Interest point recommendation method based on user multi-behavior enhancement and efficient rich information negative sampling

Abstract

This invention discloses an interest point recommendation method based on user multi-behavior enhancement and efficient rich-information negative sampling, comprising: acquiring and preprocessing user multi-behavior data; constructing three types of context graphs using the user multi-behavior data; learning multi-view representations of users and interest points using a Skip-Gram model; constructing a pairwise dataset, calculating the probability score of each user-interest point check-in pair using a pairwise ranking model, and calculating the BPR loss function; replacing the negative sampling part in the above two models with a negative sampling method based on a baseline point graph; designing a joint learning framework that combines the loss of the Skip-Gram model with the loss function of BPR; calculating the probability score of a user checking in at any interest point in the future using a pairwise ranking model, and recommending interest points to the user. This invention's method can recommend interest points to users more accurately, and its graph construction method can be flexibly applied to different recommendation scenarios.

Description

Technical Field

[0001] This invention belongs to the field of Point-of-Interest (POI) recommendation technology, and in particular, it is a point-of-interest recommendation method based on user multi-behavior enhancement and efficient information-rich negative sampling. Background Technology

[0002] With the rapid development of internet technology, online information is experiencing explosive growth. Filtering out information of interest to users from this massive amount of data has become increasingly difficult, leading to the emergence of personalized recommendation systems. Statistics show that by 2020, my country had over 986 million mobile internet users. These users continuously provide the government, businesses, and various institutions with mobile information related to public transportation, urban development, and more. Based on GPS information, location-based networking services (LBNS) have been developed, promoting route planning, point-of-interest location search, and public services, providing convenience and security for urban residents, guiding urban planning towards development goals, making business site selection more rational, and generating considerable commercial benefits.

[0003] Point-of-interest (POI) recommendation, a highly challenging research area in urban computing, plays a crucial role in LBNS (Blockchain-Based Networking). With the maturity of wireless communication and location acquisition technologies, numerous location-based applications have emerged, such as Didi Chuxing, Google Maps, and Uber. Users of these applications generate a wealth of behavioral records (e.g., check-in information, movement trajectories, and map queries), which document users' real-life experiences. The task of POI recommendation is to leverage these behavioral records to personalize recommendations for users, recommending places they are interested in, or even places they haven't yet visited. This helps users explore urban life, provides travel services, improves their quality of life, and simultaneously drives new commercial value.

[0004] Traditional point-of-interest (POI) recommendation methods are based on collaborative filtering (CF) and utilize user check-in data to understand user preferences. However, severe data sparsity significantly degrades POI recommendation performance due to the difficulty in collecting sufficient check-in logs and limitations in user activity range. To address this issue, some studies have constructed ranked pairs to more fully describe user preferences. Meanwhile, Jiao et al. utilized contextual information, such as the category and geographical influence of POIs, to better understand user browsing behavior. Recently, Xie, Zhang et al. proposed several graph-based POI recommendation methods, aiming to extract relationship graphs from various contexts and learn representations through graph embeddings or GNN models for more accurate modeling.

[0005] While existing research has yielded numerous studies on point-of-interest (POI) recommendation, most rely solely on historical user check-in data. Furthermore, the sparsity of this data severely limits the analysis of user check-in behavior. Additionally, these works neglect effective research into user map query behavior and valid negative samples. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art by providing a point-of-interest (POI) recommendation method based on enhanced user multi-behavior and efficient information-rich negative sampling. This method is applicable to user check-in datasets and map query datasets, enabling more complete extraction of user multi-behavior relationships. Furthermore, it explores the impact of more informative negative samples collected based on geographical factors on users' travel POI selection, thereby enhancing the performance of POI recommendation and recommending POIs that better match users' preferences.

[0007] The technical solution to achieve the purpose of this invention is: an interest point recommendation method based on user multi-behavior enhancement and efficient rich information negative sampling, the method comprising the following steps:

[0008] Step 1: Obtain user multi-behavior data and preprocess the data to exclude outlier data;

[0009] Step 2: Construct three types of contextual relationship graphs using user multi-behavior data: Check-in interest point relationship graph G and query interest point relationship graph G. p-q User-Query Interest Relationship Graph G u-q And the relationship diagram G between check-in interest points and check-in interest points. p-p In the relationship graph, each node represents a user or point of interest.

[0010] Step 3: Use the Skip-Gram model to mine user multi-behavior relationships in three context relationship graphs and learn multi-view representations of users and points of interest;

[0011] Step 4: Construct a pairwise dataset, use a pairwise ranking model to calculate the probability score of each user-point of interest check-in pair, and calculate the BPR loss;

[0012] Step 5: Replace the negative sampling portion in Skip-Gram and the pairwise sorting model with a negative sampling method based on the baseline point map;

[0013] Step 6: Design a joint learning framework that combines the loss function of the Skip-Gram model and the BPR method to jointly learn the representations of users and points of interest in the check-in and query spaces.

[0014] Step 7: Use a pairwise ranking model to calculate the probability score of a user checking in at any point of interest in the future, and recommend points of interest to the user based on this score.

[0015] An interest point recommendation system based on user multi-behavior enhancement and efficient rich information negative sampling, the system comprising:

[0016] The first module is used to acquire user multi-behavior data and preprocess the data to exclude outlier data.

[0017] The second module is used to construct three types of contextual relationship graphs using user multi-behavior data: a relationship graph G between check-in interest points and query interest points. p-q User-Query Interest Relationship Graph G u-q And the relationship diagram G between check-in interest points and check-in interest points. p-p In the relationship graph, each node represents a user or point of interest.

[0018] The third module is used to mine user multi-behavior relationships in three types of context graphs using the Skip-Gram model, and to learn multi-view representations of users and points of interest.

[0019] The fourth module is used to construct the pairwise dataset, use the pairwise ranking model to calculate the probability score of each user-point of interest check-in pair, and calculate the BPR loss;

[0020] The fifth module is used to replace the negative sampling portion in Skip-Gram and pairwise sorting models with a negative sampling method based on the baseline plot;

[0021] The sixth module is used to design a joint learning framework that combines the loss functions of the Skip-Gram model and the BPR method to jointly learn the representations of users and points of interest in the check-in and query spaces.

[0022] The seventh module is used to calculate the probability score of a user checking in at any point of interest in the future using a pairwise ranking model, and to recommend points of interest to the user based on this score.

[0023] Compared with the prior art, the significant advantages of this invention are:

[0024] 1) Innovatively explore the implicit information and inherent relationships of user multi-behavioral data. By combining user check-in behavior and map query behavior, the problem of data scarcity can be alleviated.

[0025] 2) Based on user check-in behavior and map query behavior, construct three context relationship graphs to learn more accurate user and point of interest representations.

[0026] 3) A negative sampling method based on a benchmark map was used, which incorporates geographical factors into the sampling process to generate more informative negative samples, thereby improving the model's convergence speed and the accuracy of representation learning.

[0027] 4) A joint learning framework was used to avoid overfitting by sharing representations of users and points of interest.

[0028] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0029] Figure 1 The flowchart shows the interest point recommendation method based on user multi-behavior enhancement and efficient rich information negative sampling.

[0030] Figure 2 Figure (a) shows the performance comparison of the present invention and four variants on Precision, and Figure (b) shows the performance comparison of the present invention and four variants on Recall.

[0031] Figure 3 Figure (a) shows the performance comparison results of the present invention and four variants on MAP, and Figure (b) shows the performance comparison results of the present invention and four variants on NDCG. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] It should be noted that if the embodiments of the present invention involve descriptions such as "first" and "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0034] In one embodiment, an interest point recommendation method based on user multi-behavior enhancement and efficient rich-information negative sampling is provided, the method comprising the following steps:

[0035] Step 1: Obtain user multi-behavior data (user interest point check-in data and user map query data), and preprocess the data to exclude outlier data;

[0036] Here, the user multi-behavior data includes:

[0037] User interest point check-in data includes user ID, check-in time, interest point ID, interest point name, interest point category, and interest point location;

[0038] User map query data includes user ID, query time, point of interest ID, point of interest name, point of interest category, and point of interest location.

[0039] Step 2: Construct three types of contextual relationship graphs using user multi-behavior data: Check-in interest point relationship graph G and query interest point relationship graph G. p-q User-Query Interest Relationship Graph G u-q , Check-in Interest Points and Relationship Diagram G p-p Each node in the diagram represents a user or point of interest.

[0040] Step 3: Use the Skip-Gram model to mine user multi-behavior relationships in three context relationship graphs and learn multi-view representations of users and points of interest;

[0041] Step 4: Construct a pairwise dataset, use a pairwise ranking model to calculate the probability score of each user-point of interest check-in pair, and calculate the BPR loss;

[0042] Step 5: Replace the negative sampling portion in Skip-Gram and the pairwise sorting model with a negative sampling method based on the baseline point map;

[0043] Here, since ordinary negative sampling methods cannot filter through the large number of negative samples, selecting more important and informative negative samples can improve the convergence speed of the model and the accuracy of representation learning. Therefore, a negative sampling method based on a benchmark point map is adopted to improve the accuracy of representation learning.

[0044] Step 6: Design a joint learning framework that combines the loss function of the Skip-Gram model and the BPR method to jointly learn the representations of users and points of interest in the check-in and query spaces.

[0045] Step 7: Use a pairwise ranking model to calculate the probability score of a user checking in at any point of interest in the future, and recommend points of interest to the user based on this score.

[0046] Furthermore, in one embodiment, the three context relationship graphs constructed in step 2 are specifically as follows:

[0047] Step 2-1: First, divide the day into four time periods: (1) the morning period, from 3:00 AM to 10:59 AM; (2) the afternoon period, from 11:00 AM to 3:59 PM; (3) the evening period, from 4:00 PM to 9:59 PM; and (4) the midnight period, from 10:00 PM to 2:59 AM. Combine this with the weekday and weekend periods to obtain eight time periods.

[0048] Step 2-2: Based on the above time periods, construct three types of context relationship diagrams:

[0049] The check-in interest point and query interest point map G p-q Specifically, edges are established between points of interest that exhibit check-in and query behavior within the same day or time period, represented as follows: in Indicates point of interest v j In the sign-in area, Indicates point of interest v k Representation in the query space; the graph can be represented as G p-q =(V, E) p-q W p-q ), where V represents the set of points of interest, E p-q Let W represent the set of edges in the graph. p-q Record the number of times the edges occur in the graph;

[0050] The user-interest relationship graph G u-q Specifically, this involves establishing edges between a user and their points of interest queried within the same time period, represented as... Where u i Representing user i, Indicates point of interest v k Representation in the query space; the graph can be represented as G u-q = (U, V, E) u-q W u-q ), where U represents the set of users and V represents the set of points of interest E. u-q Let W represent the set of edges in the graph. u-q Record the number of times the edges occur in the graph;

[0051] The relationship diagram G between check-in interest points and check-in interest points p-p Specifically, this involves establishing edges between points of interest that a user checks in within the same day or time period, represented as follows: in and These represent points of interest, v. j and v k The representation in the check-in space; this diagram can be represented as G. p-p =(V, E) p-p Wp-p ); where V represents the set of points of interest, E p-p Let W represent the set of edges in the graph. p-p Record the number of times an edge occurs in the graph.

[0052] Furthermore, in one embodiment, the Skip-Gram model used in step 3 is specifically:

[0053] Regarding the check-in interest points and query interest point map G mentioned in step 2 p-q The loss function of the Skip-Gram model is:

[0054]

[0055] In the formula, σ is the logistic function: σ(x)=1 / (1+e^(-x)), L represents the number of negative samples collected, and P v This represents the distribution of negative samples. and v q Both are vector representations of points of interest in the query space.

[0056] For the user-query interest graph G u-q The loss function of the Skip-Gram model is:

[0057]

[0058] In the formula, It is the vector representation of user i in the query space;

[0059] For the check-in interest points and the relationship diagram G p-p The loss function of the Skip-Gram model is:

[0060]

[0061] In the formula, All are vector representations of points of interest within the sign-in space.

[0062] Furthermore, in one embodiment, step 4, constructing the pairwise dataset, specifically involves:

[0063]

[0064] In the formula, Represents the set of points of interest that user i has checked in; (u i v j v k )∈D s This indicates that, for user i, his / her points of interest v that he / she has visited... j Preferences for unvisited points of interest are greater than those for previously visited points of interest.k It can also be represented as v j >u i，poi v k , where >u i，poi It satisfies 1) completeness, 2) antisymmetry, and 3) transitivity.

[0065] Furthermore, in one embodiment, step 4, which involves calculating the probability score of user-point-of-interest check-in pairs using a pairwise ranking model, specifically includes:

[0066] Point of interest v j and v k The ordering relationship between them can also be expressed as:

[0067]

[0068] In the formula, v j Indicates user u i Visited points of interest, v k Indicates points of interest that the user has not yet visited, x ij and x ik These represent the points of interest that users will check in at in the future. j and v k The score; x ij The calculation formula is as follows:

[0069]

[0070] In the formula, w represents the hyperparameter used to balance user query preferences and check-in preferences; and These are the vector representations of user i in the check-in and query spaces, and the interest point v, respectively. j Vector representation in the sign-in space; here, no vector representation is used. This is because it calculates the probability of a user querying a point of interest, not the probability of checking in at that point of interest. The specific formula for the loss function in BPR is:

[0071]

[0072] In the formula, D s For paired datasets, σ ​​is the logistic function: σ(x) = 1 / (1+e^(-x)).

[0073] Furthermore, in one embodiment, step 5 replaces the negative sampling portion in the Skip-Gram and pairwise ranking models with a negative sampling method based on a benchmark map. The specific process includes:

[0074] Step 5-1: Randomly select interest points from the interest point set V at a certain proportion τ. These interest points are called reference points, and the reference point set is defined as V.base ;

[0075] Step 5-2: Using the reference point as the center, record other points of interest within a radius r. Define that all points of interest in this area are geographically adjacent to each other, and construct a reference point map G based on this. base =(V, E) nei ), where V represents the set of points of interest, E nei For the edges in the baseline graph, the form is:

[0076] E nei ={(v i v j )|v base ∈V base ,dis(v i v base )<r,dis(v j v base )<r}

[0077] In the formula, V base This represents the set of reference points, and dis(x, y) represents the geographical distance between two points of interest.

[0078] Step 5-3, for the context relationship graph G p-p and G p-q Each edge (v) j v k If v j Edge E located in the baseline graph nei In the middle, then the set of neighboring points V of that point... j-nei ={v|(v j ,v)∈E nei In the context graph G, L negative samples are randomly selected; otherwise, random sampling is performed. u-q Define user i's historical check-in trajectory S i ={s1, s2, ... s n The center point of the trajectory is v. c v c The formula for calculating the coordinates is as follows:

[0079]

[0080]

[0081] In the formula, and These represent the latitude and longitude coordinates of the user's j-th check-in record, respectively.

[0082] definition Let be the set of negative sample points collected for user i whose distance from the center point of its trajectory does not exceed r. The definition is as follows:

[0083]

[0084] In the formula, dis(x, y) represents the distance between two points of interest;

[0085] Step 5-4: The loss function of the Skip-Gram model for the three context graphs in Step 3 can be rewritten as follows:

[0086]

[0087]

[0088]

[0089] In the formula, V j-nei Indicate points of interest The set of neighboring sample points, Indicates point of interest v neg Vector representation in the query space Indicates point of interest v neg Vector representation in the sign-in space;

[0090] Step 5-5, then the loss function of BPR in step 4 can be rewritten as:

[0091]

[0092] In the formula, D base The pairwise sorted training set, collected using a negative sampling method based on the benchmark map, is shown below:

[0093]

[0094] In the formula, V j-nei Indicates point of interest v j The set of neighboring sample points.

[0095] Furthermore, in one embodiment, the calculation formula for the joint learning framework described in step 6 is:

[0096]

[0097] In the formula, α, β, and γ are regularization parameters that control the contributions of the three context graphs, and the parameter set θ to be trained by the model includes U c V c U q y q, represent the vector representations of users and points of interest in the check-in space and query space, respectively, with the remaining symbols being the same as above.

[0098] Furthermore, in one embodiment, step 7 uses a pairwise ranking model to calculate a score indicating the likelihood of a user checking in at any point of interest in the future, and recommends points of interest to the user accordingly, specifically:

[0099] Step 7-1: Based on the user and interest point representation U learned through the joint learning framework in Step 6. c V c U q The following formula in step 4 is used to calculate the user i's future check-in interest points v. j Score:

[0100]

[0101] In the formula, w represents the hyperparameter used to balance user query preferences and check-in preferences; and They are user u i Vector representation of check-in and query space and interest point v j Vector representation in the sign-in space;

[0102] Step 7-2: After calculating the scores of all the points of interest that user i has not visited, sort them from highest to lowest score and recommend the points of interest with the highest score to the user.

[0103] In one embodiment, an interest point recommendation system based on user multi-behavior enhancement and efficient rich-information negative sampling is provided, the system comprising:

[0104] The first module is used to acquire user multi-behavior data and preprocess the data to exclude outlier data.

[0105] The second module is used to construct three types of contextual relationship graphs using user multi-behavior data: a relationship graph G between check-in interest points and query interest points. p-q User-Query Interest Relationship Graph G u-q And the relationship diagram G between check-in interest points and check-in interest points. p-p In the relationship graph, each node represents a user or point of interest.

[0106] The third module is used to mine user multi-behavior relationships in three types of context graphs using the Skip-Gram model, and to learn multi-view representations of users and points of interest.

[0107] The fourth module is used to construct the pairwise dataset, use the pairwise ranking model to calculate the probability score of each user-point of interest check-in pair, and calculate the BPR loss;

[0108] The fifth module is used to replace the negative sampling portion in Skip-Gram and pairwise sorting models with a negative sampling method based on the baseline plot;

[0109] The sixth module is used to design a joint learning framework that combines the loss functions of the Skip-Gram model and the BPR method to jointly learn the representations of users and points of interest in the check-in and query spaces.

[0110] The seventh module is used to calculate the probability score of a user checking in at any point of interest in the future using a pairwise ranking model, and to recommend points of interest to the user based on this score.

[0111] Specific limitations regarding the interest point recommendation system based on user multi-behavior enhancement and efficient rich-information negative sampling can be found in the limitations of the interest point recommendation method based on user multi-behavior enhancement and efficient rich-information negative sampling mentioned above, and will not be repeated here. Each module in the aforementioned interest point recommendation system based on user multi-behavior enhancement and efficient rich-information negative sampling can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0112] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0113] This embodiment uses a real-world city user check-in and user map query dataset for experiments. The check-in and query dataset includes 474,969 check-in records and 443,404 query records from July 1, 2018 to September 30, 2018. Each record includes user ID, check-in / query time, point of interest ID, point of interest category, point of interest location, and point of interest name. The dataset is divided chronologically, with the user check-in records from the last week used as the test set, the records from the second-to-last week used as the validation set, and the remaining data used as the training set. The learning rate is set to 0.001. Since the number of edges in the three context graphs is different, the data scale of each task is also different. To balance them, graph G... p-q And Figure G u-q Set the batch size to 128, and set the image G. p-p The batch size is 512. In the model, the latent representation dimension d for users and points of interest is set to 32. The model is optimized using Skip-Gram and BPR loss functions, minimizing them through backpropagation. The model's performance is evaluated using precision, recall, mean precision (MAP), and normalized depreciation cumulative gain (NDCG).

[0114] This embodiment selects different popular point-of-interest (POI) recommendation methods as comparison methods for comparative experiments: Weighted Regularized Matrix Factorization (WRMF), Bayesian Personalized Ranking (BPR), Personalized Top-N Sequence Recommendation Model Based on Convolutional Sequence Embedding (CASER), User Preference and Contextual Information Embedding Model (PACE), Graph Neural Network-Based Collaborative Filtering Method (NGCF), Disentangled Graph Collaborative Filtering Method (DGCF), Self-Attention Autoencoder Based on Proximity Influence Awareness (SAE-NAD), Collaborative Recommendation Incorporating Geographic Factor Influence (USG), Negative Sampling Model Based on Graph Collaborative Filtering Method (MixGCF), and Ultra Simplified Graph Convolutional Model (UltraGCN). The method of this invention is denoted as UMPRec, and the variant using ordinary random negative sampling in this model is denoted as UMPRec-r.

[0115] Tables 1 and 2 show the results of different methods for recommending points of interest on real datasets.

[0116] Table 1. Performance comparison of different methods on real datasets (Precision@k, Recall@k)

[0117]

[0118] Table 2. Performance comparison of different methods on real datasets (MAP@k, NDCG@k)

[0119]

[0120] As can be seen from Tables 1 and 2, the method proposed in this invention (UMPRec) performs best. From the tables, it can be seen that: (1) CASER and PACE may have poor performance due to the excessive sparsity of the data. In addition, the poor performance of PACE may also be due to the lack of effectiveness of PACE in extracting features from sparse data using multilayer perceptron (MLP); (2) Compared with BPR, WRMF performs better. The main reason is that the negative sampling technique enables BPR to effectively distinguish between positive and negative samples after the model is trained, making the pairwise sorting method superior to point-based optimization algorithms in implicit data; (3) Both USG and SAE-NAD algorithms consider the influence of geographical factors, so they have better performance than most of the comparison methods; (4) NGCF, DGCF and UltraGCN are based on GNN and collaborative filtering algorithms. The performance of the first two is similar, mainly due to the severe data scarcity and the fact that they both use similar GNN structures; UltraGCN performs better because it filters out redundant relationships between users and points of interest, preventing the addition of too much noise; (5) MixGCF also achieves better performance because it utilizes effective negative samples; (6) Finally, the UMPRec model of this invention outperforms other comparison methods in all indicators; at the same time, the comparison with the UMPRec-r variant shows the superiority of the key negative sample selection method based on the benchmark point map, proving that geographical distance is a very important influencing factor when users select points of interest.

[0121] To explore the impact of user check-in and query behavior on the performance of point of interest (POI) recommendation, this invention, in addition to the complete model "All" which considers all contextual relationships, sets up three variants of this invention: (1) the BPR method, which does not consider other contextual information and only uses user check-in data for POI recommendation; (2) WithPOI-POI, which only considers the contextual relationship graph G between POIs. p-p (3) WithQuery, which means only considering the context graph G related to the map query behavior. u-q And Figure G p-q It is worth noting that since both the UMPRec model and the UMPRec-r model consider multiple user behavior relationships and differ only in their negative sampling methods, the results of the two models are not significantly different. Therefore, the analysis results of the UMPRec model are mainly presented here. The final experimental results are as follows: Figure 2 , Figure 3 As shown.

[0122] Depend on Figure 2 and Figure 3It can be seen that, firstly, models employing multiple user behavior relationships consistently outperform the BPR method that does not consider any contextual information. Secondly, WithPOI-POI performs better than WithQuery because user check-in behavior clearly indicates user preferences for points of interest, while user map query behavior contains more noise and may not necessarily translate into check-in behavior. Finally, the complete model All proposed in this invention outperforms the three variants regardless of the value of k, i.e., when recommending k points of interest to the user. This demonstrates that extracting three contextual relationships from multiple user behaviors is highly effective, indicating a complementary relationship between user check-in and query behaviors, which can jointly promote point of interest recommendation.

[0123] In summary, the proposed interest point recommendation method based on enhanced user multi-behavior data and efficient, information-rich negative sampling fully utilizes user multi-behavior data to mine three types of contextual relationships, alleviating data scarcity and enhancing the performance of interest point recommendation. Simultaneously, the use of a negative sampling method based on a baseline point graph extracts more informative negative samples, enabling efficient training and learning of more accurate user and interest point representations. Finally, comparison with other related algorithms further verifies that the method of this invention can more accurately recommend user interests.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1.A method for point of interest recommendation based on multi-behavior enhancement and efficient informative negative sampling of users, characterized in that, The method comprises the following steps: Step 1, obtaining user multi-behavior data and pre-processing the data to exclude outlier data; Step 2, three kinds of context relationship graphs are respectively constructed by using user multi-behavior data: check-in interest point and query interest point relationship graph G p-q , user and query interest point relationship graph G u-q , and check-in interest point and check-in interest point relationship graph G p-p ; wherein each node in the relationship graph represents a user or an interest point; Step 3, using a Skip-Gram model to mine the user multi-behavior relationship in the three kinds of context relationship graphs and learning the multi-view representation of the user and the interest point; Step 4, constructing a pair data set, using a pair ranking model to calculate the possibility score of each user-interest point check-in pair and calculating the BPR loss; Step 5, replacing the negative sampling part in the Skip-Gram and pair ranking model with a negative sampling method based on the reference point graph; Step 6, designing a joint learning framework to combine the loss functions of the Skip-Gram model and the BPR method to jointly learn the representation of the user and the interest point in the check-in and query space; Step 7, using a pair ranking model to calculate the possibility score of the user checking in any interest point in the future and recommending the interest point to the user according to the score. 2.The method of claim 1, wherein, The user multi-behavior data in step 1 comprises: User interest point check-in data, including user ID, check-in time, interest point ID, interest point name, interest point category and interest point location; User map query data, including user ID, query time, interest point ID, interest point name, interest point category and interest point location. 3.The method of claim 1, wherein, In step 2, three kinds of context relationship graphs are constructed using user multi-behavior data, specifically: Step 2-1, dividing a day into four time periods: (1) morning period, i.e. 3:00 to 10:59, (2) afternoon period, i.e. 11:00 to 15:59, (3) evening period, i.e. 16:00 to 21:59, and (4) midnight period, i.e. 22:00 to 2:59; and combining workdays and weekends to obtain eight time periods; Step 2-2, constructing three kinds of context relationship graphs according to the time periods divided in step 2-1: The check-in interest point and the query interest point relationship graph G p-q Specifically, an edge is established between the interest points in which the check-in and query behaviors occur in the same day or the same time period, and is represented as Wherein represents the interest point v j In the representation of the check-in space, represents the interest point v k In the representation of the query space; the graph can be represented as G p-q =(V, E p-q , W p-q ), wherein V represents a set of interest points, E p-q represents a set of edges in the graph, and W p-q records the number of times of occurrence of the edges in the graph; The user and the query interest point relation graph G u-q Specifically: the edge between the user and the interest point queried by it in the same time period is established, denoted as Where u i represents the user i, represents the interest point v k The representation of the query space; the graph can be represented as G u-q =(U, V, E u-q , W u-q ), wherein U represents a user set, V represents an interest point set, E u-q represents a set of edges in the graph, and W u-q records the number of times the edge in the graph occurs; The check-in interest point and the check-in interest point relationship graph G p-p Specifically, an edge is established between the interest points to which the user checks in on the same day or in the same time period, denoted as Wherein And Respectively represent the interest point v j And v k In the representation of the check-in space; the graph can be represented as G p-p =(V, E p-p , W p-p ); wherein V represents a set of points of interest, E p-p represents a set of edges in the graph, W p-p records the number of times an edge in the graph occurs. 4.The method of claim 3, wherein, The Skip-Gram model used in step 3 is specifically: For step 2, the check-in interest point and the query interest point relationship graph G p-q The loss function of the Skip-Gram model is: where σ is the logistic sigmoid function: σ(x) = 1 / (1 + e^(-x)), L denotes the number of negative samples collected, P v denotes the distribution of negative samples collected, and v q are vector representations of points of interest in the query space; For a user and a query point-of-interest relationship graph G u-q , the loss function of the Skip-Gram model is: wherein is the vector representation of user i under the query space; For the check-in point of interest and the check-in point of interest relationship graph G p-p , the loss function of the Skip-Gram model is: wherein v c are vector representations of points of interest under the check-in space. 5.The method of recommending POI based on user multi-behavior enhancement and efficient information-rich negative sampling according to claim 4, characterized in that, The pair data set in step 4 is specifically: wherein, represents the set of interest points that user i has checked in; (u i , v j , v k ) ∈ D s represents that, for user i, its preference for the interest point v j that has been visited is greater than the interest point v k that has not been visited, which can also be expressed as v j > u i，poi v k , wherein > u i，poi satisfies 1) completeness, 2) anti-symmetry, and 3) transitivity. 6.The method of recommending POI based on user multi-behavior enhancement and efficient information-rich negative sampling according to claim 5, characterized in that, The use of a pair ranking model to calculate the possibility score of a user-interest point check-in pair in step 4 is specifically: Point of interest v j And v k The ordering relationship between the points of interest v and v is expressed as: where v j represents the interest points visited by user i, v k represents the interest points not visited by user i, x ij and x ik represent the scores of the interest points v j and v k that user i will check in in the future; and x ij is calculated as follows: where w is a hyper-parameter used to balance the user query preference and the check-in preference; and are the vector representations of user i in the check-in and query spaces, respectively, and j is the vector representation of interest point v in the check-in space. The loss function formula of BPR is specifically: where D s is the pair-wise dataset, and σ is the logistic function: σ(x) = 1 / (1 + e^(-x)). 7.The method of recommending POI based on user multi-behavior enhancement and efficient information-rich negative sampling according to claim 6, characterized in that, In step 5, the negative sampling part in the Skip-Gram and pair ranking model is replaced with a negative sampling method based on the reference point graph, and the specific process comprises: Step 5-1, randomly select interest points from the interest point set V with a certain proportion τ, called reference points, and define the reference point set as V base ; Step 5-2, record other interest points existing within a radius r centered on the reference point, define all interest points within this area have proximal relationship in geographical distance between each other, and build a reference point graph G with this base = (V, E nei ), where V represents the set of interest points, E nei is the edge in the reference point graph, in the form of: E nei = {(v i , v j ) | v base ∈ V base , dis(v i , v base ) < r, dis(v j , v base ) < r} wherein V base represents a set of reference points, dis(v i , v base ) represents a geographic distance between two points of interest v i , v base . Step 5-3, for the context relationship graph G p-p and G p-q Each edge (v) j v k If v j Edge E located in the baseline graph nei In the middle, then the set of neighboring points V of that point... j-nei ={v|(v j ,v)∈E nei In the context graph G, L negative samples are randomly selected; otherwise, random sampling is performed. u-q Define user i's historical check-in trajectory S i ={s1, s2, ... s n The center point of the trajectory is v. c v c The formula for calculating the coordinates is as follows: In the formula, and respectively represent the longitude and latitude coordinates of the jth check-in record of the user. Definitions Let S i be the set of negative sample points collected for user i that are within a distance r from the center point of their trajectory, i.e. is defined as follows: where dis(v, v c ) represents the distance between two points of interest v, v c , represents the user's check-in history data; Step 5-4, rewriting the loss function of the Skip-Gram model of the three kinds of context relationship graphs in step 3 as: where V j-nei represents a vector representation of a point of interest represents a set of neighboring sample points of a point of interest, represents a vector representation of a point of interest v neg in a query space, represents a vector representation of a point of interest v neg in a check-in space; Step 5-5, rewriting the loss function of BPR in step 4 as: In the formula, D base is a pair-wise ranking training set collected according to a negative sampling method based on a reference point diagram, as follows: In the formula, V j-nei represents a set of neighboring sample points of the interest point v j . 8.The method of recommending POI based on user multi-behavior enhancement and efficient information-rich negative sampling according to claim 7, characterized in that, The joint learning framework calculation formula in step 6 is: where a, b and g are regularization parameters controlling the contribution of the three context graphs, and the parameter set of the model to be trained, 0, includes U c , V c , U q , V q , respectively, represent the vector representation of users and points of interest in the check-in space and the query space. 9.The method of claim 8, wherein, Step 7 uses a pair ranking model to calculate the possibility score of the user checking in any interest point in the future and recommends the interest point to the user according to the score, specifically: Step 7-1: Based on the user and interest point representation U learned through the joint learning framework in Step 6. c V c U q The following formula in step 4 is used to calculate the user i's future check-in interest points v. j Score: where w is a hyper-parameter used to balance the user query preference and the check-in preference; and are the vector representations of user i in the check-in and query spaces, respectively, and is the vector representation of the point of interest in the check-in space. Step 7-2, after calculating the scores of all interest points that the user i has not visited, sorting them from high to low and recommending the interest point with the highest score to the user. Step 7-2, after calculating the scores of all interest points that the user i has not visited, sorting them from high to low and recommending the interest point with the highest score to the user.

Images