Next interest point recommendation method fusing global periodicity and local regularity

Abstract

A next interest point recommendation method fusing global periodicity and local regularity comprises the steps that 1, historical data are obtained and preprocessed, and a data set is obtained; 2, extracting related attributes in the data set to form a sign-in sequence of the user; 3, constructing a graph structure for a graph convolutional network by using the distance span information in the short-term sign-in sequence of the user, and updating the embedded representation of the interest point; 4, coding the interest point sign-in sequence of the user to generate a sign-in embedded representation, and dividing the sign-in embedded representation into sign-in sequence embedded representations taking sign-in as a unit and taking day as a unit; 5, according to the embedded representation of the sign-in sequence, constructing a short-term preference embedded representation and a long-term preference embedded representation representing the user; 6, predicting the probability that the next interest point of the user is a certain place according to the long-term and short-term preference embedding representation of the user; and 7, selecting several optimal interest points for recommendation to obtain a recommendation set. By means of the method, more personalized and more accurate recommendation of the next interest point is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of interest point recommendation in social networks, and in particular, relates to a next interest point recommendation method integrating global periodicity and local regularity. Background Art

[0002] With the widespread popularity of social media and the rapid development of location information technology, the generation and application of location data has become a major feature of contemporary society. The popularity of smartphones and other mobile devices has led to a rapid accumulation of location data, greatly improving the convenience of daily life and optimizing the accuracy and convenience of location-based services. However, with the increasing amount of information, it has become an urgent need to accurately recommend locations of interest to users and improve their offline engagement. On the one hand, personalized location recommendation systems can effectively improve user engagement and satisfaction; on the other hand, personalized location recommendations provide businesses with an effective promotion platform, enabling them to accurately locate potential customer groups. Therefore, accurately recommending places of interest to users from massive location data can not only significantly improve the quality of life of residents, but also stimulate consumption potential, accelerate the construction of a new development pattern dominated by the domestic market, and further promote the modernization of the national governance system and governance capabilities.

[0003] In recent years, many scholars have conducted in-depth research on the location recommendation problem in location-based social networks. These studies mainly focus on modeling the geographic information, text information, temporal information, and social information of user check-in data, and use advanced deep learning technologies such as Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM) and Graph Neural Network (GNN) for location recommendation.

[0004] In social networks, users' preferences for points of interest may change at different times, and their preferences for distance spans may also vary. For example, users may travel during holidays, but stay near home at other times. In addition, users' preferences for points of interest also have periodic patterns, such as visiting specific places on certain days of the month, or showing different preferences for points of interest on weekends and weekdays.

[0005] However, the existing POI recommendation methods do not effectively combine the user's global periodicity and local regularity to recommend the next POI, resulting in the need to improve the recommendation effect. Summary of the invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a next point of interest recommendation method that integrates global periodicity and local regularity, so as to provide a more personalized and accurate next point of interest recommendation method.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A next point of interest recommendation method integrating global periodicity and local regularity, the steps of which are:

[0009] Step 1: Obtain the user's historical check-in data from the public dataset and preprocess it to obtain a new dataset.

[0010] Step 101: Clean the historical check-in data of the original user;

[0011] Step 102: Setting a threshold for the length of a sign-in sequence;

[0012] Step 103: Clean out the sign-in sequences that do not meet the threshold, that is, the sign-in sequence length is less than the threshold.

[0013] Step 2: Extract relevant attributes from the dataset: user, visited POIs, geographic location of POIs, and visit time to form the user's check-in sequence.

[0014] Step 201: For each check-in record, extract the user ID, the POI ID, the longitude and latitude of the POI, and the specific check-in time;

[0015] Step 202: Convert the specific sign-in time into two attributes, namely, the number of hours of the sign-in time in a day and whether the day is a weekend;

[0016] Step 203: Different sign-in records of the same user are combined into a sign-in sequence of the user in chronological order.

[0017] Step 3: Use the distance span information in the user’s short-term check-in sequence to build a graph structure, which is used in the graph convolutional network to update the embedding representation of the interest points.

[0018] The graph structure is constructed by using the distance span information in the user's short-term check-in sequence, including:

[0019] Step 301: Calculate the distance span of each location transfer based on the user's short-term check-in sequence

[0020] Step 302: Divide the distance interval according to the hyperparameter γ and the maximum value of the distance span;

[0021] Step 303: Determine the distance interval to which the distance span of each location transfer of the user belongs

[0022] Step 304: According to The number of times different distance intervals appear in the , assign weights to each different distance interval, and obtain a set of non-repeated intervals and their weights

[0023] Step 305: Construct a user weight matrix A with an initial value of 0. In the (distance interval, weight) pair, if the distance between interest point i and interest point j is within the distance interval, then A i,j Set it as the weight corresponding to the distance interval until all (distance interval, weight) pairs are traversed to obtain a weight matrix A as the graph structure of the graph convolution operation.

[0024] Using graph convolutional networks to update the embedded representation of interest points specifically includes:

[0025] Step 306: Set the number of graph convolution layers, use A to perform graph convolution operations on the embedded representation of the interest point, and obtain output vectors of different layers;

[0026] Step 307: Aggregate the output vectors of different layers into This is the final embedding representation of the interest point.

[0027] Step 4: Encode the user’s POI check-in sequence to generate an embedded representation of the check-in, and divide it into check-in sequence embedding representations in units of check-ins and days.

[0028] Generating the embedded representation of the check-in specifically includes:

[0029] Step 401: Generate an embedding representation of the user, time, and whether it is the weekend using the embedding layer;

[0030] Step 402: Connect the embedded representations of the user, the point of interest, the time, and whether it is the weekend to obtain the embedded representation of the check-in.

[0031] Converting the check-in embedding representation into the check-in sequence embedding representation in units of check-ins and days specifically includes:

[0032] Step 403: Divide the user check-in sequence into days, and embed the check-ins within a day into a representation E day Input into Mamba Block, and then get the updated embedding representation in sign-in units;

[0033] Step 404: Aggregate the embedded representations obtained in step 403 to obtain an embedded representation in units of days representing the day;

[0034] Step 405: The output result of the check-in sequence of all days after the processing of step 403 is used as the check-in sequence embedding representation H of the user sequence in units of check-ins check-in The output result of the check-in sequence of all days after processing in step 404 is used as the check-in sequence embedding representation H of the user sequence in units of days day .

[0035] Step 5: Based on the embedding representation of the check-in sequence, construct the short-term preference embedding representation and long-term preference embedding representation representing the user.

[0036] Constructing a user's short-term preference embedding representation specifically includes:

[0037] Step 501: Based on the preference embedding representation of the most recent day, attention weights are constructed by the similarity with the preference embedding representations of other days;

[0038] Step 502: Using each weight, aggregate the preference embedding representations of different days as the user's short-term preference embedding representation h Short .

[0039] Constructing a user's long-term preference embedding representation specifically includes:

[0040] Step 503: using Fast Fourier Transform (FFT) to capture the user's long-term periodic preference features, and calculating the weight of each check-in embedding representation according to the periodic preference;

[0041] Step 504: Using the weights of the check-in embeddings, aggregate all preference embeddings to form the user's long-term preference embedding h Long .

[0042] Step 6: Based on the user’s long-term and short-term preference embedding representations, predict the probability that the user’s next point of interest will be certain locations.

[0043] Step 601: embed the user's short-term preference into h Short and the long-term preference embedding representation h Long Merge into user preference embedding representation h u ;

[0044] Step 602: h u Input to a fully connected layer and perform softmax processing on the output to get the predicted probability of the next point of interest

[0045] Step 7: According to the recommendation probability of the points of interest, select the best points of interest for recommendation to obtain a recommendation set.

[0046] Step 701: sorting the interest point candidate set in descending order according to the recommendation probability;

[0047] Step 702: Select interest points in a TOP-K manner to obtain an ordered set of interest point recommendations.

[0048] The beneficial effects created by the present invention are: 1. The present invention proposes a novel graph construction method for graph convolutional networks, which can effectively capture the user's recent distance span preferences and provide strong support for subsequent sequence modeling; 2. The present invention designs a novel method for user long-term and short-term sequence modeling, which utilizes discrete Fourier transform and attention mechanism to capture the user's long-term periodic preferences and short-term preferences; 3. The present invention constructs a novel model for next point of interest recommendation, which captures the user's near-term and long-term global periodic preferences and short-term local regularity preferences, and models the user's long-term and short-term sequences to achieve more personalized and accurate next point of interest recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 This is a structural diagram of the model recommended by the present invention;

[0051] Figure 3 for Figure 2 Detailed structure diagram of the Manba-based encoding layer. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. The following embodiments are applicable to the present invention, but are not intended to limit the scope of the present invention.

[0053] like Figure 1 As shown, the present invention proposes a method for recommending the next point of interest by combining the user's short-term span preference and long-term periodic preference. The model structure of the invention is as follows: Figure 2 As shown, the following steps are included:

[0054] Step 1: Obtain the user's historical check-in data from the public dataset and preprocess it to obtain a new dataset.

[0055] Step 101: Clean the historical check-in data of the original user;

[0056] Step 102: Set a threshold for the length of the sign-in sequence, requiring that the sign-in sequence has at least two days with a sign-in count greater than or equal to 2 times in the past 7 days;

[0057] Step 103: Clean out the sign-in sequences that do not meet the threshold, that is, the sign-in sequence length is less than the threshold.

[0058] Step 2: Extract relevant attributes from the dataset: user, visited POIs, geographic location of POIs, and visit time to form the user's check-in sequence.

[0059] Step 201: For each check-in record, extract the user ID, the POI ID, the longitude and latitude of the POI, and the specific check-in time;

[0060] Step 202: Convert the specific sign-in time into two attributes, namely, the number of hours of the sign-in time in a day and whether the day is a weekend, denoted as t and d.

[0061] Step 203: Different sign-in records of the same user are combined into a sign-in sequence S(u) = {c1, c2, ..., c |S|}, where c = (u, l, t, d), u represents the user, l represents the point of interest, including the point of interest ID and the longitude and latitude of the point of interest.

[0062] Step 3: Use the distance span information in the user’s short-term check-in sequence to build a graph structure, which is used in the graph convolutional network to update the embedding representation of the interest points.

[0063] Using the distance span information in the user's short-term check-in sequence to construct a graph structure specifically includes:

[0064] Step 301: Based on the user's short-term check-in sequence Calculate the distance span of each location transfer in Represents the distance between two points of interest;

[0065] Step 302: Divide the distance interval according to the hyperparameter γ = 0.5 (kilometers) and the maximum value of the distance span, starting from 0, and divide the distance into intervals of length γ {r1, r2, ..., r last},make sure

[0066] Step 303: Determine the distance interval to which the distance span of each location transfer of the user belongs

[0067] Step 304: According to The number of times different distance intervals appear in the , assign weights to each different distance interval, and obtain a set of non-repeated intervals and their weights where w i = count(r i ) / (m-1);

[0068] Step 305: Construct a user weight matrix A with an initial value of 0. In the (distance interval, weight) pair, if the distance between interest point i and interest point j is within the distance interval, then A i,j Set it as the weight corresponding to the distance interval. Until all (distance interval, weight) pairs are traversed, a weight matrix A is obtained as the graph structure of the graph convolution operation.

[0069] Using graph convolutional networks to update the embedded representation of interest points specifically includes:

[0070] Step 306: Set the number of graph convolution layers l = 2, and use A to embed the interest point to represent E and perform graph convolution operation. The formula is:

[0071]

[0072] Where E 0 The element e in l is the initial embedding representation of the interest point obtained by the embedding layer, with dimension d l , D is a diagonal matrix representing the degree of nodes in A, used for regularization operation, σ is the activation function, and then the output vectors of different layers {E 0 ,E 1 ,…,E l};

[0073] Step 307: Aggregate the output vectors of different layers into This is the final embedding representation of the interest point

[0074] Step 4: Encode the user’s POI check-in sequence to generate an embedded representation of the check-in, and divide it into check-in sequence embedding representations in units of check-ins and days.

[0075] Generating the embedded representation of the check-in specifically includes:

[0076] Step 401: Generate an embedding representation of the user, time, and whether it is a weekend using the embedding layer u 、e t and, e d Its dimensions are d u d t and d d ;

[0077] Step 402: Connect the embedded representations of the user, point of interest, time, and whether it is a weekend to obtain the embedded representation of the check-in e = [e u ;e l ;e t ;e d ], where the dimension of e is d and satisfies d = d u +d l +dt +d d .

[0078] Converting the check-in embedding representation into the check-in sequence embedding representation in units of check-ins and days specifically includes:

[0079] Step 403: Divide the user's check-in sequence into days, and embed the check-in sequence of a certain day into E day ={e1,e2,...,e m}, use Mamba Block to update the sign-in embedding representation of each day, and get H = {h1,h2,...,h m}, the formula is:

[0080] H=MambaBlock(E day )

[0081] Mamba Block is the computational module of Mamba. Mamba introduces a data-dependent selection mechanism based on SSM (state space model) and integrates hardware-aware parallel algorithms in its loop mode. This mechanism enables Mamba to capture contextual information in long sequences while maintaining computational efficiency. As an approximate linear perplexity sequence model, Mamba shows potential in terms of efficiency improvement and performance improvement in long sequence tasks.

[0082] Step 404: Aggregate the embedded representations obtained in step 403 to obtain an embedded representation in units of days representing the day. The calculation formula is:

[0083] h day =aggregation(h1,h2,...,h m )

[0084] The aggregation operation uses the summing and averaging method.

[0085] Step 405: The output result of the check-in sequence of all days after processing in step 403 is used as the check-in sequence embedding representation of the user sequence in units of check-ins The output result of the check-in sequence of all days after processing in step 404 is used as the check-in sequence embedding representation of the user sequence in units of days.

[0086] Step 5: Based on the embedding representation of the check-in sequence, construct the short-term preference embedding representation and long-term preference embedding representation representing the user.

[0087] Constructing a user's short-term preference embedding representation specifically includes:

[0088] Step 501: Based on the preference embedding representation of the most recent day, the attention weight is constructed by the similarity with the preference embedding representation of other days. The calculation formula is as follows:

[0089]

[0090] Step 502: Use the weights to aggregate the preference embedding representations of different days. The formula is as follows:

[0091]

[0092] h Short As a user's short-term preference embedding representation.

[0093] Constructing a user's long-term preference embedding representation specifically includes:

[0094] Step 503: Use Fast Fourier Transform (FFT) to capture the user's long-term periodic preference characteristics, the formula is as follows:

[0095]

[0096] Where L is the length of the user sequence, τ i =t L+1 -t i , τ' i =t i -t1, j is an imaginary unit, if l L = l n Then s[n] is 1, otherwise it is 0.

[0097] This probability is used to infer the probability that the next point of interest that the user will visit is a point of interest that he has visited before a certain time, that is,

[0098] p(l L+1 = l i |τ i )=p(l L = l i-1 |τ i-1 ).

[0099] Then, weights are assigned to check-ins at different times according to periodic preferences, using the following formula:

[0100] β ti = p(l L+1 = l i |τ i )

[0101]

[0102] Step 504: Using the weights of each sign-in, aggregate all preference embedding representations to serve as the user's long-term preference embedding representation. The formula is as follows:

[0103]

[0104] Step 6: Based on the user’s long-term and short-term preference embedding representations, predict the probability that the user’s next point of interest will be certain locations.

[0105] Step 601: embed the user's short-term preference into h Short and the long-term preference embedding representation h Long Merge into user preference embedding representation, the formula is as follows:

[0106] h u =h Long +h Short

[0107] Step 602: h u Input to a fully connected layer and perform softmax processing on the output to get the predicted probability of the next point of interest The formula is as follows:

[0108]

[0109] Step 7: According to the recommendation probability of the points of interest, select the best points of interest for recommendation to obtain a recommendation set.

[0110] Step 701: sorting the interest point candidate set in descending order according to the recommendation probability;

[0111] Step 702: Select interest points in a TOP-K manner to obtain an ordered set of interest point recommendations.

Claims

1. A next point of interest recommendation method integrating global periodicity and local regularity, characterized in that: The steps are: Step 1: Obtain the user's historical check-in data from the public dataset and preprocess it to obtain a new dataset; Step 2: Extract relevant attributes from the dataset: user, visited POIs, geographic location of POIs, and visit time to form the user's check-in sequence; Step 3: Use the distance span information in the user's short-term check-in sequence to construct a graph structure for use in the graph convolutional network to update the embedded representation of the interest points; Step 4: Encode the user's POI check-in sequence to generate an embedded representation of the check-in, and divide it into check-in sequence embedding representations in units of check-ins and in units of days; Step 5: Based on the embedding representation of the check-in sequence, construct the short-term preference embedding representation and long-term preference embedding representation representing the user; Step 6: Based on the user’s long-term and short-term preference embedding representations, predict the probability that the user’s next point of interest will be certain locations; Step 7: According to the recommendation probability of the POI, select the best POI for recommendation and obtain the recommendation set.

2. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In the step 1, the specific method is: Step 101: Clean the historical check-in data of the original user; Step 102: Setting a threshold for the length of a sign-in sequence; Step 103: Clean out the check-in sequences that do not meet the threshold and obtain filtered historical check-in data.

3. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In the step 2, the specific method is: Step 201: For each check-in record, extract the user ID, the POI ID, the longitude and latitude of the POI, and the specific check-in time; Step 202: Convert the specific sign-in time into two attributes, namely the number of hours of the sign-in time in a day and whether the day is a weekend.

4. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In step 3, the specific method of constructing the graph structure using the distance span information in the user's short-term check-in sequence is: Step 301: Calculate the distance span of each location transfer based on the user's short-term check-in sequence Step 302: Divide the distance interval according to the hyperparameter γ and the maximum value of the distance span; Step 303: Determine the distance interval to which the distance span of each location transfer of the user belongs Step 304: According to The number of times different distance intervals appear in the , assign weights to each different distance interval, and obtain a set of non-repeated intervals and weights Step 305: Construct a user weight matrix A with an initial value of 0. In the (distance interval, weight) pair, if the distance between interest point i and interest point j is within the distance interval, then A i,j Set it as the weight corresponding to the distance interval until all (distance interval, weight) pairs are traversed to obtain a weight matrix A as the graph structure of the graph convolution operation.

5. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In step 3, the specific method of updating the embedding representation of the interest point using the graph convolutional network is as follows: Step 306: Set the number of graph convolution layers, use A to perform graph convolution operations on the embedded representation of the interest point, and obtain output vectors of different layers; Step 307: Aggregate the output vectors of different layers into This is the final embedding representation of the interest point.

6. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In the step 4, the specific method is: Generate an embedding representation of the check-in: Step 401: Generate an embedding representation of the user, time, and whether it is the weekend using the embedding layer; Step 402: Connect the embedded representations of the user, the point of interest, the time, and whether it is the weekend to obtain the embedded representation of the check-in; Convert the check-in embedding representation into a check-in sequence embedding representation in units of check-ins and days: Step 403: Divide the user check-in sequence into days, and embed the check-ins within a day into a representation E day Input into MambaBlock, and then get the updated embedding representation in sign-in units; Step 404: Aggregate the embedded representations obtained in step 403 to obtain an embedded representation in units of days representing the day; Step 405: The output result of the check-in sequence of all days after the processing of step 403 is used as the check-in sequence embedding representation H of the user sequence in units of check-ins check-in , the output result of the check-in sequence of all days after processing in step 404 is used as the check-in sequence embedding representation H of the user sequence in units of days day .

7. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In the step 5, the specific method is: Construct user short-term preference embedding representation: Step 501: Based on the preference embedding representation of the most recent day, attention weights are constructed by the similarity with the preference embedding representations of other days; Step 502: Using each weight, aggregate the preference embedding representations of different days as the user's short-term preference embedding representation h Short . Constructing user long-term preference embedding representation: Step 503: Use Fast Fourier Transform (FFT) to capture the user's long-term periodic preference features, and calculate the weight of each sign-in embedding representation according to the periodic preference; Step 504: Using the weights of the check-in embeddings, aggregate all preference embeddings to form the user's long-term preference embedding h Long .

8. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In the step 6, the specific method is: Step 601: embed the user's short-term preference into h Short and the long-term preference embedding representation h Long Merge into user preference embedding representation h u ; Step 602: h u Input to a fully connected layer and perform softmax processing on the output to get the predicted probability of the next point of interest 9. The next point of interest recommendation method integrating global periodicity and local regularity according to claim 1, characterized in that: In the step 7, the specific method is: Step 701: sorting the interest point candidate set in descending order according to the recommendation probability; Step 702: Select interest points in a TOP-K manner to obtain an ordered set of interest point recommendations.

Images