Point of interest recommendation method based on improved LSTM and location jumping

By improving the combination of LSTM and position jump algorithm, we can capture the user's long-term and short-term interest behavior dependence, solve the problem of insufficient accuracy of interest point recommendation in the existing technology, and achieve more accurate interest point recommendation.

CN114896481BActive Publication Date: 2024-11-26CHONGQING TONGQI FUTURE URBAN PLANNING & DESIGN INST CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210481261.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-05
Publication Date
2024-11-26
Estimated Expiration
2042-05-05

AI Technical Summary

Technical Problem

Existing LSTMs are not accurate enough to capture the user's long-term interest behavior dependence, and it is difficult to fully capture the user's interest point access preferences, especially non-subjective access noise, resulting in poor accuracy of interest point recommendations.

Method used

The improved LSTM and position jump algorithm are adopted to capture the user's long-term interest behavior dependence by combining the time similarity of the historical trajectory sequence and the current trajectory sequence, and to capture the user's short-term interest behavior dependence through the position jump algorithm of non-continuous non-adjacent interest points, integrating the two to improve the accuracy of recommendations for points of interest.

Benefits of technology

It effectively improves the comprehensiveness and accuracy of long-term and short-term interest behavior dependence, improves the accuracy of interest point recommendations, and can better handle non-subjective noise information in user access sequences.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114896481B_ABST
    Figure CN114896481B_ABST
Patent Text Reader

Abstract

The present invention specifically relates to a method for recommending points of interest based on improved LSTM and position jumping, including: obtaining a user's target trajectory sequence and inputting it into a recommendation model, and outputting a corresponding point of interest prediction result: firstly, dividing the target trajectory sequence into a historical trajectory sequence and a current trajectory sequence, and performing embedding processing; then, by using an improved LSTM capable of learning LSTM irrelevant context representation, capturing the user's long-term interest behavior dependency; then, by using a position jumping algorithm capable of learning non-continuous and non-adjacent points of interest of different importance, capturing the user's short-term interest behavior dependency; finally, fusing the user's long-term interest behavior dependency and the user's short-term interest behavior dependency, and then obtaining a corresponding point of interest prediction result; using the point of interest prediction result as the point of interest recommendation result for the corresponding user. The present invention can improve the accuracy and comprehensiveness of capturing the user's long-term and short-term interest behavior dependency, thereby improving the accuracy of the user's point of interest recommendation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of interest point recommendation, and in particular to an interest point recommendation method based on improved LSTM and position jumping. Background Art

[0002] With the rapid development of the Internet and the rapid popularization of mobile smart devices, people are paying more and more attention to their own location and the information of the surrounding geographical locations. Therefore, location-based service applications (LBS) have attracted a large number of users. Location-based service applications can quickly obtain the user's geographical location information through mobile smart devices and provide corresponding location services in a timely manner.

[0003] Points of interest (POI) refer to all geographical objects that can be abstracted as points, especially some geographical entities that are closely related to people's lives, such as shopping malls, parks, hotels, playgrounds, parking lots, etc. People can use mobile smart devices to share their location anytime and anywhere, check in at their favorite points of interest, and share their experiences at these points of interest on the Internet, which generates a large amount of interactive information and rich auxiliary information such as time, location, and comments. This information makes it possible to recommend points of interest to users in a personalized way.

[0004] Among them, the Chinese patent with publication number CN111949877A discloses "A personalized point of interest recommendation method and system", which mines the user's point of interest access preference by integrating external factors such as time, weather, and space through LSTM, and can recommend personalized points of interest to target users under specified time and weather conditions to help explore points of interest. However, the existing LSTM does not fully interact with the input and hidden states of multiple previous moments. LSTM is only associated with the input and previous hidden states at the current moment, making it difficult to fully capture the user's long-term interest behavior dependence, which in turn leads to poor accuracy in user point of interest recommendation. At the same time, the applicant found that there may be some non-subjective access noise in the user's access sequence (for example, the user visits a distant office building and surrounding points of interest due to work arrangements, which is not because the user is interested in these points of interest), making it difficult to accurately capture the user's short-term interest behavior dependence, which also leads to poor accuracy in user point of interest recommendation. Therefore, how to design a method that can improve the accuracy and comprehensiveness of capturing the user's long-term and short-term interest behavior dependence to improve the accuracy of user point of interest recommendation is a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: how to provide a method for recommending points of interest based on improved LSTM and position jumping to improve the accuracy and comprehensiveness of capturing the long-term and short-term interest behavior dependencies of users, thereby improving the accuracy of user point of interest recommendations.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] The POI recommendation method based on improved LSTM and location jumping includes the following steps:

[0008] S1: Obtain the user's target trajectory sequence;

[0009] S2: Input the user's target trajectory sequence into the trained recommendation model and output the corresponding POI prediction results;

[0010] The recommendation model first divides the target trajectory sequence into a historical trajectory sequence and a current trajectory sequence, and performs embedding processing; then, the improved LSTM that can learn LSTM irrelevant context representation is combined with the time similarity of the historical trajectory sequence and the current trajectory sequence to capture the user's long-term interest behavior dependency; then, the position jumping algorithm that can learn non-continuous and non-adjacent interest points of different importance is combined with the current trajectory sequence to capture the user's short-term interest behavior dependency; finally, the user's long-term interest behavior dependency and the user's short-term interest behavior dependency are integrated to obtain the corresponding interest point prediction results;

[0011] S3: Recommending points of interest to the corresponding user based on the points of interest prediction results output by the recommendation model.

[0012] Preferably, in step S2, embedding processing is implemented through context information features in four dimensions: user, point of interest, time and space;

[0013] The historical trajectory sequence embedding is represented as

[0014] The current trajectory sequence embedding is represented as

[0015] Preferably, in step S2, the user's long-term interest behavior dependency is generated through the following steps:

[0016] S201: Sequence the historical trajectory and the current trajectory sequence The input is improved in LSTM to retain the sequential dependency and obtain the correlation expression of the input sequence context information, and then output the corresponding historical trajectory hidden state and the current trajectory hidden state

[0017] S202: Hide the historical track and the current trajectory hidden state Input into the multi-head self-attention network, output the corresponding historical trajectory self-attention representation and the current trajectory self-attention representation

[0018] S203: Self-attention representation based on time weight combined with historical trajectory Calculate the target historical trajectory ts after considering the time factor h ;

[0019] S204: Self-attention representation of the current trajectory Perform average pooling to generate the target current trajectory representation ts n ;

[0020] S205: Based on each target historical trajectory ts h Current trajectory of the target ts n The influence of , calculate the corresponding user long-term interest behavior dependence

[0021] Preferably, in step S201, the hidden state h of LSTM is improved t =Mogrify(x,h t-1 );

[0022] Where: Mogrify(x,h t-1 ) represents an LSTM that is updated by considering the hidden state and output state at the previous time step;

[0023] in,

[0024] Where: x ↑ and respectively represent the highest index x in the alternating sequence j and c prev represents the cell state at the previous moment in LSTM; Mogrify represents improving the representation of LSTM; h t-1 Represents the hidden state of the previous moment of the improved LSTM;

[0025]

[0026] Where: x -1 = x and cycle Represents the hyperparameters, when r=0, it returns to the basic LSTM; and represents a trainable matrix; ⊙ represents the product of elements; σ represents the logical sigmoid function; oddj represents when j is an odd number in the alternating sequence; even j represents when j is an even number in the alternating sequence.

[0027] Preferably, in step S202, the following calculation is performed in the multi-head self-attention network:

[0028]

[0029] Where: yes or and After passing through the multi-head self-attention network, it is expressed as and Both represent the projection matrix in the i-th attention mechanism; head i represents the i-th head of multi-head self-attention; d represents Dimension of W i KT Represents the projection matrix W i K The transpose of .

[0030] Preferably, in step S203, the target historical trajectory

[0031] Where: Indicates the c and Tt The temporal similarity relationship between time slots;

[0032] in,

[0033] Where: W t represents the weight matrix; T i represents the set of check-in points of interest of the user at the i-th time point; T j Represents the set of check-in points of interest of the user at the jth time point.

[0034] Preferably, in step S204, the target current trajectory

[0035] Preferably, users’ long-term interest behavior dependence

[0036] Where: d n,h Indicates the location of the point of interest l t-1 With track TS h The geographical distance between

[0037] in,

[0038] Where: and represents the median coordinate of the target’s historical trajectory; W h Represents a trainable projection weight matrix; arcsin represents the inverse sine function in mathematics;

[0039] Preferably, in step S2, the user short-term interest behavior dependency is generated through the following steps:

[0040] S211: Construct distance matrix Store the distance between each point of interest; then calculate the point of interest closest to the current point of interest in the current trajectory sequence based on the distance matrix, and combine the deep traversal method to construct an improved LSTM input set at time t-1

[0041] S212: Input set Input into the improved LSTM to consider the impact of non-continuous and non-adjacent interest points on the user's short-term interest behavior dependence, and then output the corresponding current trajectory hidden state

[0042] in,

[0043] Where: x t-1 Represents input data; Represents the hidden state, and the superscript geo indicates that it is used in the position jumping algorithm.

[0044] S213: Hide the current track In the input attention mechanism layer, the current trajectory weighted vector representation corresponding to the output of the attention mechanism is To learn non-contiguous and non-adjacent points of interest with different importance, and then consider the non-subjective noise in the user's access sequence;

[0045] Attention mechanism layer,

[0046]

[0047] Where: express Output after passing through the neural network; express The transpose of w Represents the bias term in the neural network; W w and u w All represent randomly initialized matrices;

[0048] S214: Weighted vector representation based on current trajectory Calculate the corresponding user short-term interest behavior dependency

[0049] Where: W (1) 、b (1) , W (2) 、b (2) They all represent learnable parameters.

[0050] Preferably, in step S2, the user's long-term interest behavior dependency and the user's short-term interest behavior dependency are fused by the following formula:

[0051]

[0052] Where: R output Indicates the prediction result of interest points; Indicates the user's long-term interest behavior dependence; Indicates user's short-term interest behavior dependence; Represents the connection between the user's long-term interest behavior dependency and the user's short-term interest behavior dependency; A trainable projection matrix representing all interest points.

[0053] Compared with the prior art, the method for recommending points of interest in the present invention has the following beneficial effects:

[0054] The present invention captures the user's long-term interest behavior dependency through an improved LSTM capable of learning the LSTM irrelevant context representation, combining the time similarity of the historical trajectory sequence and the current trajectory sequence, that is, the irrelevant context of the LSTM can be learned by interacting with the input and hidden layers of the previous rounds, so that the LSTM can be associated with the input and hidden states of the previous moments, and then the user's long-term interest behavior dependency can be fully captured, thereby improving the comprehensiveness of capturing the user's long-term and short-term interest behavior dependency, thereby improving the accuracy of the user's interest point recommendation.

[0055] The present invention captures the user's short-term interest behavior dependency by combining the current trajectory sequence with a position jumping algorithm that can learn non-continuous and non-adjacent interest points of different importance, that is, it can consider the non-subjective access noise in the user's access sequence by learning non-continuous and non-adjacent interest points of different importance, and selectively obtain useful information, thereby accurately capturing the user's short-term interest behavior dependency, improving the accuracy of capturing the user's long-term and short-term interest behavior dependency, and thus can further improve the accuracy of user interest point recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to make the purpose, technical solution and advantages of the invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0057] Figure 1The logical block diagram of the method for recommending points of interest based on multi-context information;

[0058] Figure 2 This is a logical diagram of the operation of the recommended model;

[0059] Figure 3 Schematic diagram for improving the operation logic of LSTM. DETAILED DESCRIPTION

[0060] The following is a further detailed description through specific implementation methods:

[0061] Example:

[0062] This embodiment discloses a method for recommending points of interest based on improved LSTM and position jumping.

[0063] like Figure 1 As shown, the POI recommendation method based on improved LSTM and position jumping includes the following steps:

[0064] S1: Obtain the user's target trajectory sequence;

[0065] S2: Input the user's target trajectory sequence into the trained recommendation model and output the corresponding POI prediction results;

[0066] like Figure 2 As shown in the figure, the recommendation model first divides the target trajectory sequence into the historical trajectory sequence and the current trajectory sequence, and performs embedding processing; then, the improved LSTM that can learn the LSTM irrelevant context representation is combined with the time similarity of the historical trajectory sequence and the current trajectory sequence to capture the user's long-term interest behavior dependency; then, the position jumping algorithm that can learn non-continuous and non-adjacent interest points of different importance is combined with the current trajectory sequence to capture the user's short-term interest behavior dependency; finally, the user's long-term interest behavior dependency and the user's short-term interest behavior dependency are integrated to obtain the corresponding interest point prediction results;

[0067] S3: Recommending points of interest to the corresponding user based on the points of interest prediction results output by the recommendation model.

[0068] The present invention captures the user's long-term interest behavior dependency by combining the time similarity of the historical trajectory sequence and the current trajectory sequence through an improved LSTM capable of learning the LSTM irrelevant context representation, that is, it can learn the irrelevant context of the LSTM through interaction with the input and hidden layers of the previous rounds, so that the LSTM can be associated with the input and hidden states of the previous moments, and then can fully capture the user's long-term interest behavior dependency, improve the comprehensiveness of capturing the user's long-term and short-term interest behavior dependency, and thus can improve the accuracy of user interest point recommendation. At the same time, the present invention captures the user's short-term interest behavior dependency by combining the current trajectory sequence through a position jump algorithm capable of learning non-continuous and non-adjacent interest points of different importance, that is, it can consider the non-subjective access noise in the user's access sequence by learning non-continuous and non-adjacent interest points of different importance, and selectively obtain useful information, and then can accurately capture the user's short-term interest behavior dependency, improve the accuracy of capturing the user's long-term and short-term interest behavior dependency, and thus can further improve the accuracy of user interest point recommendation.

[0069] In this embodiment, the user set and the interest point set are respectively represented as U = {u1,u2,…,u U} and L={l1,l2,…,l L}; The geographical location of each point of interest is represented by (lon l ,lat l ) tuple encoding.

[0070] For each user u in u∈U, its target trajectory sequence is expressed as TS={TS1,TS2,…,TS n}; where n is the index of the most recent sequence, each trajectory TS i ∈TS consists of a series of points of interest visited by the user in a continuous order according to the user's real check-in sequence,

[0071] Based on the target user's trajectory sequence, the historical trajectory sequence TS of user u is h ={TS1,TS2,…,TS n-1};

[0072] For the target user u∈U, its most recent trajectory is the current trajectory sequence Among them, L t-1 Represents the points of interest recently visited by user u.

[0073] During the specific implementation process, the multi-feature embedding module is used to embed context information features from four dimensions: user, point of interest, time and space.

[0074] Simple models such as collaborative filtering can only represent the transition between independent states such as location and time. The choice of the next point of interest of the target user is affected by many factors such as long-term and short-term behavior dependencies, different days of the week, different time periods of the day, and the location of the point of interest. Therefore, after embedding each context information feature, the present invention performs a connection operation on it to obtain the hidden representation of each trajectory point (such as Figure 2 shown at the bottom).

[0075] Among them, the point of interest set

[0076] The historical trajectory sequence (encoded) is embedded as

[0077] The current trajectory sequence (encoded) is embedded as

[0078] The target trajectory sequence (encoded) is embedded as

[0079] During the specific implementation process, the user's long-term interest behavior dependency is generated through the long-term behavior dependency module.

[0080] The most intuitive way to capture the long-term behavior dependency of users is to selectively use more prominent information from the input sequence of user behavior trajectories. Therefore, the present invention uses all the behavior trajectory sequences of the target user to obtain the long-term behavior dependency of the user. Figure 2 As shown in FIG. 1 , this is inspired by a good improved LSTM. The present invention uses an improved deep network to capture the user's interest behavior dependency. The improved LSTM of the present invention is selected as follows Figure 3 The mogrifier LSTM shown.

[0081] The specific steps include:

[0082] S201: Sequence the historical trajectory and the current trajectory sequence The input is improved in LSTM to retain the sequential dependency and obtain the correlation expression of the input sequence context information, and then output the corresponding historical trajectory hidden state and the current trajectory hidden state

[0083] Improve the hidden state h of LSTM t =Mogrify(x,h t-1 );

[0084] Where: Mogrify(x,h t-1 ) represents an LSTM that is updated by considering the hidden state and output state at the previous time step;

[0085] in,

[0086] Where: x ↑ and respectively represent the highest index x in the alternating sequence j and c prev represents the cell state at the previous moment in LSTM; Mogrify represents improving the representation of LSTM; h t-1 Represents the hidden state of the previous moment of the improved LSTM;

[0087]

[0088] Where: x -1 = x and cycle Represents the hyperparameters, when r=0, it returns to the basic LSTM; and represents a trainable matrix; ⊙ represents the product of elements; σ represents the logical sigmoid function; oddj represents when j is an odd number in the alternating sequence; even j represents when j is an even number in the alternating sequence.

[0089] S202: Hide the historical track and the current trajectory hidden state Input into the multi-head self-attention network, output the corresponding historical trajectory self-attention representation and the current trajectory self-attention representation

[0090] The long-term behavior of users has a certain periodicity, and it is necessary to consider interest point sets with different degrees of significance in the historical sequence. In order to reduce the amount of calculation, the output after the mogrifier LSTM is placed in the multi-head attention network first. The present invention uses a multi-head attention mechanism to process them.

[0091] The following calculations are performed in the multi-head self-attention network:

[0092]

[0093] Where: yes or and After passing through the multi-head self-attention network, it is expressed as and Both represent the projection matrix in the i-th attention mechanism; head i represents the i-th head of multi-head self-attention; d represents Dimension of W i KT Represents the projection matrix W i K The transpose of .

[0094] S203: Self-attention representation based on time weight combined with historical trajectory Calculate the target historical trajectory ts after considering the time factor h ;

[0095] In real life, the popularity of points of interest will change greatly due to time changes. For example, people may look for restaurants at lunch time and then go to entertainment venues in the evening. Therefore, in the time dimension, the present invention develops a time-weighted operation. Specifically, the present invention arranges the weekly trajectories into 48 time periods, 24 time periods on weekdays, and 24 time periods on weekends (we believe that the user's weekday behavior is roughly periodic, and the weekend trajectory can better reflect the user's behavioral dependence). For each slot i, we consider a set of points of interest The points of interest that the user has not visited are filtered out. Then the temporal similarity relationship Sim between the i-th and j-th time slots is calculated by cosine similarity i,j , as shown below:

[0096]

[0097] Where: W t represents the weight matrix, that is, according to the time similarity calculated between time slots, different weights are assigned to the user interest behavior dependencies at different times based on the time factor; T i represents the set of check-in points of interest of the user at the i-th time point; T j Represents the set of check-in points of interest of the user at the jth time point.

[0098] Intuitively speaking, the more the points of interest overlap between two time periods, the more similar the user's interest behavior dependencies may be in these two time periods. Then, when recommending to users, the weight of the points of interest visited by users who are more similar to the target user in the same time period will be increased:

[0099] Target history track

[0100] Where: Indicates the c and Tt The temporal similarity relationship between time slots.

[0101] S204: Self-attention representation of the current trajectory Perform average pooling to generate the target current trajectory representation ts n ;

[0102] After considering the temporal context information, the target history trajectory sequence is expressed as ts h ∈{ts1,ts2,…,ts n-1 The present invention only performs time-weighted operation on the target historical trajectory sequence.

[0103] For the current trajectory sequence TS of the target n , which can better represent the user's real-time interest behavior dependence, and average pooling saves all interest point information in TS n It is more conducive to modeling the long-term and short-term behavior dependencies of users:

[0104] Target current trajectory

[0105] S205: Based on each target historical trajectory ts h Current trajectory of the target ts n The influence of , calculate the corresponding user long-term interest behavior dependence

[0106] In the spatial dimension, it is understandable that the next POI recommendation will be affected by the geographical environment of the target user's real-time trajectory. If the user updates the check-in record of other cities, the POI sets of other cities should be recommended based on the user behavior dependency. Therefore, inspired by LSTPM and non-local networks in computer vision tasks, the solution of the present invention considers each historical trajectory sequence TS h ∈{TS1,TS2,…,TS n-1} for the current trajectory sequence TS n We construct the geo-local operation layer through the following equation:

[0107] User long-term interest behavior dependence

[0108] Where: d n,h Indicates the location of the point of interest l t-1 With track TS h The geographical distance between

[0109] in,

[0110] Where: and represents the median coordinate of the target’s historical trajectory; W h Represents a trainable projection weight matrix; arcsin represents the inverse sine function in mathematics; Represents the aggregation of the hidden states of the user's long-term interest behavior dependencies and the most current points of interest based on the trajectory.

[0111] The present invention improves LSTM to process historical trajectory sequences and current trajectory sequences, so that the sequential dependency can be retained and the correlation expression of the context information of the input sequence can be better obtained, and then the irrelevant context of LSTM can be learned by interacting with the input and hidden layers of the previous rounds, so that LSTM can better associate with the input and hidden states at each moment, thereby more comprehensively capturing the user's long-term interest behavior dependency. Secondly, the present invention uses the output of the improved LSTM as the input of the multi-head self-attention network to generate an attention representation, so that the amount of calculation in the process of capturing the user's long-term interest behavior dependency can be reduced. Finally, the present invention calculates the time weight of the interest point by considering the popularity of the interest point caused by time changes, obtains the user's real-time interest behavior dependency by average pooling, and then generates the user's long-term interest behavior dependency by combining the impact of each target historical trajectory on the target's current trajectory, which can more accurately and comprehensively capture the user's long-term interest behavior dependency, thereby further improving the accuracy of the user's interest point recommendation.

[0112] During the specific implementation process, the user's short-term interest behavior dependency is generated through the short-term behavior dependency module.

[0113] When acquiring long-term behavior dependencies, the present invention considers the current trajectory sequence TS of the target user. n However, the current trajectory sequence has limitations in capturing the long-term behavior dependency of the target user. One of the reasons is that the trajectory sequence of the target user TS = {TS1, TS2, …, TS n}Too many and long, long-term behavior dependencies are the general interests of users, and the current trajectory sequence only accounts for a small part, so the impact is more limited than expected. On the contrary, for obtaining the short-term behavior dependencies of users, the current trajectory sequence TS n Therefore, the present invention proposes a position hopping algorithm which considers the non-subjective access noise in the user's access sequence by learning non-continuous and non-adjacent points of interest.

[0114] The specific steps include:

[0115] S211: Construct distance matrix Store the distance between each point of interest; then calculate the point of interest closest to the current point of interest in the current trajectory sequence based on the distance matrix, and combine the deep traversal method to construct an improved LSTM input set at time t-1 in,

[0116] S212: Input set Input into the improved LSTM to consider the impact of non-continuous and non-adjacent interest points on the user's short-term interest behavior dependence, and then output the corresponding current trajectory hidden state

[0117] in,

[0118] Where: x t-1 Represents input data; Represents the hidden state, and the superscript geo indicates that it is used in the position jumping algorithm;

[0119] S213: Hide the current track In the input attention mechanism layer, the current trajectory weighted vector representation corresponding to the output of the attention mechanism is To learn non-contiguous and non-adjacent points of interest with different importance, and then consider the non-subjective noise in the user's access sequence;

[0120] To consider non-adjacent spatial factors with different levels of importance, we add an attention mechanism layer to the attention network:

[0121]

[0122] Where: express Output after passing through the neural network; express The transpose of w Represents the bias term in the neural network; W w and u w All represent randomly initialized matrices;

[0123] S214: Weighted vector representation based on current trajectory Calculate the corresponding user short-term interest behavior dependency

[0124] Where: W (1) 、b (1) , W (2) 、b (2) They all represent learnable parameters.

[0125] The present invention learns non-continuous and non-adjacent points of interest with different degrees of importance through a position jumping algorithm, and then considers the non-subjective access noise in the user's access sequence, and selectively obtains useful information, so that the user's short-term behavior dependency can be captured more accurately, thereby improving the comprehensiveness of capturing the user's long-term and short-term interest behavior dependency. At the same time, the present invention considers non-adjacent spatial factors with different degrees of importance by adding an attention mechanism layer, thereby further improving the accuracy of capturing the user's short-term behavior dependency.

[0126] In the specific implementation process, the following formula is used to integrate the user's long-term interest behavior dependency and the user's short-term interest behavior dependency;

[0127]

[0128] Where: R output Indicates the prediction result of interest points; Indicates the user's long-term interest behavior dependence; Indicates user's short-term interest behavior dependence; Represents the connection between the user's long-term interest behavior dependency and the user's short-term interest behavior dependency; A trainable projection matrix representing all interest points.

[0129] The objective function when training the recommendation model is as follows:

[0130]

[0131] Where: N represents the total number of training samples, p i Represents the probability of the true interest point generated by the recommendation model for the i-th training sample.

[0132] In order to better illustrate the advantages of the technical solution of the present invention, the following experiments are disclosed in this embodiment.

[0133] 1. Experimental purpose

[0134] The purpose of our experiment is to answer the following questions:

[0135] RQ1: How does the performance of the STMLA model (i.e., the recommendation model of the present invention) compare with other state-of-the-art recommendation algorithms?

[0136] RQ2: What is the impact of key components on the performance of the STMLA model?

[0137] 2. Dataset

[0138] This experiment conducts comprehensive experiments on our proposed STMLA model on two real LBSNs datasets, Foursquare and Gowalla2. These two datasets have been widely used in the field of next POI (i.e., point of interest) recommendation.

[0139] The Foursquare dataset was collected in New York from February 2010 to January 2011, and the Gowalla dataset was collected worldwide from February 2009 to October 2010. They cover a large number of check-ins (user ID, POI ID, time, location, etc.). In both datasets, in order to take into account the impact of data sparsity, unpopular POIs with less than 10 users and less than 10 POIs are deleted. Target user trajectories with less than 3 check-ins in a day and target user trajectories with no more than 5 total check-in trajectories will also be deleted.

[0140] 3. Evaluation Metrics

[0141] We use two commonly used evaluation metrics, Recall@N and Normalized Discounted Cumulative Gain (NDCG@N), to evaluate the recommendation performance, where N∈{5,10}. Recall@N is used to evaluate the accuracy of model predictions, and NDCG@N is used to evaluate the performance of the model. 2) Baseline: We use Python's Pytorch library to implement our STMLA model and compare it with the following baseline algorithms.

[0142] 3.1 Baseline Model:

[0143] 1) LSTM: A variant of RNN that can capture users’ long-term and short-term interest behavior dependencies.

[0144] 2) ST-RNN: This is a RNN model that utilizes temporal and spatial transition matrices to model the local spatiotemporal context of each layer.

[0145] 3) Time-LSTM: The time gate is used to model time intervals and is combined with LSTM to capture the long-term and short-term interest behavior dependencies of users.

[0146] 4) DRCF: This model exploits the dot product of latent factors to model user-place interactions and uses a deep recurrent collaborative filtering framework to capture users’ long-term and short-term interest behavior dependencies.

[0147] 5) DeepMove: This model divides the user's check-in sequence into historical trajectories and current trajectories, and uses GRU and attention modules to capture the user's long-term and short-term interest behavior dependencies.

[0148] 6) STGN: This method is a variant of LSTM, which adds two pairs of time gates and distance gates to jointly control the update of user interest behavior dependencies.

[0149] 7) LSTPM: is a state-of-the-art model that captures long-term interest behavior dependencies through a two-layer non-local network and short-term interest behavior dependencies through a geographically extended LSTM network.

[0150] 3.2 Parameter settings:

[0151] In the experiment, 80% of each dataset is randomly selected as the training set, and the remaining 20% ​​is used as the test set. The inspiration for dividing the user's check-in sequence into historical tracks and current tracks comes from DeepMove. According to the DeepMove model, the public hyperparameters shared by all models are set: the dimension of the hidden state and embedding is 500, all parameters are optimized using Adam, the batch size is set to 32, and the learning rate is set to 0.0001. The number of heads of multi-head attention is set to 2.

[0152] The experimental results are shown in Table 1.

[0153] Table 1 Performance comparison of Recall@N and NDCG@N between models on Gowalla and Foursquare datasets

[0154]

[0155] 4. About RQ1

[0156] As shown in Table 1, on the Foursquare and Gowalla datasets, the STMLA algorithm is comprehensive and significantly outperforms other baseline models. Among them, compared with the best baseline LSTPM algorithm, the average performance improvement of the STMLA algorithm on the Foursquare and Gowalla datasets is 4.83% and 33.67% respectively. With all parameter settings unified, the results of the STMLA model fully demonstrate the remarkable effectiveness of the algorithm.

[0157] Among all the baselines, the results of LSTPM, DeepMove and DRCF algorithms are better than other baselines. Because they capture the long-term and short-term user interest behavior dependencies in different ways, which strongly shows that although a lot of research focuses on the dynamic capture of short-term user interest behavior dependencies, it is very critical to capture the long-term user interest behavior dependencies at the same time. From different perspectives, LSTPM, DeepMove and DRCF algorithms have captured the long-term periodicity of users, which well demonstrates the importance of time factors for POI recommendation. In addition, we also solve the problem of insufficient contextual interaction of LSTM algorithm and the problem of different emphasis on non-adjacent and non-continuous POIs, which is the key to determine the advantages of STMLA algorithm over other baselines.

[0158] Since Gowalla’s data is collected worldwide and the number of POIs far exceeds the number of users, the Gowalla dataset is very sparse. We can observe that all models perform lower on Gowalla than on Foursquare. However, our STMLA has a higher improvement on the Gowalla dataset, which also illustrates the importance of considering non-contiguous and non-adjacent POIs with different levels of importance when facing the very sparse Gowalla dataset. It is worth noting that the DeepMove algorithm has the second best performance on the Foursquare baseline. Unfortunately, the DeepMove algorithm does not include spatial information, resulting in poor performance on the Gowalla dataset. This also well illustrates the importance of spatial factors for POI recommendation.

[0159] 5. About RQ2

[0160] To verify the impact of each of our main components on the performance of STMLA, we divide the STMLA-based model into five simplified versions for ablation experiments.

[0161] 1) STML: Only the attention network is omitted, and the long-term and short-term interest behavior dependencies of users are captured based on the improved LSTM module.

[0162] 2) STLA: omitting the improved LSTM module, using ordinary LSTM, and capturing the user’s long-term and short-term interest behavior dependencies based on the attention network

[0163] 3) STMLL-S: Based on the improved LSTM module, the position jumping algorithm is used to capture short-term user interest behavior dependencies.

[0164] 4) STMLA_L: Based on the improved LSTM module and attention network, the component that captures short-term interest behavior dependency is omitted, and only the long-term interest behavior dependency of users is considered.

[0165] 5) STMLA_S: Based on the improved LSTM module and attention network, it omits the components that capture long-term interest behavior dependencies and only considers users’ short-term interest behavior dependencies.

[0166] The experimental results are shown in Table 2.

[0167] Table 2 Comparison of the impact of key components of the STMLA model on performance

[0168]

[0169] As shown in Table 2, the performance of STMLA_L and STMLA_S algorithms on both datasets is not satisfactory. The experimental results show that it is not enough to only consider capturing long-term interest behavior dependency or short-term interest behavior dependency (even though STMLAS outperforms the best baseline), both of which are very important for the next POI recommendation.

[0170] Compared with the STMLA algorithm, the performance of the STML and STLA algorithms on the two datasets is not outstanding because they cannot capture both long-term and short-term interest behavior dependencies. However, compared with the best baseline LSTPM algorithm, especially on the Gowalla dataset, the two simplified models have significant improvements. This once again proves the effectiveness of the improved LSTM module and attention network. Since the Gowalla dataset is relatively sparse, the performance of the STLA algorithm is better than that of the STML algorithm.

[0171] Since the contextual interactions of the LSTM modules are fully considered, the performance of STMLL-S shows that the location hopping algorithm has a significant contribution to the STMLA algorithm by comparison with the STML algorithm. This is because we automatically determine the relevant non-contiguous and non-adjacent POIs of the input in a dynamic manner according to the user's current check-in sequence, and assign higher weights to the nearest POIs, which is undoubtedly meaningful. From another perspective, the performance of the STMLL-S algorithm is lower than that of the STMLA algorithm, which shows that even if the user's behavior has long-term periodicity, it is necessary to consider POIs with different levels of importance in the historical sequence. In addition, the performance of both STMLAL and STMLAS algorithms is worse than that of the STMLL-S algorithm, because the STMLL-S algorithm is a simplified version that considers both long-term and short-term user behavior dependencies. In general, each key component of the STMLA algorithm has a positive impact on the user's personalized selection of the next POI.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A method for recommending points of interest based on improved LSTM and position jumping, characterized in that: The following steps are involved: S1: Obtain the user's target trajectory sequence; S2: Input the user's target trajectory sequence into the trained recommendation model and output the corresponding POI prediction results; The recommendation model first divides the target trajectory sequence into a historical trajectory sequence and a current trajectory sequence, and performs embedding processing; then, the improved LSTM that can learn LSTM irrelevant context representation is combined with the time similarity of the historical trajectory sequence and the current trajectory sequence to capture the user's long-term interest behavior dependency; then, the position jumping algorithm that can learn non-continuous and non-adjacent interest points of different importance is combined with the current trajectory sequence to capture the user's short-term interest behavior dependency; finally, the user's long-term interest behavior dependency and the user's short-term interest behavior dependency are integrated to obtain the corresponding interest point prediction results; In step S2, embedding is performed through the context information features of four dimensions: user, point of interest, time and space; The historical trajectory sequence embedding is represented as TS h ={TS1,TS2,…,TS |n-1| } represents the historical trajectory sequence, n is the index of the most recent sequence; The current trajectory sequence embedding is represented as Generate user long-term interest behavior dependency through the following steps: S201: Sequence the historical trajectory and the current trajectory sequence The input is improved in LSTM to retain the sequential dependency and obtain the correlation expression of the input sequence context information, and then output the corresponding historical trajectory hidden state and the current trajectory hidden state S202: Hide the historical track and the current trajectory hidden state Input into the multi-head self-attention network, output the corresponding historical trajectory self-attention representation and the current trajectory self-attention representation S203: Self-attention representation based on time weight combined with historical trajectory Calculate the target historical trajectory ts after considering the time factor h ; S204: Self-attention representation of the current trajectory Perform average pooling to generate the target current trajectory representation ts n ; S205: Based on each target historical trajectory ts h Current trajectory of the target ts n The influence of , calculate the corresponding user long-term interest behavior dependence In step S201, the hidden state h of LSTM is improved t =Mogrify(x,h t-1 ); Where: Mogrify(x,h t-1 ) represents an LSTM that is updated by considering the hidden state and output state at the previous time step; in, Where: x ↑ and respectively represent the highest index x in the alternating sequence j and c prev represents the cell state at the previous moment in LSTM; Mogrify represents improving the representation of LSTM; h t-1 Represents the hidden state of the previous moment of the improved LSTM; Where: x -1 = x and cycle Represents the hyperparameters, when r=0, it returns to the basic LSTM; and represents a trainable matrix; ⊙ represents the product of elements; σ represents the logical sigmoid function; odd j represents when j is an odd number in the alternating sequence; even j represents when j is an even number in the alternating sequence; S3: Recommending points of interest to the corresponding user based on the points of interest prediction results output by the recommendation model.

2. The method for recommending points of interest based on improved LSTM and position jumping as claimed in claim 1, characterized in that: In step S202, the following calculation is performed in the multi-head self-attention network: Where: yes or and After passing through the multi-head self-attention network, it is expressed as and Both represent the projection matrix in the i-th attention mechanism; head i represents the i-th head of multi-head self-attention; d represents Dimensions; Represents the projection matrix The transpose of .

3. The method for recommending points of interest based on improved LSTM and position jumping as claimed in claim 2, characterized in that: In step S203, the target historical trajectory Where: Indicates the c and Tt The temporal similarity relationship between time slots; in, Where: W t represents the weight matrix; T i represents the set of check-in points of interest of the user at the i-th time point; T j Represents the set of check-in points of interest of the user at the jth time point.

4. The method for recommending points of interest based on improved LSTM and position jumping as claimed in claim 2, characterized in that: In step S204, the target current trajectory 5. The method for recommending points of interest based on improved LSTM and position jumping as claimed in claim 4, characterized in that: In step S205, the user's long-term interest behavior depends on Where: d n,h Indicates the location of the point of interest l t-1 With track TS h The geographical distance between in, Where: and represents the median coordinate of the target’s historical trajectory; W h Represents a trainable projection weight matrix; arcsin represents the inverse sine function in mathematics; 6. The method for recommending points of interest based on improved LSTM and position jumping as claimed in claim 5, characterized in that: In step S2, the user short-term interest behavior dependency is generated through the following steps: S211: Construct distance matrix Store the distance between each point of interest; Then, according to the distance matrix, the point of interest closest to the current point of interest in the current trajectory sequence is calculated, and the input set of the improved LSTM at time t-1 is constructed by combining the deep traversal method. S212: Input set Input into the improved LSTM to consider the impact of non-continuous and non-adjacent interest points on the user's short-term interest behavior dependence, and then output the corresponding current trajectory hidden state in, Where: x t-1 Represents input data; Represents the hidden state, and the superscript geo indicates that it is used in the position jumping algorithm; S213: Hide the current track In the input attention mechanism layer, the current trajectory weighted vector representation corresponding to the output of the attention mechanism is To learn non-contiguous and non-adjacent points of interest with different importance, and then consider the non-subjective noise in the user's access sequence; Attention mechanism layer, Where: express Output after passing through the neural network; express The transpose of w Represents the bias term in the neural network; W w and u w All represent randomly initialized matrices; S214: Weighted vector representation based on current trajectory Calculate the corresponding user short-term interest behavior dependency Where: W (1) 、b (1) , W (2) 、b (2) They all represent learnable parameters.

7. The method for recommending points of interest based on improved LSTM and position jumping as claimed in claim 6, characterized in that: In step S2, the user's long-term interest behavior dependency and the user's short-term interest behavior dependency are fused by the following formula; Where: R output Indicates the prediction result of interest points; Indicates the user's long-term interest behavior dependence; Indicates user's short-term interest behavior dependence; Represents the connection between the user's long-term interest behavior dependency and the user's short-term interest behavior dependency; A trainable projection matrix representing all interest points.