A next interest point recommendation method combining user long-term and short-term preferences

Abstract

This invention belongs to the field of next interest recommendation in social networks, and relates to a method for recommending next interest points by combining users' long-term and short-term preferences. The method includes: acquiring raw data; preprocessing the raw data to obtain a new dataset; extracting relevant attributes from the new dataset to generate user visit interest point check-in records; extracting the spatiotemporal trajectory representation of user visit interest points based on the check-in records; constructing a user preference knowledge graph and a user preference heterogeneous graph based on the spatiotemporal trajectory of user visit interest points; inputting the user preference knowledge graph and user preference heterogeneous graph into a sequence model to obtain recommendation prediction probabilities; obtaining an ordered recommendation set based on the recommendation prediction probabilities; and selecting the optimal interest point from the ordered recommendation set for recommendation. This invention integrates user preference information, spatiotemporal information, and check-in information to recommend interest points that match the target user's long-term and short-term preferences.

Description

Technical Field

[0001] This invention belongs to the field of social network interest recommendation, specifically involving a method for recommending the next point of interest by combining users' short-term and long-term preferences. Background Technology

[0002] With the widespread use of mobile devices in daily life, an increasing number of location-based social services have emerged. People can share their life experiences and post check-in records on location-based social networks (LBSNs) such as Foursquare, Yelp, Facebook, and Brightkite. Domestic LBSN services, such as Dianping, Meituan, Ele.me, and Weibo, are developing rapidly. Statistics show that Meituan's user transaction volume increased by 12.5% ​​in 2019, with millions of reviews covering a wide range of interests including restaurants, hotels, and travel agencies. This rapid development of location-based social networks has led to a rapid increase in merchant and user data, bringing huge opportunities to both users and service providers. Analyzing massive amounts of LBSN data can help service providers recommend precise Points of Interest (POIs) to users. Predicting the next recommended point after a user visits an POI at a specific time has become a research hotspot in the recommendation field because its application scenarios are highly time-sensitive. POI recommendations can not only bring economic benefits to LBSN service providers but also help users discover potential interests, thus bringing enjoyment and convenience to their lives. For users, accurate and effective point-of-interest (POI) recommendations help them better understand their surroundings, make more informed travel decisions, enhance their enjoyment of life, and better understand their city. For businesses, providing precise POI recommendations can bring economic benefits and increase platform visibility. Furthermore, this service can offer personalized services, thereby improving user experience. By analyzing users' personalized preferences and identifying the geographical scope of popular POIs, businesses can achieve targeted marketing and further increase revenue.

[0003] In recent years, location-based social network point-of-interest (POI) recommendation methods have gained increasing popularity in the Internet of Things (IoT) and e-commerce. Among these methods, mining users' short-term and long-term preferences plays a crucial role in improving user experience and service quality. This approach can better meet user needs, increase user satisfaction, and consequently boost service usage and merchant revenue. Existing POI recommendation technologies can be categorized into collaborative filtering, matrix factorization, Markov chain-based methods, and deep learning-based methods. Collaborative filtering methods start with users or POIs, calculate the similarity between them, and then recommend POIs to the target user. Matrix factorization methods, on the other hand, primarily start with users' check-in records for POIs, construct a check-in matrix, and mine user preferences and interest attributes through matrix factorization. They can recommend personalized content based on users' historical behavior and interests, improving user satisfaction and user experience. Markov chain-based methods are recommendation methods based on modeling user access patterns to POIs. They record user access sequences and assume that the next POI access is related to the previous one, using modeling to mine sequence patterns for recommendations. Deep learning-based interest recommendation methods integrate various contextual information, such as time, space, interest category, and social relationships. They construct recurrent neural networks and incorporate attention mechanisms to uncover user interests and their attributes, recommending potentially relevant content. Both methods can recommend personalized content based on user history and contextual information, improving user satisfaction and overall experience.

[0004] Although extensive research has been conducted on interest-based recommendation models and significant progress has been made, some challenges remain:

[0005] 1. Existing spatiotemporal dynamic knowledge graphs assign the same preference weight to each POI in the check-in order, ignoring the fact that user preferences change according to POI; obviously, ignoring or weakening users' expression of item preferences is not advisable for recommendation systems.

[0006] 2. In the field of point of interest recommendation, the proportion of points of interest accessed in the user's trajectory to the total number of points of interest in the dataset is very small, which leads to a serious data sparsity problem.

[0007] 3. The existing user information is too generic and cannot reflect the different preferences users have for different points of interest (POIs) while on the go, which weakens the personalization of recommendations in the recommendation system. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for recommending the next point of interest by combining users' short-term and long-term preferences, comprising:

[0009] S1. Obtain the raw data and preprocess it to obtain a new dataset; the raw data includes user access record data, check-in sequence data, and spatiotemporal information data.

[0010] S2. Extract relevant attributes from the new dataset and generate check-in records for user visits to points of interest based on the relevant attributes; the relevant attributes include the time information and latitude and longitude coordinate information of the user's visits to the points of interest.

[0011] S3. Extract the spatiotemporal trajectory representation of the user's access to points of interest based on the user's check-in records, and construct a user preference knowledge graph and a user preference heterogeneous graph with spatiotemporal characteristics based on the spatiotemporal trajectory representation of the user's access to points of interest.

[0012] S4. Input the spatiotemporal trajectory representation of user access points of interest, user preference knowledge graph and user preference heterogeneous graph into the sequence model to obtain the recommendation prediction probability of points of interest;

[0013] S5. Select interest points based on the predicted probability of the interest points to obtain an ordered recommendation set;

[0014] S6. Select the optimal point of interest from the ordered recommendation set for recommendation.

[0015] Step S1 preprocesses the original data to obtain a new dataset, including:

[0016] Clean the original data; set a check-in threshold; based on the check-in threshold, count the number of times each user checks in on a point of interest and the number of times each point of interest is checked in by a user in the cleaned data. If a user's check-in count is lower than the check-in threshold, delete that user's check-in data. If a point of interest is checked in by a user less than the check-in threshold, delete that point of interest and obtain a new dataset.

[0017] Step S2 extracts relevant attributes from the new dataset, including:

[0018] S21. Obtain user and point of interest ID information, check-in timestamp, and spatial latitude and longitude coordinates of point of interest from user access record data, check-in sequence data, and spatiotemporal information data in the new dataset.

[0019] S22. Convert the check-in timestamps from the obtained data into two time periods: Monday to Friday as one time period and the weekend as another time period.

[0020] Extracting the spatiotemporal trajectory representation of user access points of interest, including:

[0021] L = {l1, l2, ..., l n}

[0022] Where L represents the spatiotemporal trajectory of a user's access to points of interest, li This represents each check-in record of a user, l i ={p1,p2,……,p n}, p i This represents the point of interest, and n represents the sign-in order in ascending chronological order.

[0023] The user preference knowledge graph includes: user preference triples and spatiotemporal triples; the user preference triples are (p i ,W,p j The spatiotemporal triples are (u, t, p) and (p, (Δt, Δd), p); where (Δt, Δd) represents the spatiotemporal interval between points of interest, Δt represents the time interval, Δd represents the spatial interval, u represents the current user, p represents the current POI, w represents the preference weight of a single user for the POI, and p i Let p represent the i-th POI. j Let t represent the next POI to be transferred, t represent time, and W represent the global user preference weight for the POI.

[0024] Step S3, constructing a heterogeneous user preference graph based on the spatiotemporal trajectory of user access points of interest, includes:

[0025] S31. Calculate each user's preference value for each POI based on the spatiotemporal trajectory of the user's visits to POIs, and save the preference value to a matrix. middle;

[0026] S32. Construct a user-POI interaction graph G based on the spatiotemporal trajectory of users visiting points of interest;

[0027] S33, will By combining it with the user-POI interaction graph G, a heterogeneous graph of user preferences was obtained.

[0028] The sequence model includes: a knowledge graph embedding module, a graph convolution module, a GRU sequence module, and a fully connected layer. The GRU sequence module includes a long-term preference module and a short-term preference module. The process by which the sequence model processes the spatiotemporal trajectory representation of user access points of interest, the user preference knowledge graph, and the user preference heterogeneous graph includes:

[0029] S41. Input the heterogeneous user preference graph into the graph convolution module to obtain the user preference vector;

[0030] S42. Input the user preference knowledge graph into the knowledge graph embedding module for knowledge graph embedding to obtain entity vectors and relation vectors; the relation vectors include spatiotemporal relation vectors and preference relation vectors.

[0031] S43. Input the user preference vector, entity vector and relation vector into the long-term preference module to obtain the user's long-term preference vector;

[0032] S44. Input the spatiotemporal trajectory representation of the user's access to points of interest, the user preference vector, the entity vector, and the relation vector into the short-term preference module to obtain the continuous short-term preference vector and the non-continuous short-term preference vector.

[0033] S45. Input the user's long-term preference vector, continuous short-term preference vector, and discontinuous short-term preference vector into the fully connected layer to obtain the recommendation prediction probability of the interest point.

[0034] The short-term preference module includes: GRU, user preference relation path encoding module, short-term attention mechanism module, and Time-Geo-dilated GRU module; where Time-Geo-dilated GRU is a spatiotemporal GRU; the short-term preference module processes the spatiotemporal trajectory representation of user access points of interest, user preference vector, entity vector, and relation vector, including:

[0035] S441. Input the spatiotemporal relationship vector into the Time-Geo-dilated GRU to obtain the discontinuous short-term preference vector;

[0036] S442. Input the user preference vector and the spatiotemporal relationship vector into the user preference relationship path encoding module for combination to obtain the combined vector;

[0037] S443. Input the spatiotemporal trajectory representation L of the user's access to points of interest into the GRU to obtain the current sequence vector z;

[0038] S444. Input the combined vector and the current sequence vector into the short-term attention mechanism module to obtain a continuous short-term preference vector.

[0039] The long-term preference module includes: a local attention mechanism module, a non-local operation module, and GRU; the long-term preference module processes user preference vectors, entity vectors, and relation vectors in the following ways:

[0040] S431. Add the spatiotemporal relation vector and the entity vector to obtain vector v. st The user preference vector and the entity vector are added together to obtain the vector v. up ;

[0041] S432, v st and v up The input is fed into the local attention mechanism module to obtain the output vector;

[0042] S433. Input the output vector into the GRU to obtain the GRU output vector;

[0043] S434. Input the GRU output vector into the nonlocal operation module to perform average pooling operation to obtain the long-term preference vector.

[0044] Step S6 selects interest points based on the predicted probability of the interest points, resulting in an ordered recommendation set including:

[0045] Obtain all points of interest in the new dataset to obtain a candidate point of interest set; sort the candidate points of interest in the candidate point of interest set in descending order according to the recommendation prediction probability of the points of interest to obtain a sorted set; select the points of interest in the sorted set in the TOP-N manner to obtain an ordered recommendation set.

[0046] The beneficial effects of this invention are:

[0047] 1. This invention utilizes user check-in point of interest (POI) behavior data to construct a novel user preference knowledge graph. This knowledge graph possesses spatiotemporal features built upon users' historical movement trajectories, capturing users' dynamic spatiotemporal transition characteristics and adding user preferences for POI visits and path selection between two POIs, thus solving the data sparsity problem. 2. This invention integrates user preference information, spatiotemporal information, and check-in information for user visits to POIs, learning users' long-term and short-term preferences. It recommends POIs that match the long-term and short-term preferences of target users, helping users quickly explore and plan their travel purposes. It also helps businesses understand the distribution of user behavior characteristics, analyze potential users, and achieve significant economic and social benefits. Attached Figure Description

[0048] Figure 1 This is a flowchart of a method for recommending the next point of interest that combines users' short-term and long-term preferences, according to the present invention.

[0049] Figure 2 This is a diagram illustrating the user preference and interest extraction structure of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] like Figure 1 As shown, this invention proposes a method for recommending the next point of interest by combining users' short-term and long-term preferences, including the following steps:

[0052] S1. Obtain the raw data and preprocess it to obtain a new dataset; the raw data includes user access record data, check-in sequence data, and spatiotemporal information data.

[0053] S2. Extract relevant attributes from the new dataset to generate check-in records for user visits to points of interest; the relevant attributes include the time information and latitude and longitude coordinates of the user's visits to the points of interest;

[0054] S3. Extract the spatiotemporal trajectory representation of the user's access to points of interest based on the user's check-in records, and construct a user preference knowledge graph and a user preference heterogeneous graph with spatiotemporal characteristics based on the spatiotemporal trajectory of the user's access to points of interest.

[0055] S4. Input the spatiotemporal trajectory representation of user access points of interest, user preference knowledge graph and user preference heterogeneous graph into the sequence model to obtain the recommendation prediction probability of points of interest;

[0056] S6. Select interest points based on the predicted probability of the interest points to obtain an ordered recommendation set;

[0057] S7. Select the optimal point of interest from the ordered recommendation set for recommendation.

[0058] Step S1 preprocesses the original data to obtain a new dataset, including:

[0059] Clean the original data; set a check-in threshold; based on the check-in threshold, count the number of times each user checks in on a point of interest and the number of times each point of interest is checked in by a user in the cleaned data. If the number of times a user checks in on a point of interest is less than the check-in threshold, delete the user's check-in data. If the number of times a point of interest is checked in by a user is less than the check-in threshold, delete the point of interest and obtain a new dataset.

[0060] The check-in thresholds are set to 5 and 10. If a user checks in a point of interest less than 5 times and all points of interest are checked in by the user less than 10 times, then the user's check-in data and the point of interest are considered not representative.

[0061] Step S2 extracts relevant attributes from the new dataset, including:

[0062] S21. Obtain the user and point of interest ID information, check-in timestamp, and spatial latitude and longitude coordinates of the point of interest from the obtained user access record data, check-in sequence data, and spatiotemporal information data.

[0063] S22. Convert the check-in timestamps from the acquired data into two time periods: Monday to Friday as one time period, and the weekend as another. Represent the check-in time information as a specific hour of the week.

[0064] Step S3 extracts the spatiotemporal trajectory representation of the user's access to points of interest, including:

[0065] L = {l1, l2, ..., l n}

[0066] Where L represents the spatiotemporal trajectory of a user's access to points of interest, l i This represents each check-in record of a user, l i ={p1,p2,……,p n}, p i This represents the point of interest, and n represents the sign-in order in ascending chronological order.

[0067] The user preference knowledge graph includes: user preference triples and spatiotemporal triples; the user preference triples are (p i ,W,p j ), (u,w,p), the spatiotemporal triple is (u,t,p), (p i ,(Δt,Δd),p j );

[0068] Where (Δt, Δd) represents the spatiotemporal interval between points of interest, Δt represents the time interval, Δd represents the spatial interval, u represents the user, p represents the POI, and w represents the user's preference weight for the POI. i Let p represent the i-th POI. j Indicates the POI of the next transition, t represents time, and W represents p. i For p j Preference weights.

[0069] Step S3, constructing a heterogeneous user preference graph based on the spatiotemporal trajectory of user access points of interest, includes:

[0070] S31. Based on the spatiotemporal trajectory L = {l1, l2, ..., l...} of the user's access to points of interest, ... n} Calculate the preference value μ for each user for each POI. ij :

[0071]

[0072]

[0073] μ ij =κ ij -κ j ;

[0074] S32, Add a matrix Save each user's preference value μ for each POI. ij ;

[0075] S33. Construct a user-POI interaction graph G based on the spatiotemporal trajectory of users visiting points of interest;

[0076] S34, will By combining it with the user-POI interaction graph G, a heterogeneous graph of user preferences was obtained.

[0077]

[0078] Among them, P ij The value represents the frequency with which user i accesses POIj, κ ij κ represents the proportion of user i's check-ins on POIj. j It is the average sign-in rate for each POI, μ ij is the preference value of user i when accessing POIj, n is the number of POIs, and m is the number of users;

[0079] like Figure 2 As shown, the sequence model includes: a knowledge graph embedding module, a GCN graph convolution module, a GRU sequence module, and a fully connected layer. The GRU sequence module includes a long-term preference module and a short-term preference module. The short-term preference module includes: GRU, Time-Geo-dilated GRU, User Preference Transfer Relation Path Encoding module, and Short-Term Relation Attention module. User Preference Transfer Relation Path Encoding is the encoding of user preference relationship paths, Short-Term Relation Attention is a short-term attention mechanism, and Time-Geo-dilated GRU is a spatiotemporal GRU. The long-term preference module includes: a Local Attention module, a Nonlocal Operation module, and GRU. Local Attention is a local attention mechanism, and Nonlocal Operation is a nonlocal operation.

[0080] The process by which the sequence model processes the spatiotemporal trajectory representation of user access points of interest, the user preference knowledge graph, and the user preference heterogeneous graph includes:

[0081] S41. Input the heterogeneous user preference graph into the graph convolution module to obtain the user preference vector;

[0082] S42. Input the user preference knowledge graph into the knowledge graph embedding module for knowledge graph embedding to obtain entity vectors and relation vectors; the relation vectors include spatiotemporal relation vectors and preference relation vectors.

[0083] S43. Input the user preference vector, entity vector and relation vector into the long-term preference module to obtain the user's long-term preference vector;

[0084] S44. Input the spatiotemporal trajectory representation of the user's access to points of interest, the user preference vector, and the relation vector into the short-term preference module to obtain the continuous short-term preference vector and the non-continuous short-term preference vector.

[0085] S45. Based on the user's long-term preference vector, continuous short-term preference vector, and discontinuous short-term preference POI sequence, obtain the candidate interest point set;

[0086] S46. Input the candidate interest point set into the fully connected layer to obtain the recommendation prediction probability of the candidate interest points.

[0087] Step S44 inputs the spatiotemporal trajectory representation of the user's access to points of interest, the user preference vector, the entity vector, and the relation vector into the short-term preference module, including:

[0088] S441. Input the spatiotemporal relation vector into the Time-Geo-dilated GRU module to obtain the discontinuous short-term preference vector z'. n :

[0089] z' n =Time-Geo-dilated GRU(p n-1 ,z' n-γ )

[0090] Time-Geo-dilated GRU explores the impact of spatiotemporal dependencies between discontinuous Points of Interest (POIs) in the current trajectory through spatiotemporal relationships. This includes: calculating the similarity π between POIs based on the spatiotemporal relationship vector: π = λ*Δd + (1-λ)Δt; obtaining the jump parameter for each POI based on the similarity π; for example, if the starting point is p1, and p1 and p4 have high similarity, then the jump parameter is 3; inputting the spatiotemporal relationship vector and jump parameter into the GRU, the GRU performs jump connections between nodes of highly similar POIs in the spatiotemporal relationship vector based on the jump parameter, obtaining the discontinuous short-term preference vector z'. n Where λ is a parameter for adjusting time and geographical location, and p n-1 For the POI after the jump, z' n-γ The output of the hidden layer is γ, which represents the skip parameter.

[0091] S442. Input the user preference vector and spatiotemporal relationship vector into the user preference relationship path encoding module. Combine the user preference vector and spatiotemporal relationship vector using knowledge graph path reasoning technology to achieve inference of user preferences for continuous POIs, and obtain the combined vector z. up :

[0092]

[0093] S443. Input the spatiotemporal trajectory representation L of the user's access to points of interest into the GRU to obtain the current sequence vector z:

[0094] z = GRU(L)

[0095] S444, z up The current sequence vector z is input into the short-term attention mechanism module to obtain a continuous short-term preference vector p. s ;

[0096]

[0097]

[0098] Where, β i The attention coefficient represents the short-term attention mechanism, Δt represents the time interval, Δd represents the spatial interval, u' is the user preference vector, and r st represents the spatiotemporal relationship vector in the knowledge graph, where n is the number of POIs;

[0099] Step S43 inputs the user preference vector, entity vector, and relation vector into the long-term preference module, including:

[0100] S431. Add the entities in the spatiotemporal relation vector and the entity vector to obtain the vector v. st The vector v is obtained by adding the entities in the user preference vector and the entity vector. up ;

[0101] S432, v st and v up The input is fed into the local attention mechanism module to obtain the output vector p. L :

[0102]

[0103]

[0104] S433, output vector p L Input into the GRU, and obtain the GRU output sequence.

[0105]

[0106] S435, Output sequence of GRU The input nonlocal operation module performs average pooling to obtain the long-term preference vector;

[0107] Where α is the attention coefficient in the local attention mechanism, and n is the number of points of interest (POIs).

[0108] Step S6 selects interest points based on the predicted probability of the interest points, resulting in an ordered recommendation set including:

[0109] S61. Obtain all points of interest set(POI) = {p1, p2, p3…p n}, where p1 represents the first interest point in the set of interest points, and n represents the number of interest points;

[0110] S62. Sort the candidate interest points in the candidate interest point set in descending order according to the recommendation prediction probability of the interest points, and recommend the set of interest points that the user is most likely to visit at the next time point to the target user in the TOP-N method, so as to obtain an ordered recommendation set; the TOP-N method is to select the N interest points with the smallest prediction probability.

[0111] The recommendation prediction probability of candidate interest points is obtained through a fully connected approach, including:

[0112] y' = liner(y)

[0113] prob = softmax(y')

[0114] Where y represents the candidate interest point set vector obtained by concatenating the user's long-term preference vector, continuous short-term preference vector, and discontinuous short-term preference vector, y' represents the interest point set vector after full connection, softmax() represents the activation function, and prob represents the recommendation prediction probability of the candidate interest points.

[0115] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recommending the next point of interest by combining users' short-term and long-term preferences, characterized in that, include: S1. Obtain the raw data and preprocess it to obtain a new dataset; the raw data includes user access record data, check-in sequence data, and spatiotemporal information data. S2. Extract relevant attributes from the new dataset and generate check-in records for user visits to points of interest based on the relevant attributes; the relevant attributes include the time information and latitude and longitude coordinate information of the user's visits to the points of interest. S3. Extract the spatiotemporal trajectory representation of the user's access to points of interest based on the user's check-in records, and construct a user preference knowledge graph and a user preference heterogeneous graph with spatiotemporal characteristics based on the spatiotemporal trajectory representation of the user's access to points of interest. The user preference knowledge graph includes: user preference triples and spatiotemporal triples; the user preference triples are... , The spatiotemporal triple is , ;in, Indicates the spatiotemporal interval between points of interest. Indicates time interval, Indicates spatial interval, Indicates the current user. Indicates the current POI. This represents the preference weight of a single user for a POI. This represents the i-th POI. Indicates the POI for the next transfer. Indicates time, This represents the global user preference weight for POIs; Step S3, constructing a heterogeneous user preference graph based on the spatiotemporal trajectory of user access points of interest, includes: S31. Calculate each user's preference value for each POI based on the spatiotemporal trajectory of the user's visits to POIs, and save the preference value to a matrix. middle; S32. Construct a user-POI interaction graph based on the spatiotemporal trajectory of users visiting points of interest. ; S33, will Interaction diagram with user-POI By combining these, a heterogeneous graph of user preferences was obtained. ; S4. Input the spatiotemporal trajectory representation of user access points of interest, user preference knowledge graph and user preference heterogeneous graph into the sequence model to obtain the recommendation prediction probability of points of interest; The sequence model includes: a knowledge graph embedding module, a graph convolution module, a GRU sequence module, and a fully connected layer. The GRU sequence module includes a long-term preference module and a short-term preference module. The process by which the sequence model processes the spatiotemporal trajectory representation of user access points of interest, the user preference knowledge graph, and the user preference heterogeneous graph includes: S41. Input the heterogeneous user preference graph into the graph convolution module to obtain the user preference vector; S42. Input the user preference knowledge graph into the knowledge graph embedding module for knowledge graph embedding to obtain entity vectors and relation vectors; the relation vectors include spatiotemporal relation vectors and preference relation vectors. S43. Input the user preference vector, entity vector and relation vector into the long-term preference module to obtain the user's long-term preference vector; S44. Input the spatiotemporal trajectory representation of the user's access to points of interest, the user preference vector, and the relation vector into the short-term preference module to obtain the continuous short-term preference vector and the non-continuous short-term preference vector. S45. Input the user's long-term preference vector, continuous short-term preference vector, and discontinuous short-term preference vector into the fully connected layer to obtain the recommendation prediction probability of the point of interest. S5. Select interest points based on the predicted probability of the interest points to obtain an ordered recommendation set; S6. Select the optimal point of interest from the ordered recommendation set for recommendation.

2. The method for recommending the next point of interest by combining users' short-term and long-term preferences as described in claim 1, characterized in that, Step S1 preprocesses the original data to obtain a new dataset, including: Clean the original data; set a check-in threshold; based on the check-in threshold, count the number of times each user checks in on a point of interest and the number of times each point of interest is checked in by a user in the cleaned data. If a user's check-in count is lower than the check-in threshold, delete that user's check-in data. If a point of interest is checked in by a user less than the check-in threshold, delete that point of interest and obtain a new dataset.

3. The method for recommending the next point of interest by combining users' short-term and long-term preferences as described in claim 1, characterized in that, Step S2 extracts relevant attributes from the new dataset, including: S21. Obtain user and point of interest ID information, check-in timestamp, and spatial latitude and longitude coordinates of point of interest from user access record data, check-in sequence data, and spatiotemporal information data in the new dataset. S22. Convert the check-in timestamps from the obtained data into two time periods: Monday to Friday as one time period and the weekend as another time period.

4. The method for recommending the next point of interest by combining users' short-term and long-term preferences as described in claim 1, characterized in that, Extracting the spatiotemporal trajectory representation of user access points of interest, including: in, This represents the spatiotemporal trajectory of a user's visits to points of interest. This represents each of a user's check-in records. , Indicate points of interest, This indicates the sign-in order in ascending chronological order.

5. The method for recommending the next point of interest by combining users' short-term and long-term preferences according to claim 1, characterized in that, The long-term preference module includes: a local attention mechanism module, a non-local operation module, and GRU; the long-term preference module processes user preference vectors, entity vectors, and relation vectors in the following ways: S431. Add the spatiotemporal relation vector and the entity vector to obtain the vector. The vector is obtained by adding the user preference vector and the entity vector. ; S432, will and The input is fed into the local attention mechanism module to obtain the output vector; S433. Input the output vector into the GRU to obtain the GRU output vector; S434. Input the GRU output vector into the nonlocal operation module to perform average pooling operation to obtain the long-term preference vector.

6. The method for recommending the next point of interest by combining users' short-term and long-term preferences according to claim 1, characterized in that, The short-term preference module includes: GRU, user preference relation path encoding module, short-term attention mechanism module, and Time-Geo-dilated GRU; where Time-Geo-dilated GRU is a spatiotemporal GRU; the short-term preference module processes the spatiotemporal trajectory representation of user access points of interest, user preference vector, and relation vector, including: S441. Input the spatiotemporal relationship vector into the Time-Geo-dilated GRU to obtain the discontinuous short-term preference vector; S442. Input the user preference vector and the spatiotemporal relationship vector into the user preference relationship path encoding module for combination to obtain the combined vector; S443. Represent the spatiotemporal trajectory of user visits to points of interest. Input into the GRU to obtain the current sequence vector; S444. Input the combined vector and the current sequence vector into the short-term attention mechanism module to obtain a continuous short-term preference vector.

7. The method for recommending the next point of interest by combining users' short-term and long-term preferences according to claim 1, characterized in that, Step S6 selects interest points based on the predicted probability of the interest points, resulting in an ordered recommendation set including: Obtain all points of interest in the new dataset to obtain a candidate point of interest set; sort the candidate points of interest in the candidate point of interest set in descending order according to the recommendation prediction probability of the points of interest to obtain a sorted set; select the points of interest in the sorted set in the TOP-N manner to obtain an ordered recommendation set.

Images