, leading to information overload. To improve information utilization and enable businesses to provide more effective personalized recommendations, recommendation system technology has emerged, achieving a win-win situation for both users and businesses.

[0003] In location-based social networks (LBSNs), a point of interest (POI) represents a real location, such as a shopping mall or school. Existing technologies typically employ continuous POI recommendation, which possesses strong temporal and spatial properties. This means that based on a user's past check-in records, it predicts the user's possible choices in the current state. Furthermore, since check-in requires the user to physically visit the location, different distances also influence the user's choice, giving continuous POI recommendation its spatiotemporal characteristics.

[0004] Currently, recommendations for continuous points of interest still face the following problems: the sparsity of check-in data due to the increase in the number of users and points of interest as the scale of the LBSN expands; the omission of information due to only considering one or two types of meta-information and ignoring other meta-information, resulting in insufficient breadth of meta-information utilization; and the inaccuracy of recommendations for cold users when there are no check-in records of social network data for a new user who has recently joined the LBSN. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a continuous interest point recommendation method and system based on social spatiotemporal information and user preferences. Its purpose is to solve the technical problems of sparsity of check-in data, insufficient breadth of meta-information utilization, and inaccurate recommendations for cold users.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for recommending continuous points of interest based on social spatiotemporal information and user preferences is provided, comprising:

[0007] Obtain the long-term and short-term preference characteristics of users in the user check-in sequence;

[0008] Collect the latitude and longitude information of all points of interest in each user's check-in sequence dataset, and use the haversine function to calculate the actual physical distance between every two points of interest;

[0009] Set a distance threshold, construct an adjacency matrix based on the actual geographical distances, and obtain the neighbor set of each point of interest through the adjacency matrix;

[0010] Iterate through the neighbor set of each point of interest, calculate the attention coefficient and distance coefficient between each point of interest and its neighboring nodes, and aggregate them into the geospatial features of each point of interest.

[0011] Obtain the user social information features at each position in the user check-in sequence;

[0012] The long-term preference features, short-term preference features, geospatial features, and user social information features are added and fused into multi-semantic features, which are then input into the recommendation model to generate a list of interest points that match the user's preferences.

[0013] Optionally, obtaining the long-term and short-term preference characteristics of users in the user check-in sequence includes:

[0014] The input user check-in sequence is feature extracted using a multi-head attention mechanism, and the user's long-term preference features are learned from multiple subspaces.

[0015] The user check-in sequence is processed by a recurrent neural network (RNN) to extract features, model the user's recent check-in behavior, and learn the user's short-term preference features.

[0016] Optionally, the step of traversing the neighbor set of each point of interest, calculating the attention coefficient and distance coefficient between each point of interest and its neighboring nodes, and aggregating the corresponding geospatial features of each point of interest includes:

[0017] Obtain the initial features of the target interest point and the set of its neighbors;

[0018] One point of interest is obtained sequentially from the neighbor set, and the attention coefficient between the obtained point of interest and the target point of interest is calculated sequentially: Among them, s ij For the attention scoring mechanism, a scoring function is used to calculate the correlation between two points, resulting in a score, which is then normalized using softmax; a(·,·) represents the similarity calculation function, W represents the trainable parameter matrix, and h j It is a feature of node j; This represents the set of neighboring interest points of interest point i;

[0019] Calculate the distance coefficient between the acquired points of interest and the target points of interest sequentially: Where d(i,j) represents the distance between two points of interest;

[0020] Update the features of the target interest point based on the attention coefficient and distance coefficient obtained in each calculation:

[0021] After traversing the neighbor set, a geospatial feature vector of the target point of interest is constructed based on its location information; and geospatial features of each point of interest are constructed.

[0022] Optionally, obtaining the user social information features at each position in the user check-in sequence includes:

[0023] Construct a social network graph based on the preference relationships between different users' interest information;

[0024] The similarity matrix of users is obtained through the social network graph;

[0025] The corresponding set of neighboring users is obtained based on the similarity matrix, and features are extracted from the set of neighboring users using a random neighbor sampling algorithm.

[0026] The extracted features are added to the target users to form the user's social information features.

[0027] Optionally, the preference association is represented using Jaccard-like similarity;

[0028] The formula for the similarity between the Jaccard class is: If the class-Jaccard similarity is greater than the user similarity threshold δ, then it is determined that there is a preference association between the two corresponding users.

[0029] Optionally, the step of adding and fusing the long-term preference features, the short-term preference features, the geospatial features, and the user's social information features into a multi-semantic feature, and inputting it into the recommendation model to generate an interest point recommendation list that matches the user's preferences, includes:

[0030] The feature vector information of the long-term preference features, the short-term preference features, the geospatial features, and the user's social information features are concatenated to generate point of interest metadata;

[0031] The probability distribution of the candidate set for the next point of interest is calculated based on the aforementioned interest point metadata.

[0032] Select the top k items with the highest probability from the candidate set as the user's interest point recommendation list.

[0033] Optionally, the candidate set of points of interest includes a category candidate set and a location candidate set.

[0034] Optionally, before obtaining the long-term and short-term preference characteristics of users in the user check-in sequence, the method further includes:

[0035] The scalar information of multimodal features in the user check-in sequence is transformed into low-dimensional dense multimodal feature vector information.

[0036] According to another aspect of the present invention, a continuous point-of-interest recommendation system based on social spatiotemporal information and user preferences is provided, comprising:

[0037] The user preference tracking module uses recurrent neural networks (RNNs) and multi-head attention mechanisms to analyze and process the user's historical check-in sequence, and obtain the user's long-term and short-term preference features in the user check-in sequence;

[0038] The geographic distance perception module uses a graph neural network (GNN) to obtain the geographic spatial features of each location in the user check-in sequence;

[0039] The social information influence module uses a random neighbor sampling algorithm to extract features from the user's social network graph and obtain the user's social information features at each position in the user check-in sequence;

[0040] The candidate set prediction module is used to add and fuse the feature information obtained by the user preference tracking module, the geographic distance perception module and the social information influence module into multi-semantic features, and input them into the recommendation model to generate a list of interest points that match user preferences.

[0041] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0042] 1. Hierarchical feature extraction was performed on the user check-in sequence to comprehensively learn the user's behavioral characteristics; combined with multiple characteristic information, multi-semantic features were formed to enrich and enhance the ability of features to express contextual semantic information, solve the problem of sparsity of check-in data and insufficient breadth of meta-information utilization, and make it more conducive for the model to learn the distribution characteristics of users' interests, preferences, social information and other characteristics.

[0043] 2. By constructing a neighbor set for each user's point of interest, the geospatial features of the point of interest, i.e., spatiotemporal state information, can be obtained. Adding a distance attention coefficient when constructing the neighbor set can make the distance relationship between different points of interest more significant and more in line with the spatial semantics of point of interest recommendation.

[0044] 3. When constructing user social information features, the random neighbor sampling algorithm is used to construct features. The random sampling method can simulate the propagation of user preferences in the social network, and at the same time absorb the preference features of similar users, thereby enhancing the user's feature expression; the established social network graph and the random neighbor sampling algorithm can solve the problem of cold user initiation recommendation.

[0045] 4. By integrating numerous influencing factors and features into a single multi-semantic feature, the recommendation model can utilize the information from each feature component to enable continuous interest point recommendations to possess spatiotemporal characteristics, thereby improving recommendation performance. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating a continuous interest point recommendation method based on social spatiotemporal information and user preferences disclosed in Embodiment 1 of the present invention.

[0047] Figure 2 This is an example diagram illustrating the composition of the user check-in sequence structure disclosed in Embodiment 1 of the present invention;

[0048] Figure 3 This is an example diagram of the distance relationship between points of interest disclosed in Embodiment 1 of the present invention;

[0049] Figure 4 This is an example diagram of the random sampling strategy disclosed in Embodiment 1 of the present invention;

[0050] Figure 5 This is a schematic diagram of the structure of a continuous interest point recommendation system based on social spatiotemporal information and user preferences disclosed in Embodiment 2 of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0052] Example 1

[0053] like Figure 1 As shown, a continuous interest point recommendation method based on social spatiotemporal information and user preferences includes:

[0054] S100: Obtain the long-term and short-term preference characteristics of users in the user check-in sequence;

[0055] S200: The latitude and longitude information of all points of interest in each user's check-in sequence dataset was statistically analyzed, and the actual physical distance between each pair of points of interest was calculated using the haversine function.

[0056] S300. Set a distance threshold, construct an adjacency matrix based on the actual geographical distance relationship between each point of interest, and obtain the neighbor set of each point of interest through the adjacency matrix;

[0057] S400. Traverse the neighbor set of each point of interest, calculate the attention coefficient and distance coefficient between each point of interest and its neighbor nodes, and aggregate them into the geospatial features of the target point of interest.

[0058] S500: Obtain the social information features of users at each position in the user check-in sequence;

[0059] S600: Long-term preference features, short-term preference features, geospatial features, and user social information features are added and fused into multi-semantic features, which are then input into the recommendation model to generate a list of interest points that match user preferences.

[0060] Optionally, prior to S100, it also includes:

[0061] The scalar information of multimodal features in the user check-in sequence is transformed into low-dimensional dense multimodal feature vector information.

[0062] like Figure 2 As shown, a user check-in sequence is constructed, and a single check-in record can be represented as... The mathematical form of a user check-in record of length n can be represented as:

[0063]

[0064] in, Indicates user u i The j-th record in the check-in sequence, u i l j t τ c k These represent the IDs for the user, POI, time range, and POI category, respectively, and the timestamp t. τ Strictly sorted from smallest to largest.

[0065] The user check-in sequence contains multimodal scalar information, which needs to be processed into vector information. The multimodal feature embedding and aggregation module maps these multimodal scalar identification information into low-dimensional dense vectors through an embedding layer. Specifically, a two-dimensional embedding matrix is ​​created for each modality. For example, the embedding matrices for user, interest point, time interval, and category information are respectively: M U MP M T M C Then, based on the corresponding modal scalar information, the low-dimensional dense vector representation is obtained from the subscripts corresponding to the embedding matrix:

[0066]

[0067] Among them, u i l j t τ c k These represent the IDs of the user, POI, time interval, and POI category, respectively. Embedding(·) represents the embedding layer, which embeds the corresponding matrix information and obtains the corresponding feature vector through a scalar. These represent feature vectors for users, points of interest (POIs), time intervals, and POI categories, respectively, where d represents the dimension of the feature vector. The multimodal feature embedding aggregation module specifically includes: 1) a user ID embedding layer, used to process user ID scalars into feature vectors; 2) a POI ID embedding layer, used to process POI ID scalars into feature vectors; 3) a timestamp interval ID embedding layer, used to process timestamp interval ID scalars into feature vectors; and 4) a POI category ID embedding layer, used to process POI category ID scalars into feature vectors.

[0068] After transforming the multimodal scalar information in the user check-in sequence into low-dimensional dense multimodal feature vector information, hierarchical feature extraction was performed on the long-term preference features, short-term preference features, geospatial features, and user social information features in the user check-in sequence. This facilitates a more comprehensive learning of user behavior characteristics, and statistical analysis provides evidence for the hierarchical approach.

[0069] Then, long-term preference features, short-term preference features, geospatial features, and user social information features are added and fused into multi-semantic features. These features are then concatenated according to certain rules to achieve multi-modal feature embedding and aggregation, while enriching user personal information and increasing the expressive power of the feature vectors. These multi-semantic features are then input into the recommendation model to generate a list of interest-based recommendations that match user preferences.

[0070] Optionally, S100 specifically includes:

[0071] S110. Feature extraction is performed on the input user check-in sequence through a multi-head attention mechanism to learn the user's long-term preference features from multiple subspaces;

[0072] S120. Features are extracted from the input user check-in sequence using a recurrent neural network (RNN), with a focus on modeling the user's recent check-in behavior to learn the user's short-term preference features.

[0073] User check-in records exhibit a hierarchical structure, reflecting both long-term and short-term changes in user preferences. For instance, user check-in behavior often exhibits periodicity; users are likely to do the same thing within similar timeframes. More recent records reflect recent check-in behavior, from which short-term preferences can be analyzed; while older records reflect potential check-in intentions, allowing for the analysis of long-term preferences.

[0074] For user u i Long-term historical check-in sequence of length l The data is processed through a multimodal feature embedding aggregation module to obtain a multimodal embedding representation of the long-term historical sequence. Then, a multi-head self-attention mechanism is used for modeling, as shown below:

[0075]

[0076]

[0077]

[0078] in, This represents the latent representation corresponding to the long-term historical check-in sequence embedded by the multimodal feature aggregation module, where |l| represents the length of the long-term check-in sequence. Let represent the long-term historical features learned through the self-attention mechanism in the j-th subspace, and h represent the number of subspaces in the multi-head attention mechanism. Indicates a splicing symbol. The comprehensive characteristics representing the long-term historical features of h subspaces.

[0079] A multi-head attention mechanism is employed to extract features from the input long-term check-in sequence. By learning long-term user preference features from multiple subspaces, the extracted long-term features do not significantly change over time, preserving the user's long-term potential check-in behavior characteristics. Furthermore, the attention mechanism can uncover relationships between non-adjacent records in the check-in sequence, while the multi-head mechanism can model the user's long-term features from different perspectives, thus achieving better extraction results.

[0080] For user u i Short-term current sign-in sequence of length s By processing the multimodal feature embedding aggregation module, a multimodal embedding representation of the short-term current sequence can be obtained. In this embodiment, the GRU model from the recurrent neural network (RNN) is used for modeling, as shown below:

[0081]

[0082]

[0083] in, This indicates that the current short-term check-in sequence is embedded into the latent representation corresponding to the aggregation module through multimodal feature embedding, |s| represents the length of the short-term check-in sequence, GRU(·) represents the GRU model, and P s This represents the final short-term current preference feature vector. and These represent the initial and final hidden states, respectively.

[0084] The Gated Recurrent Unit (GRU) is used to model and extract features from the input recent check-in sequence. The GRU can discover the periodic behavior of users in the check-in records, thereby modeling the user's check-in preferences in a recent period of time, learning the user's short-term preference features, which have significant periodicity and timeliness, and reflect the characteristics of the user's check-in behavior in a recent period of time.

[0085] For convenience, both long-term and short-term sequences are taken to have the same length, i.e., |l|=|s|=n. The next step is to process the input sequence... By capturing long-term historical characteristics and short-term preference characteristics, their long-term characteristics are obtained respectively. and short-term characteristics And these are then integrated into the final preference feature representation.

[0086] In this embodiment, by performing hierarchical modeling and feature extraction on the user's long-term check-in sequence and the user's short-term check-in sequence respectively, it is possible to learn the user's hidden preference features in the check-in records more effectively, learn the user's behavioral characteristics more comprehensively, and help the model better capture feature information.

[0087] Obtain the geospatial features of each location in the user check-in sequence, specifically including three steps: S200, S300, and S400.

[0088] Users' choice of their next check-in location is influenced by its distance; they tend to choose closer locations over more distant ones. Distance information needs to be represented by the distance between each different point of interest, such as... Figure 3 As shown, node 'a' represents the current location, node 'b' represents the location the user is most likely to go to, and node 'c' represents the location the user is less likely to go to. The locations the user is likely to go to are related to distance, such as... Figure 3 As shown, within a circular area A of a certain radius centered at node 'a', all points of interest are points of interest to be selected next. When studying the geospatial characteristics of a user's points of interest, it is necessary to construct an adjacency graph of those points.

[0089] A Point-of-Interest (POI) adjacency graph, also known as a "POI-POI" adjacency graph, describes the distance relationships between two different POIs.

[0090] G g = <V g E g A g >

[0091] Among them, V g E represents the POI node information in the adjacency graph. g This represents the edge information between different POI nodes. The edge relationships between POI nodes are derived from their distance relationships. First, the latitude and longitude information of all POIs in each dataset is collected. Then, the actual physical distance between each pair of POIs is calculated based on their latitude and longitude information using the haversine function. A distance threshold d is set. δ When the actual geographic distance between two POIs is less than or equal to d δ If the condition indicates a relationship between two POIs, then an edge is added to the two POI nodes. For example, POI node l i and POI node l j The distance between them is less than d δ Then in graph G g The set of edges E g A new bidirectional edge ε has been added. ij ∈E g A g This represents the feature set of each node, i.e. With V g The POI nodes in the table correspond one-to-one, where a j = <l j ,c k >,l j and c k Indicates the location and its category.

[0092] Based on the distance to the target interest point i being less than d δ Constructing the adjacency matrix W from points of interest g Then, based on the adjacency matrix W g Construct the neighbor set for each POI node i A distance metric, similarity β, is introduced to calculate the similarity between two points of interest (POIs). The distance similarity decreases as the distance increases. The neighbor set of the target POI is traversed, and the attention coefficient and distance coefficient between the target POI and its neighbor nodes are calculated, aggregating them into the geospatial features of the target POI. After traversing all POIs, the geospatial features of each POI are constructed.

[0093] Optionally, the S400 specifically includes:

[0094] S410. Obtain the initial features of the target interest point and the neighbor set of the target interest point;

[0095] S420. Obtain one interest point from the neighbor set in sequence, and calculate the attention coefficient between the obtained interest points and the target interest point in sequence: Among them, s ij For the attention scoring mechanism, a scoring function is used to calculate the correlation between two points, resulting in a score, which is then normalized using softmax; a(·,·) represents the similarity calculation function, W represents the trainable parameter matrix, and h j It is a feature of node j; This represents the set of neighboring interest points of interest point i;

[0096] S430. Calculate the distance coefficients between the acquired points of interest and the target points of interest sequentially: Where d(i,j) represents the distance between two points of interest;

[0097] S440. Update the features of the target interest point based on the attention coefficient and the distance coefficient obtained in each calculation:

[0098] S450. After traversing the neighbor set, construct the geospatial feature vector of the target interest point based on its location information; and construct the geospatial features of each interest point.

[0099] Where, α ij β represents the normalized attention coefficient between two POI nodes, while β ij This represents the normalized distance coefficient between two POI nodes. As the distance increases, β... ij This will decrease. a(·,·) represents the similarity calculation function, and d(i,j) represents the distance between two POI nodes. This represents the set of neighboring POI nodes of POI node i. Finally, by unifying the attention coefficient and distance coefficient, they are aggregated into the geospatial feature vector of the current POI.

[0100] Specifically, after traversing the information of each neighbor node, the geospatial feature vector is updated once. After traversing the information of all neighbor nodes of the target point of interest, the final geospatial feature vector h′ is obtained. i .

[0101] An adjacency graph is constructed based on the physical distance relationships between different points of interest (POIs). A graph attention network is then used to model the relationships between each POI and its neighbors by obtaining node attention coefficients. Furthermore, distance attention coefficients are obtained based on the distance between different POIs. These node attention coefficients and distance attention coefficients are used together to model the spatial features of each POI in the check-in sequence. In traditional graph attention networks, the neighborhood set does not contain specific distance relationships but is only an abstract representation of neighbors. Adding distance attention coefficients makes the distance relationships between different POIs more significant and better aligns with the spatial semantics of POI recommendation.

[0102] Optionally, the S500 specifically includes:

[0103] S510. Construct a social network graph based on the preference relationships between different users' interest information;

[0104] S520. Obtain the user similarity matrix through the social network graph;

[0105] S530. Obtain the corresponding set of neighboring users based on the similarity matrix, and extract features from the set of neighboring users using a random neighbor sampling algorithm;

[0106] S540. Add the extracted features to the target user to form the user's social information features.

[0107] A user's interests and preferences may spread within their social circle of friends, and users are also influenced by the preferences of their neighbors within the social network. A user's social network graph describes the preference relationships between different users. Since a user's interests can be influenced by the preferences of their neighbors within the social network, it is necessary to construct the social network graph relationships between different users.

[0108] G u = <V u E u A u >

[0109] Among them, V u E represents the user node information in the adjacency graph. u This represents the edge information between different user nodes. Represents the feature set of a user node. Represents the total number of users, where u i This is the user's ID.

[0110] In a user social network graph, if there is a preference relationship between two users, a social network graph can be constructed. The preference relationship is represented by Jaccard-like similarity. If the Jaccard-like similarity is greater than or equal to a set similarity threshold s, the relationship is considered valid. δThis indicates that these users belong to a group with similar interests. For example, user u i and user u j The check-in sets are respectively and If the similarity between two check-in sets is greater than or equal to the threshold s δ This indicates that there is a bidirectional edge relationship ε between the two users. ij ∈E u The formula for similarity to Jaccard is:

[0111] Through user social network graph G u A user similarity matrix can be constructed. Used to represent user u i and user u j If there is a relationship between them, the value is 1; otherwise, it is 0.

[0112] After obtaining the similarity matrix of the target user, it is also necessary to extract the user's social information features from it. Since the similarity matrix of the target user has at least one neighbor, in order to enhance the social feature information of the target user, it is also necessary to extract the features of the neighbors and enrich the target user's features. In order to improve the speed of obtaining user social information features, this embodiment uses a random neighbor sampling algorithm to extract features from the set of neighbor users, which is implemented by the social information influence module, such as... Figure 4 As shown.

[0113] The specific operation of the random neighbor sampling algorithm is as follows: for user u i Based on the user similarity matrix W u Get its neighbor user set And randomly sample a neighboring user u j At the same time, randomly obtain u j Information about a POI that I'm interested in. j Finally, a multimodal feature embedding aggregation layer is used to analyze user u. i Neighbor u j and POI node l j Vectorize and obtain respectively and And update user u i Its own information characteristics:

[0114] This study utilizes the similarity between historical check-in records of different users to construct a social network. A random neighbor sampling algorithm is used to process the network's information. For each user, the algorithm selects one of their neighbors and a point of interest that the neighbor is interested in. This information—including the current user, the neighbor, and the neighbor's point of interest—is then integrated into the current user's state, achieving the goal of interest propagation. Similar users exhibit similar preference changes within the social network. Random sampling can simulate the propagation of user preferences within the network and also absorb the preference characteristics of similar users, thereby enhancing the representation of user features.

[0115] Optionally, the S600 specifically includes:

[0116] S610. The feature vector information of long-term preference features, short-term preference features, geospatial features and user social information features are concatenated to generate interest point metadata.

[0117] S620. Calculate the probability distribution of the candidate set of the next interest point based on the interest point metadata;

[0118] S630. Select the top k with the highest probability in the candidate set as the user's preferred interest point recommendation list.

[0119] After obtaining the features of each modality, the modality information is then concatenated to obtain the feature representation e of the interest point metadata. j : Concat(·) represents the concatenation function, and the final embedding feature representation of the entire sequence is E: E = [e1, e2, ..., e n ].

[0120] The candidate set prediction module combines the contextual semantic information obtained from all the different modules mentioned above to perform analysis and calculation, and generates a candidate set for recommending the next POI to the user. For a given check-in sequence of length n... We can obtain input sequences of length n-1: and output sequence: Representing preference features Combined with current geospatial characteristics and user social information characteristics This constitutes the metadata of points of interest.

[0121] Input the metadata of the points of interest into the recommendation model to calculate the probability distribution of the candidate set of the next points of interest. The probability distribution of the candidate set of the next points of interest is obtained as follows: The candidate set of interest points includes the category candidate set C = [c1, c2, ..., c nAnd the candidate set of locations L = [l1, l2, ..., l n The top k items with the highest probabilities in the candidate set are selected as the user's preferred interest point recommendation list, where k is a positive integer and can be customized by the user.

[0122] The technical solution of this invention integrates long-term preference features, short-term preference features, geospatial features, and social information features of users to obtain user interest information in historical check-in information, spatiotemporal status information, and social networks. It comprehensively analyzes user interest data, generates a recommendation list of interest points that match user preferences, and makes reasonable recommendations for the user's next interest point. This solves the technical problems of sparsity in check-in data, insufficient breadth of metadata utilization, and inaccurate recommendations for inactive users. It achieves hierarchical feature extraction of user check-in sequences, adds geospatial features, and enhances user social information features. This enables full utilization of check-in data, comprehensive learning of user behavioral characteristics, accurate recommendation of the next interest point, and optimization of the user experience.

[0123] Example 2

[0124] like Figure 5 As shown, a continuous interest point recommendation system based on social spatiotemporal information and user preferences includes:

[0125] The user preference tracking module uses recurrent neural networks (RNNs) and multi-head attention mechanisms to analyze and process the user's historical check-in sequence, and obtain the user's long-term and short-term preference features in the user check-in sequence;

[0126] The geographic distance perception module uses a graph neural network (GNN) to obtain the geographic spatial features of each location in the user's check-in sequence;

[0127] The social information influence module uses a random neighbor sampling algorithm to extract features from the user's social network graph and obtain the user's social information features at each position in the user's check-in sequence;

[0128] The candidate set prediction module is used to add and fuse the feature information obtained from the user preference tracking module, the geographic distance perception module, and the social information influence module into multi-semantic features, and input them into the recommendation model to generate a list of interest points that match user preferences.

[0129] The continuous point of interest recommendation system based on social spatiotemporal information and user preferences provided in this embodiment of the invention can execute the continuous point of interest recommendation method based on social spatiotemporal information and user preferences provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0130] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for recommending continuous points of interest based on social spatio-temporal information and user preferences, characterized in that, The method comprises the following steps: obtaining long-term preference features and short-term preference features of a user in a user check-in sequence; The longitude and latitude information of all the interest points in each user check-in sequence dataset is counted, and the function is used to calculate the actual geographic distance between each two interest points. The longitude and latitude information of all the interest points in each user check-in sequence dataset is counted, and the function is used to calculate the actual geographic distance between each two interest points. setting a distance threshold, constructing an adjacency matrix according to the actual geographical distances, and obtaining a neighbor set of each interest point through the adjacency matrix; traversing the neighbor set of each interest point, respectively calculating the attention coefficient and the distance coefficient between each interest point and the neighbor node, and corresponding aggregation into the geographical space features of each interest point; obtaining user social information features of each position in the user check-in sequence; adding and fusing the long-term preference features, the short-term preference features, the geographical space features and the user social information features into multi-semantics features, and inputting into a recommendation model to generate an interest point recommendation list meeting the user's preference; wherein the obtaining of the long-term preference features and the short-term preference features of the user in the user check-in sequence comprises: extracting features of the input user check-in sequence through a multi-head attention mechanism, learning long-term preference features of the user from multiple subspaces; extracting features of the input user check-in sequence through a recurrent neural network (RNN), modeling recent check-in behaviors of the user, and learning short-term preference features of the user; the traversing of the neighbor set of each interest point, the respective calculation of the attention coefficient and the distance coefficient between each interest point and the neighbor node, and the corresponding aggregation into the geographical space features of each interest point comprise: obtaining initial features of a target interest point and a neighbor set of the target interest point; Obtain an interest point from the neighbor set in turn, and calculate the attention coefficient between the obtained interest point and the target interest point in turn: ; wherein, is an attention scoring mechanism, which calculates the correlation between two points through a scoring function to obtain a score, and then normalizes the score using softmax; represents a similarity calculation function, represents a trainable parameter matrix, is the feature of a node ; represents the neighbor interest point set of an interest point ; The distance coefficient between the acquired interest point and the target interest point is calculated in sequence: ; wherein, represents the distance between the two interest points; updating the feature of the target interest point according to the attention coefficient and the distance coefficient obtained each time: ; After traversing the neighbor set, a geographical space feature vector of the target interest point is constructed according to position information of the target interest point, and a geographical space feature of each interest point is constructed ; the adding and fusing of the long-term preference features, the short-term preference features, the geographical space features and the user social information features into multi-semantics features, and the inputting into the recommendation model to generate the interest point recommendation list meeting the user's preference comprise: splicing feature vector information of the long-term preference features, the short-term preference features, the geographical space features and the user social information features to generate interest point metadata; calculating a probability distribution of a candidate set of a next interest point based on the interest point metadata; Select the top of the candidate set as the user's favorite interest point recommendation list.

2. The method of claim 1, wherein, the obtaining of the user social information features of each position in the user check-in sequence comprises: constructing a social network graph according to preference association relationships between interest information of different users; obtaining a similarity matrix of the user through the social network graph; obtaining a corresponding neighbor user set according to the similarity matrix, and extracting features from the neighbor user set by using a random neighbor sampling algorithm; adding the extracted features to a target user to constitute user social information features.

3. The method of claim 2, wherein, The preference association relationship is expressed by a Jaccard-like similarity; The Jaccard-like similarity formula is: ; if the jaccard-like similarity is greater than a user similarity threshold then a preference association between the corresponding two users is determined to exist.

4. The method of claim 3, wherein, the candidate set of the interest point comprises a category candidate set and a location candidate set.

5. The method of claim 1, wherein, Before the obtaining of the long-term preference features and the short-term preference features of the user in the user check-in sequence, the method further comprises: converting scalar information of multiple modes in the user check-in sequence into low-dimensional dense feature vector information of multiple modes.

6. A continuous point-of-interest recommendation system based on social spatio-temporal information and user preferences, configured to perform the method of any of claims 1-5. The method comprises the following steps: a user preference tracking module uses a recurrent neural network (RNN) and a multi-head attention mechanism to analyze and process a user historical check-in sequence, and obtains long-term preference features and short-term preference features of a user in a user check-in sequence; a geographical distance perception module, which uses a graph neural network (GNN) to obtain geographical spatial features at each location in the user check-in sequence; a social information influence module, which uses a random neighbor sampling algorithm to extract features from a user social network graph to obtain user social information features at each location in the user check-in sequence; a candidate set prediction module, which is configured to add and fuse the feature information obtained by the user preference tracking module, the geographical distance perception module, and the social information influence module into multi-semantics features, and input the multi-semantics features into a recommendation model to generate a point of interest recommendation list that meets the user's preferences.

Images