A point of interest recommendation method combining user periodic preferences and short-term interests

Abstract

The present application belongs to the field of social network interest point recommendation, and particularly relates to an interest point recommendation method combining user periodic preference and short-term interest, comprising: data processing, processing original data into input data required by the model; extracting a non-continuous time-space interval matrix of interest point history according to user access to the interest point and combining the access history track of the interest point type; inputting the user history track access and the non-continuous time-space interval matrix into a network structure based on deep learning to obtain the behavior preference of the user; selecting a recommended interest point in a candidate interest point list according to the behavior preference of the user to obtain an ordered recommendation set; selecting the optimal recommended interest point in the ordered recommendation set; and the method is helpful for a merchant to master the user behavior characteristic distribution, analyze potential users and obtain good economic and social benefits.

Description

Technical Field

[0001] This invention belongs to the field of social network interest recommendation, specifically involving an interest recommendation method that combines users' periodic preferences and short-term interests. Background Technology

[0002] With the rapid proliferation of mobile devices in daily life, numerous location-based social services have emerged. People can share their life experiences and post check-in records on location-based social networks (LBSNs) such as Foursquare, Yelp, Facebook, and Brightkite. In China, LBSN services like Dianping, Meituan, Ele.me, and Weibo have experienced rapid growth. Statistics show that Meituan's user transaction volume increased by 12.5% ​​in 2019, with millions of reviews for services such as restaurants, hotels, and travel agencies. The rapid development of LBSNs has led to a surge in merchant and user data, presenting a significant opportunity for both users and service providers. Analyzing and mining massive amounts of LBSN data can help service providers recommend precise Points of Interest (POIs) to users. Next Point of Interest (POI) recommendation can predict the points of interest a user is most likely to visit at a given time. Its time-sensitive application makes POI recommendation a hot research topic in the recommendation field. Points of interest (POI) recommendations not only bring economic benefits to LBSN service providers but also help users uncover potential preferences, thereby adding enjoyment and convenience to their lives. For users, accurate and effective POI recommendations help them understand their living environment and make informed travel decisions, enhancing their enjoyment of life and understanding of their city. For businesses, providing accurate POI recommendations can generate economic benefits, increase platform visibility, and also enable personalized services to improve user experience. Furthermore, it allows service providers to analyze users' individual preferences and identify popular POI areas for targeted marketing.

[0003] In recent years, location-based social network point-of-interest (POI) recommendation methods have gained increasing popularity in the Internet of Things (IoT) and e-commerce. Among these methods, the idea of ​​mining users' short-term and long-term preferences plays a crucial role in improving user experience and service quality. Existing POI recommendation technologies can be categorized into collaborative filtering, matrix factorization, Markov chain-based methods, and deep learning-based methods. Collaborative filtering-based methods start with users or POIs, calculating the similarity between users or POIs before recommending POIs to the target user. Matrix factorization-based methods primarily construct a check-in matrix based on users' check-in records for POIs, mining user preferences and interest attributes through matrix factorization. Markov chain-based methods model user access POI trajectories as sequential records, assuming that the next POI visit is related to the previous one, modeling and mining sequential patterns. Deep learning-based POI recommendation methods integrate various contextual information, such as time, space, POI categories, and social relationships, constructing relevant structures like recurrent neural network models and introducing attention mechanisms to mine user interests and POI attributes, thereby recommending POIs that users may visit.

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

[0005] 1. Lack of consideration for exploring users' periodic preferences combined with interest categories: Current research methods often focus on the periodic interest of a specific interest point when studying users' periodic interests. However, users' periodic preferences combined with interest categories refer to users' fixed access preferences for a certain type of interest point or a specific interest point. In addition, when users' access goals are not clear, they may only consider a certain type of interest point instead of a specific interest point.

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

[0007] 3. The spatiotemporal intervals of recent visits can reflect users' short-term preferences. However, since the spatiotemporal interval values ​​are a set of continuous values, conventional methods such as the bucketing method require reasonable threshold settings to ensure the effectiveness of the model. Therefore, embedding continuous spatiotemporal interval values ​​is difficult. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for recommending points of interest that combines users' periodic preferences and short-term interests, comprising:

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

[0010] S2: Extract relevant attributes from the new dataset to generate check-in records for user visits to points of interest; the relevant attributes include the time information of user visits to points of interest, the category of points of interest, and latitude and longitude coordinates;

[0011] S3: Based on the user's check-in records for visiting points of interest, extract the spatiotemporal trajectory representation and discontinuous spatiotemporal interval matrix of the user's visits to points of interest through a sliding time window;

[0012] S4: Based on the periodic preference representation of users combined with interest categories, extract users' higher-order short-term interests by embedding a discontinuous spatiotemporal interval matrix using linear interpolation based on the spatiotemporal interval matrix.

[0013] S5: Based on the user's periodic preference for interest categories and the user's higher-order short-term interests, recommend interest points from the candidate interest point list to obtain an ordered recommendation set;

[0014] S6: Select the interest point with the highest probability predicted by the model from the ordered recommendation set for recommendation.

[0015] The beneficial effects of this invention are:

[0016] 1. This invention uses user check-in interest point behavior data, and mines users' periodic preferences and short-term interests based on multi-head self-attention mechanism and linear interpolation embedding method to generate a new set of interest point recommendations for target users, thus solving the data sparsity problem;

[0017] 2. This invention integrates the categories and spatiotemporal information of user access points of interest, analyzes users' periodic access patterns and short-term preferences, and recommends points of interest that match the periodic preferences and short-term interests of target users. The method described in this invention can help users quickly explore and plan their travel purposes. It also helps businesses understand the distribution of user behavior characteristics, analyze potential users, and obtain good economic and social benefits. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the interest point recommendation process of the present invention, which combines users' periodic preferences and short-term interests.

[0019] Figure 2 This is a diagram illustrating the user preference and interest extraction structure of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] An interest-based recommendation method that combines users' periodic preferences and short-term interests, such as Figure 1 As shown, it includes:

[0022] S1: Obtain the raw data and preprocess it to obtain a new dataset; the raw data includes user access record data, check-in sequence data, and spatiotemporal information data;

[0023] S2: Extract relevant attributes from the new dataset to generate check-in records for user visits to points of interest; the relevant attributes include the time information of user visits to points of interest, the category of points of interest, and latitude and longitude coordinates;

[0024] S3: Based on the user's check-in records for visiting points of interest, extract the spatiotemporal trajectory representation and discontinuous spatiotemporal interval matrix of the user's visits to points of interest through a sliding time window;

[0025] S4: Based on the periodic preference representation of users combined with interest categories, extract users' higher-order short-term interests by embedding a discontinuous spatiotemporal interval matrix using linear interpolation based on the spatiotemporal interval matrix.

[0026] S5: Based on the user's periodic preference for interest categories and the user's higher-order short-term interests, recommend interest points from the candidate interest point list to obtain an ordered recommendation set;

[0027] S6: Select the interest point with the highest probability predicted by the model from the ordered recommendation set for recommendation.

[0028] The original data is preprocessed to obtain a new dataset, including:

[0029] The preprocessing of raw data includes cleaning it. Since the raw data is unstructured and cannot be directly used for data analysis, simple data cleaning can structure most of it, such as deleting duplicate data and cleaning up invalid nodes. A check-in threshold is set; the number of times each user checks in on a point of interest and the number of times each point of interest is checked in by a user are counted after cleaning. If a user's check-in count for a point of interest is less than the threshold, that user's data is deleted; if a point of interest is checked in by a user less than the threshold, that point of interest is deleted, resulting in a new dataset.

[0030] The check-in threshold is set to 10. If the number of times a user checks in a point of interest and the total number of times all points of interest are checked in by the user are less than 10, then the user's check-in data and the point of interest are considered not representative.

[0031] Extract relevant attributes from the new dataset, including:

[0032] S21: Obtain user and point of interest ID information, check-in timestamp, point of interest category, and point of interest spatial latitude and longitude coordinates from the acquired user access record data, check-in sequence data, and spatiotemporal information data;

[0033] S22: Convert the check-in timestamp from the acquired data into a periodic time period, and represent the check-in time information as a certain hour of the week.

[0034] Extracting the spatiotemporal trajectory representation of user access points of interest, including:

[0035]

[0036] in, This represents the spatiotemporal trajectory of a user's visits to points of interest. This represents each of a user's check-in records. u represents the user. Indicates a point of interest, c k Indicates the category of interest, t k This indicates the time of the visit, and k represents the check-in order in ascending order of time.

[0037] Extract the discontinuous spatiotemporal interval matrix of user access points of interest, including:

[0038]

[0039]

[0040]

[0041] in, The spatiotemporal interval matrix representing the points of interest visited by the user. express, It means that dis(v) i ,v j () represents the point of interest v i and points of interest v j The spatial distance between them, u represents the current user, T represents the time interval, k represents the check-in order in ascending time, L represents the spatial interval, and t k The order of access is indicated by 'w', and the size of the time window is indicated by 'w'.

[0042] Obtain a periodic preference representation of users combined with interest categories, including:

[0043] S41: Based on user trajectory Each sign-in The formula after embedding is:

[0044]

[0045]

[0046]

[0047] Where C represents the user's time-based access trajectory to points of interest. The user trajectory representation vector, d k This indicates the embedding dimension size, and w represents the time window size. This represents the user embedding for the i-th check-in. This represents the embedding of the interest point for the i-th check-in. This represents the embedding of the interest category for the i-th check-in. This represents the embedding of the time of the i-th check-in. Joint represents the embedding dimension. i Let represent the joint embedding representation of the i-th check-in, and joint represent the embedding representation of the access trajectory. Indicates the t-th k The joint embedding representation of the next sign-in, where Softsign() represents the activation function;

[0048] S42: Calculate the periodic preference representation vector of the user combined with points of interest based on the user trajectory representation vector.

[0049] The periodic preference representation vector combining user points of interest is calculated based on the user trajectory representation vector, including:

[0050] head i =Attention(CW) i Q CW i K CW i V )

[0051]

[0052] P=Concat(head1,head2,…,head h W O

[0053] FFN(P)=LeakyReLU(PW1+b1)W2+b2

[0054] Period = FFN(P) + C

[0055] Where C represents the user's time-based access trajectory to points of interest. The user trajectory representation vector, W Q W V and W K Let d represent the first, second, and third trainable matrices for linear transformation of C. k Indicates the embedding dimension size, Period(U) represents the user's periodic preferences combined with interest category within a time window, and head i This represents the output of the i-th self-attention head, where Attention() represents the self-attention calculation function, Q represents the query vector, K represents the key vector, V represents the value vector, T represents the vector transpose, P represents the vector representation of all self-attention heads after linear transformation, Concat() represents the vector concatenation function, and W... O Let represent the trainable matrix with linear operations on the self-attention head, FFN(P) represent the fully connected layer operation, LeakyReLU() represent the activation function, W1 and W2 represent the trainable matrices, and b1 and b2 represent the trainable parameters. Period This represents a vector representing user periodic preferences.

[0056] Embedded discontinuous spatiotemporal interval matrices include:

[0057]

[0058]

[0059]

[0060]

[0061] in, Let e ​​represent the short-term interest representation of a user's i-th check-in. Δ This represents a short-term interest representation of the check-in user's activity. Let Δ represent the time interval matrix embedding learning and spatial interval learning for the user's i-th check-in, respectively. T and Δ L Upper(Δ) represents the time interval matrix and the spatial interval matrix, respectively. T ) and Lower(Δ T ) and Upper(Δ L ) and Lower(Δ L ) represent the maximum and minimum values ​​of time and space intervals, respectively. and as well as and Let T represent the trainingable matrix for spatiotemporal transition, LIT() represent the embedding function, L represent the spatial interval, k represent the check-in order in ascending time order, w represent the time window size, and t represent the training matrix for spatiotemporal transition. k Indicates the order of access.

[0062] Based on the user's periodic preferences for interest categories and higher-order short-term interests, recommended interest points are selected from the candidate interest point list to obtain an ordered recommendation set, such as... Figure 2 As shown, it includes:

[0063] S51: Obtain the embeddings of all points of interest as Where candidate(POI) represents the set of candidate interest points after embedding, e v This represents the embedding representation of the first interest point in the set of interest points, where N represents the number of interest points;

[0064] S52: Based on the user's periodic preference for interest categories and the user's short-term interests, obtain the recommendation prediction probability of candidate interest points through attention query. Sort the candidate interest points in descending order according to the recommendation prediction probability, and recommend the set of interest points that the user is most likely to visit at the next time point to the target user in a TOP-N manner, thus obtaining an ordered recommendation set.

[0065] The recommendation prediction probability of candidate interest points is obtained through attention query, including:

[0066] Query(Q,K,V)=V·softmax(QK T )

[0067]

[0068] Where candidate(POI) represents the set of candidate points of interest after embedding, and Period represents the periodic interest representation vector of the user combined with the category of the points of interest. For short-term user preferences, Q represents the query vector, K represents the key vector, V represents the value vector, Query() represents the recommendation function, softmax() represents the activation function, prob represents the recommendation prediction probability of candidate points of interest, and POI represents the set of candidate points of interest that are not embedded.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An interest-based recommendation method that combines users' periodic preferences and short-term interests, characterized in that, include: S1: Obtain the raw data and preprocess it to obtain a new dataset; the raw data includes user access record data, check-in sequence data, and spatiotemporal information data; S2: Extract relevant attributes from the new dataset to generate check-in records for user visits to points of interest; the relevant attributes include the time information of user visits to points of interest, the category of points of interest, and latitude and longitude coordinates; S3: Based on the user's check-in records for visiting points of interest, extract the spatiotemporal trajectory representation and discontinuous spatiotemporal interval matrix of the user's visits to points of interest through a sliding time window; Extracting the spatiotemporal trajectory representation of user access points of interest, including: ； in, This represents the spatiotemporal trajectory of a user's visits to points of interest. This represents each of a user's check-in records. , Indicates user, Indicate points of interest, Indicates the category of points of interest. Indicates the time of the visit. This indicates the sign-in order in ascending chronological order. Extract the discontinuous spatiotemporal interval matrix of user access points of interest, including: ； ； ； in, The spatiotemporal interval matrix representing the points of interest visited by the user. Indicate points of interest and points of interest Spatial distance between them Indicates time interval, Indicates spatial interval, Indicates the order of access. Indicates the size of the time window; S4: Based on the periodic preference representation of users combined with interest categories, extract users' higher-order short-term interests by embedding a discontinuous spatiotemporal interval matrix using linear interpolation based on the spatiotemporal interval matrix. Obtain a periodic preference representation of users combined with interest categories, including: S41: Based on user trajectory Each sign-in The formula after embedding is: ； ； ； in, This indicates the user's time-based access trajectory to points of interest. The user trajectory representation vector, Indicates the size of the embedding dimension. Indicates the first The user's embedded representation of the second sign-in. Indicates the first The interest points of each check-in are embedded in the representation. Indicates the first The interest category embedding representation of each check-in Indicates the first The time of each check-in is embedded in the representation. Indicates the embedding dimension. Indicates the first The joint embedding representation of the second sign-in An embedded representation of the access trajectory. Indicates the first The joint embedding representation of the second sign-in Indicates the activation function; S42: Calculate the periodic preference representation vector of the user combined with points of interest based on the user trajectory representation vector; Embedded discontinuous spatiotemporal interval matrices include: ； ； ； ； in, This indicates that the user is number 1. The number of times a sign-in indicates short-term interest. This represents a short-term interest representation of the check-in user's activity. , These represent the user's number. Learning the time interval matrix embedding and spatial interval learning for each sign-in session. and These represent the time interval matrix and the spatial interval matrix, respectively. and as well as and These represent the maximum and minimum values ​​of time and space intervals, respectively. and as well as and These represent the trainable matrices for spatiotemporal transitions, Indicates an embedded function; S5: Based on the user's periodic preference for interest categories and the user's higher-order short-term interests, recommend interest points from the candidate interest point list to obtain an ordered recommendation set; S6: Select the optimal point of interest from the ordered recommendation set for recommendation.

2. The interest point recommendation method combining user periodic preferences and short-term interests according to claim 1, characterized in that, The original data is preprocessed to obtain a new dataset, including: Clean the original data; set a check-in threshold; based on the check-in threshold, count the number of times each user checks in on a point of interest and the number of times each point of interest is checked in by a user in the cleaned data. If the number of times a user checks in on a point of interest is less than the check-in threshold, delete the user's check-in data. If the number of times a point of interest is checked in by a user is less than the check-in threshold, delete the point of interest and obtain a new dataset. The check-in threshold is set to 10. If the number of times a user checks in a point of interest and the total number of times all points of interest are checked in by the user are less than 10, then the user's check-in data and the point of interest are considered not representative.

3. The interest point recommendation method combining user periodic preferences and short-term interests according to claim 1, characterized in that, Extract relevant attributes from the new dataset, including: S21: Obtain user and point of interest ID information, check-in timestamp, point of interest category, and point of interest spatial latitude and longitude coordinates from the acquired user access record data, check-in sequence data, and spatiotemporal information data; S22: Convert the check-in timestamp from the acquired data into a periodic time period, and represent the check-in time information as a certain hour of the week.

4. The interest point recommendation method combining user periodic preferences and short-term interests according to claim 1, characterized in that, The periodic preference representation vector combining user points of interest is calculated based on the user trajectory representation vector, including: ； ； ； ； ； in, This indicates the user's time-based access trajectory to points of interest. The user trajectory representation vector, , and express The first, second, and third trainable matrices for linear transformation. Indicates the size of the embedding dimension. This indicates the user's periodic preferences within a time window, combined with categories of interest. Indicates the first Each self-attention head output result Represents the self-attention computation function. Represents the query vector. Represents the key vector. Represents a value vector. This represents the transpose of a vector. This represents the vector representation of all self-attention heads after a linear transformation. This represents a vector concatenation function. Represents a trainable matrix that performs linear operations with the self-attention head. This represents the fully connected layer operation. Represented as an activation function, , They represent trainable matrices, , They represent the trainable parameters, This represents a vector representing user periodic preferences.

5. The interest point recommendation method combining user periodic preferences and short-term interests according to claim 1, characterized in that, Based on the user's periodic preferences for interest categories and higher-order short-term interests, recommended interest points are selected from the candidate interest point list to obtain an ordered recommendation set, including: S51: Obtain the embeddings of all points of interest as ,in, This represents the set of candidate interest points after embedding. This represents the embedding representation of the first interest point in the set of interest points. Indicates the number of points of interest; S52: Based on the user's periodic preference for interest categories and the user's short-term interests, obtain the recommendation prediction probability of candidate interest points through attention query. Sort the candidate interest points in descending order according to the recommendation prediction probability, and recommend the set of interest points that the user is most likely to visit at the next time point to the target user in a TOP-N manner, thus obtaining an ordered recommendation set.

6. The interest point recommendation method combining user periodic preferences and short-term interests according to claim 5, characterized in that, The recommendation prediction probability of candidate interest points is obtained through attention query, including: ； ； in, This represents the set of candidate interest points after embedding. This represents a periodic interest representation vector that combines user interest point categories. For users' short-term preferences, Represents the query vector. Represents the key vector. Represents a value vector. This represents the recommendation function. This represents the activation function. This represents the probability of recommending candidate points of interest. This represents the set of candidate interest points that are not embedded.

Images