A point of interest recommendation method based on residual space-time cooperation network
By combining long-term and short-term user behavior with residual spatiotemporal collaboration networks, and utilizing temporal collaboration matrices and skip learning algorithms, the accuracy problem of point-of-interest recommendation systems is solved, resulting in more accurate point-of-interest recommendations and improved user satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF TECH
- Filing Date
- 2023-12-12
- Publication Date
- 2026-06-26
AI Technical Summary
Existing interest-based recommendation systems cannot effectively combine long-term and short-term user behavior, resulting in low recommendation accuracy.
A residual spatiotemporal cooperative network-based approach is adopted to capture users' long-term dependencies through historical interest point trajectory sequences and temporal cooperative matrices, capture short-term dependencies using a skip learning algorithm, and balance the relationship between the two through a sample balancer, ultimately outputting recommendation results.
It improves the accuracy of point-of-interest recommendations, recommends more locations that users are interested in, enhances user satisfaction, and reduces the impact of real-time location information on spatial dimensions.
Smart Images
Figure CN117725322B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of point of interest recommendation technology, and in particular to a point of interest recommendation method based on residual spatiotemporal cooperative networks. Background Technology
[0002] In recent years, the booming development of location-based social services, such as Foursquare and Uber, has led to unprecedented progress in the field of Points of Interest (POI) recommendation. The vast amount of user check-in behavior provides rich data for recommending the next POI. As an extension of traditional POI recommendation, next POI recommendation aims to provide more personalized content recommendations and has attracted widespread attention from researchers.
[0003] Early research primarily focused on short-term user dependencies, such as using Markov chains. Subsequently, RNN-based methods and their variants were introduced to better capture short-term trajectory features, such as ST-RNN and CARA. Meanwhile, in addition to the short-term dynamic dependencies exhibited by recent user behavior, people's behavior also displays long-term periodicity. For example, a user might shop at the supermarket every weekend, demonstrating long-term periodic behavior. However, in the next POI recommendation task, research considering both long-term and short-term user behavior is very limited, resulting in low accuracy of recommendation systems in recommending points of interest to users. Summary of the Invention
[0004] This invention aims to at least solve the technical problem of low accuracy in recommending points of interest to users in existing recommendation systems, and innovatively proposes a point of interest recommendation method based on residual spatiotemporal cooperative networks.
[0005] To achieve the above-mentioned objectives of this invention, this invention provides an interest point recommendation method based on residual spatiotemporal cooperative networks, the method comprising:
[0006] S100: Obtain the user's historical points of interest and current points of interest, sort the historical points of interest and current points of interest in chronological order, and generate a historical point of interest trajectory sequence and a current point of interest trajectory sequence.
[0007] S200. Establish a time collaboration matrix and use the historical point of interest trajectory sequence and time collaboration matrix to capture the user's long-term dependence.
[0008] S300. Based on the current interest point trajectory sequence, use a skip learning algorithm to capture the user's short-term dependence.
[0009] S400. Use a sample balancer to balance the user's long-term dependence and the user's short-term dependence, and output the recommendation results.
[0010] As an optional embodiment of the present invention, the establishment of the time cooperation matrix may include:
[0011] S201. Divide the same time period into multiple time steps, calculate the time similarity between different users in the first time step and the second time step, and use the time similarity to recommend the interest points of the second user who is highly similar to the first user to the first user.
[0012] The formula for calculating the time similarity is:
[0013]
[0014] Where, α i,j D represents the temporal similarity between the first time step i and the second time step j. i Let D represent the set of points of interest at time step i. j This represents the set of points of interest at time step j;
[0015] S202. Calculate the temporal collaboration weight, and measure the similarity of interest points within the first time step and the second time step based on the temporal collaboration weight;
[0016] The formula for calculating the time collaboration weight is as follows:
[0017]
[0018]
[0019] Among them, t h T is the current representation of the user's historical interest point trajectory. h W is the user's historical point of interest trajectory sequence. t As a time-based collaboration weight, To decode historical trajectories, α c This is the time similarity matrix.
[0020] As an optional embodiment of the present invention, the step of capturing the user's long-term dependency using the historical point of interest trajectory sequence and the time collaboration matrix may include:
[0021] S203. Use an encoder to map the historical interest point trajectory sequence of each time step onto a low-dimensional projection.
[0022] The mapping formula is:
[0023]
[0024] in, The result of the encoder processing the embedding of the trajectory sequence of i-1 historical points of interest is... For the embedding of the trajectory sequence of i-1 historical points of interest;
[0025] S204. Embed the historical point of interest trajectory sequence into the residual linear layer, and use the residual linear layer to obtain the user's long-term dependency.
[0026] As an optional embodiment of the present invention, the step of obtaining the user's long-term dependency using the residual linear layer may include:
[0027] S241. Set up two multilayer perceptrons and use the first perceptron to calculate the dynamic covariates.
[0028] The calculation formula is:
[0029]
[0030] in, Represents the trajectory sequence T at time t. i-1 r-dimensional dynamic covariates;
[0031] S242. Calculate the density of the user's historical interest point trajectory using the activation function and the second multilayer perceptron.
[0032] The calculation formula is:
[0033]
[0034]
[0035] in, Dense expressions representing historical trajectories, It is an intermediate expression of the historical trajectory. This indicates a residual connection.
[0036] As an optional embodiment of the present invention, the step of capturing users' long-term dependencies using the historical point of interest trajectory sequence and the time collaboration matrix may further include:
[0037] S205, Stack and tile the dynamic covariates of the past and future, and then... The mapping formula for the vector at each time step is as follows:
[0038]
[0039] in, This is the encoded embedding after processing by the residual linear layer, where n represents the number of residual linear layers;
[0040] S206. Utilize a multi-head attention mechanism to calculate the salience of different interest points in the user's historical interest point trajectory or current interest point trajectory. The calculation formula is as follows:
[0041]
[0042] in, This represents the output of the multi-head attention mechanism. This represents the embedding of a user's historical point of interest trajectory encoded by a residual linear layer or the embedding of the current point of interest trajectory, where γ represents the current or historical.
[0043] S207. Use a decoder to map the output of the multi-head attention mechanism to the same dimension as the historical interest point trajectory sequence;
[0044] The calculation formula is:
[0045]
[0046] in, This represents the output of the decoder. This represents the output of the historical interest point trajectory after processing by the multi-head attention mechanism.
[0047] S208. Use global residual connections to combine the historical interest point trajectory sequence with the decoder output;
[0048] The calculation formula is as follows:
[0049]
[0050] in, This is the output after global residual connection;
[0051] S209. Use average pooling to retain all user information;
[0052] The formula for average pooling is:
[0053]
[0054] Among them, t n This represents the current trajectory sequence of the user's points of interest. This represents the result after the current point of interest trajectory has been processed sequentially through the multi-head attention mechanism, decoder, and global residual connection.
[0055] S210. Utilize non-localized operations to reduce the impact of users' real-time location information on spatial dimensions, and calculate users' long-term dependencies:
[0056] The non-localized operation formula is:
[0057]
[0058]
[0059]
[0060] Where, d n,h Indicates the location of the point of interest (l) t-1 With interest point trajectory t h The distance between them, where Er represents the Earth's radius, lon represents longitude, and lat represents latitude;
[0061] The formula for calculating long-term user dependency is:
[0062]
[0063] in, W indicates long-term user dependency. h This represents the trainable projective weight matrix.
[0064] As an optional embodiment of the present invention, the step of capturing the user's short-term dependence using a skip learning algorithm based on the current point of interest trajectory sequence includes:
[0065] S301. Establish a distance matrix between each pair of interest points in the current interest point trajectory sequence described by the user, and use the distance matrix to determine the distance that the skip learning algorithm needs to skip.
[0066] S302. Process the current interest point trajectory sequence and establish an input set;
[0067] The formula for calculating the skip learning algorithm is:
[0068] h t-1 =n R ×RLL(h θ )
[0069]
[0070] in, This indicates a user's short-term dependence.
[0071] As an optional embodiment of the present invention, the calculation formula for balancing the user's long-term dependence and the user's short-term dependence using the sample balancer is:
[0072]
[0073] Where F represents the probability set of each point of interest.
[0074] As an optional embodiment of the present invention, the calculation formula for the final output recommendation data is optionally:
[0075] Next poi =Max(F)
[0076] Among them, Next poi Let F represent the point of interest that the next user is most likely to visit, and Max(F) represent the point of interest with the largest probability in the probability set.
[0077] The beneficial effects of this invention are as follows: By utilizing historical point-of-interest trajectory sequences and temporal collaboration matrices to capture users' long-term and short-term dependencies, and by employing a sample balancer to balance these dependencies, the accuracy of the recommendation system can be improved, allowing for the recommendation of more locations of interest to users, thereby increasing user satisfaction. Simultaneously, by reducing the impact of users' real-time location information on the spatial dimension through non-localized operations, long-term dependencies can be calculated more accurately. Attached Figure Description
[0078] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:
[0079] Figure 1 This is a schematic diagram of the process structure of an embodiment of the present invention.
[0080] Figure 2 This is a schematic diagram of the structure of the recommended model in an embodiment of the present invention. Detailed Implementation
[0081] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0082] like Figure 1 and 2 As shown, an interest point recommendation method based on residual spatiotemporal cooperative networks is presented, the method comprising:
[0083] S100: Obtain the user's historical points of interest and current points of interest, sort the historical points of interest and current points of interest in chronological order, and generate a historical point of interest trajectory sequence and a current point of interest trajectory sequence.
[0084] It should be noted that the set of users is U = {u1, u2, ..., u...} |U| Let |U| represent the number of users in the user set, and let the set of all user interest points be P = {p1, p2, ..., p...}. |M|Let |M| represent the number of points of interest in the set of points of interest, and each point of interest is geocoded by a tuple, i.e., (Lon, Lat). Each user u∈U has at least one available sequence of point of interest trajectories, represented by the set of point of interest trajectory sequences T={T1,T2,...,T...}. |N| Let T be the trajectory sequence of the current point of interest, where N represents the index of the trajectory sequence. Each trajectory T in T... m ∈T contains the time-ordered check-in sequence of user u in U, i.e. And p∈P. The historical point of interest trajectory sequence for each user can be represented by T. h ={T1,T2,...,T |N-1| The sequence of trajectories of the current point of interest arranged in chronological order is defined as the trajectory sequence of the current point of interest. Where P t-1 This represents the most recent point of interest for user u at time t-1.
[0085] S200. Establish a time collaboration matrix and use the historical point of interest trajectory sequence and time collaboration matrix to capture the user's long-term dependence.
[0086] It's important to note that long-term user dependence can be visualized using a time-based collaboration matrix. This matrix is a two-dimensional table where each row and column represents a different area or topic a user might be interested in. Each cell in the matrix represents the number of times a user checks in between two points of interest. By calculating the weights of the time-based collaboration matrix, the user's long-term dependence on each point of interest can be accurately derived. Using the time-based collaboration matrix not only provides quantitative data support but also clearly reveals user behavior patterns and interest preferences, enabling this invention to more accurately understand users' long-term dependence.
[0087] S300. Based on the current interest point trajectory sequence, use a skip learning algorithm to capture the user's short-term dependence.
[0088] It should be noted that the process of determining the correlation between discrete points of interest (POIs) on a map using the skip learning algorithm is as follows: A distance matrix is constructed based on the current POI trajectory sequence. A clustering algorithm is then used to perform cluster analysis on the distance matrix, grouping similar POIs into the same cluster. Based on this cluster analysis, the skip learning algorithm can be used to process the check-in data. For each cluster, a recommendation model is trained and used to predict the category of new POIs. If the category of a new POI differs from that of the current POI, the process of training the recommendation model can be skipped, thus improving efficiency. After processing by the skip learning algorithm, correlation analysis can be performed on the check-ins within each cluster. By calculating metrics such as the number of check-ins and the time interval within a cluster, the correlation between clusters can be evaluated. For example, if two clusters have a large number of check-ins and short time intervals, they can be considered to have a high correlation.
[0089] S400. Use a sample balancer to balance the user's long-term dependence and the user's short-term dependence, and output the recommendation results.
[0090] It should be noted that the sample balancer can balance the data according to the distribution of different datasets to improve the accuracy and reliability of the data. In this invention, the sample balancer weights the long-term and short-term dependencies of each user to obtain the final recommendation data. Various strategies can be used for weighting, such as weighting long-term and short-term dependencies with different weights. Specifically, the sample balancer can perform statistical analysis on each user's long-term and short-term dependencies to determine their respective importance, and then weight them according to this importance to obtain the final recommendation data. Furthermore, the sample balancer can adjust the weighting method according to the distribution of different datasets. For example, if a user has relatively little long-term dependency data, the sample balancer can increase the weight of short-term dependency data to maintain the balance between historical and current interest point trajectory sequences. By using the sample balancer, the accuracy and reliability of the data can be improved, thereby better achieving the goal of personalized recommendations. At the same time, the sample balancer can also help improve the interpretability and maintainability of the algorithm, making the algorithm more robust and reliable.
[0091] In summary, this invention captures users' long-term and short-term dependencies by utilizing historical point-of-interest trajectory sequences and temporal collaboration matrices, and balances these dependencies using a sample balancer. This improves the accuracy of the recommendation system, allowing for the recommendation of more locations of user interest and ultimately increasing user satisfaction. Furthermore, by reducing the impact of users' real-time location information on the spatial dimension through delocalized operations, long-term dependencies can be calculated more accurately.
[0092] As an optional embodiment of the present invention, the establishment of the time cooperation matrix may include:
[0093] S201. Divide the same time period into multiple time steps, calculate the time similarity between different users in the first time step and the second time step, and use the time similarity to recommend the interest points of the second user who is highly similar to the first user to the first user.
[0094] The formula for calculating the time similarity is:
[0095]
[0096] Where, α i,j This represents the temporal similarity between the first time step i and the second time step j.
[0097] D i This represents the set of points of interest at time step i;
[0098] D j This represents the set of points of interest at time step j;
[0099] |D i ∩D j | indicates D i and D j All points of interest with the same internal characteristics are represented by a modular arithmetic operation.
[0100] |D j | indicates D j Perform mold taking;
[0101] |D i | indicates D i Perform mold taking;
[0102] Specifically, since time largely determines the popularity of a point of interest at a given time, this invention constructs a time collaboration matrix that reflects the similarity between different time steps. Each week is proportionally divided into 48 time steps, with 24 time steps for weekdays and 24 time steps for weekends. Consider a set of points of interest. This indicates that the set of interest points contains points of interest that the user has not visited. Then, the similarity α between the i-th time step and the j-th time step is calculated based on the set of interest points. i,j The more shared points of interest between two time steps, the higher their similarity. When providing recommendations to users, the points of interest visited by users more similar to the user within the same time step will be given higher weight. This recommendation method can effectively improve the accuracy of recommendations and user satisfaction.
[0103] S202. Calculate the temporal collaboration weight, and measure the similarity of interest points within the first time step and the second time step based on the temporal collaboration weight;
[0104] The formula for calculating the time collaboration weight is as follows:
[0105]
[0106]
[0107] Among them, t h The current representation of the user's historical points of interest trajectory;
[0108] T h This is a sequence of the user's historical points of interest.
[0109] |T h |A collection of historical interest point trajectory sequences for the user;
[0110] W t Weighting based on time collaboration;
[0111] To decode historical trajectories;
[0112] exp() represents an exponential function with base e.
[0113] α c This is the time similarity matrix;
[0114] D represents t The time similarity matrix;
[0115] D t D represents i and D j A set;
[0116] D represents i The time similarity matrix;
[0117] c represents the current state;
[0118] h represents history.
[0119] It's important to note that in step S22, the calculation of the temporal collaboration weight measures the similarity of interest points across different time steps based on the current representation of the user's historical interest point trajectory and the temporal collaboration weight. Specifically, the user's historical interest point trajectory sequence is used to represent the user's interests. By decoding the historical trajectory, the user's interest point trajectory within the current time step can be obtained. Then, the temporal collaboration weight and the temporal similarity matrix are used to calculate the temporal collaboration weight. This weight reflects the similarity of interest points across different time steps, thereby helping the recommendation system determine more accurate recommendation results.
[0120] As an optional embodiment of the present invention, the step of capturing the user's long-term dependency using the historical point of interest trajectory sequence and the time collaboration matrix may include:
[0121] S203. Use an encoder to map the historical interest point trajectory sequence of each time step onto a low-dimensional projection.
[0122] The mapping formula is:
[0123]
[0124] in, The result of embedding the trajectory sequence of i-1 historical points of interest after processing by the encoder;
[0125] For the embedding of the trajectory sequence of i-1 historical points of interest;
[0126] T i-1 This represents the (i-1)th sequence in the historical interest point trajectory sequence;
[0127] Encoder stands for encoder;
[0128] It should be noted that long-term preferences provide information on how users evolve and adapt over time, revealing a persistent behavioral pattern. This invention designs an encoder-decoder structure to capture users' long-term preferences. The encoder will convert the data at each time step... Mapping to a low-dimensional projection of size r, which is significantly smaller than the time width R, not only reduces the data dimensionality but also preserves important long-term dependency information. This design makes the recommendation model more compact and computationally efficient.
[0129] S204. Embed the historical point of interest trajectory sequence into the residual linear layer, and use the residual linear layer to obtain the user's long-term dependency.
[0130] It's important to note that residual linear layers are a method for capturing linear relationships between features, allowing for better understanding of users' long-term dependencies. Specifically, by inputting the embedding of historical interest point trajectory sequences into a residual linear layer, the layer learns to extract linear dependencies in users' long-term preferences. These linear dependencies reflect the stability and regularity of users' long-term behavioral patterns. By utilizing the learning and extraction capabilities of residual linear layers, recommendation models can better understand users' long-term interests and behavioral patterns. This understanding and insight can help recommendation models provide users with more accurate and personalized recommendations and services, improving user satisfaction.
[0131] As an optional embodiment of the present invention, the step of obtaining the user's long-term dependency using the residual linear layer may include:
[0132] S241. Set up two multilayer perceptrons and use the first perceptron to calculate the dynamic covariates.
[0133] The calculation formula is:
[0134]
[0135] in, Represents the trajectory sequence T at time t. i-1 r-dimensional dynamic covariates;
[0136] The result of embedding the trajectory sequence of i-1 historical points of interest after processing by the encoder;
[0137] MPL stands for Multilayer Perceptron;
[0138] like Figure 2 As shown, it should be noted that in this embodiment, the present invention uses two multilayer perceptrons. The first perceptron does not apply an activation function and is used to calculate dynamic covariates. The first perceptron is a neural network capable of processing time-series data and can capture linear features in historical interest point trajectory sequences. During training, the present invention uses an encoder to process the historical interest point trajectory sequence and then directly inputs the result into the first perceptron to calculate the dynamic covariates.
[0139] S242. Calculate the density of the user's historical interest point trajectory using the activation function and the second multilayer perceptron.
[0140] The calculation formula is:
[0141]
[0142]
[0143] in, Dense expressions representing historical trajectories;
[0144] ReLU represents the activation function;
[0145] It is an intermediate expression of the historical trajectory;
[0146] LayerNorm represents the normalization operation;
[0147] Indicates residual connection;
[0148] Dropout represents a random deactivation operation.
[0149] like Figure 2 As shown, it's important to note that the second multilayer perceptron uses an activation function. Activation functions capture non-linear features in historical interest point trajectory sequences. Introducing non-linear factors through activation functions enhances the learning ability of the recommendation model. In recommendation models, user behavior and interest point attributes often exhibit complex non-linear relationships; activation functions help the model learn and process this data better. Capturing different data features: the multilayer perceptron using activation functions captures the non-linear features of the input data, while the other perceptron without activation functions captures the linear features. This setup allows the recommendation model to learn both non-linear and linear features simultaneously, leading to a more comprehensive understanding of the data. Using two perceptrons also allows raw data to be input into two different network structures for processing. This increases the generalization ability of the recommendation model, improving its adaptability and robustness to new data. The two perceptrons play different roles in the recommendation system. The perceptron using activation functions learns richer features and patterns, providing more accurate recommendation results. The perceptron without activation functions extracts more intuitive and simpler features, providing auxiliary support for the recommendation results. This setup helps improve the diversity and personalization of the recommendation system.
[0150] As an optional embodiment of the present invention, the step of capturing users' long-term dependencies using the historical point of interest trajectory sequence and the time collaboration matrix may further include:
[0151] S205, Stack and tile the dynamic covariates of the past and future, and then... The mapping formula for the vector at each time step is as follows:
[0152]
[0153] in, for Encoded embedding after processing by the residual linear layer;
[0154] n represents the number of residual linear layers;
[0155] RLL() indicates processing with a residual linear layer.
[0156] It should be noted that stacking and tiling are used to handle data with different dimensions and shapes. Stacking refers to superimposing multiple datasets along a dimensional direction to increase the depth and complexity of the data. Tiling refers to expanding a dataset along a dimensional direction to increase the breadth and diversity of the data. In this invention, stacking and tiling past and future dynamic covariates means processing the past and future dynamic covariates separately and then superimposing or expanding them in a certain order to form a more complex data structure. This processing method can capture users' long-term dependencies because users' interests and behaviors change over time. By processing past and future dynamic covariates, we can better understand users' behavioral patterns and preferences. In this invention, mapping the encoded embeddings after residual linear layer processing to vectors at each time step means arranging the processed encoded embedding data in the order of time steps to form a series of vectors, each vector corresponding to a time step. This processing method makes data processing more flexible and diverse, and can also better capture users' behavioral patterns and preferences.
[0157] S206. Utilize a multi-head attention mechanism to calculate the salience of different interest points in the user's historical interest point trajectory or current interest point trajectory. The calculation formula is as follows:
[0158]
[0159] in, This represents the output of the multi-head attention mechanism;
[0160] This represents the embedding of a user's historical interest point trajectory or the embedding of the current interest point trajectory encoded by the residual linear layer.
[0161] γ represents the current or historical state, with c representing the current state and h representing the historical state.
[0162] MultiheadAtt represents the multi-head attention mechanism.
[0163] It should be noted that replacing γ with c calculates the saliency of different interest points in the current interest point trajectory, while replacing γ with h calculates the saliency of different interest points in the historical interest point trajectory. Multi-head attention is a neural network structure that can simultaneously monitor information from the user's historical or current interest point trajectories and integrate this information into a unified output. In this invention, the multi-head attention mechanism is used to calculate the saliency of different interest points in the user's historical or current interest point trajectories, capturing the complex relationships and interactions between different interest points. By inputting the user's historical and current interest point trajectories into the multi-head attention mechanism for processing, saliency scores at different time steps can be obtained, thereby providing a better understanding of the user's behavioral patterns and preferences.
[0164] S207. Use a decoder to map the output of the multi-head attention mechanism to the same dimension as the historical interest point trajectory sequence;
[0165] The calculation formula is:
[0166]
[0167] in, This represents the output of the decoder;
[0168] This represents the output of the historical interest point trajectory after processing by the multi-head attention mechanism.
[0169] Decoder() indicates decoder processing;
[0170] It should be noted that in this invention, the decoder maps the output of the multi-head attention mechanism to the same dimension as the historical interest point trajectory sequence. This allows for the capture of the correspondence between saliency scores at different time steps and the historical interest point trajectories, thereby providing a better understanding of user behavior patterns and preferences. By comparing and analyzing the decoder's output with the historical interest point trajectory sequence, more accurate results can be obtained, leading to more personalized recommendation services.
[0171] S208. Use global residual connections to combine the historical interest point trajectory sequence with the decoder output;
[0172] The calculation formula is as follows:
[0173]
[0174] in, This is the output after global residual connection;
[0175] For the embedding of the trajectory sequence of i-1 historical points of interest;
[0176] Indicates residual connection;
[0177] LayerNorm() represents layer normalization;
[0178] It should be noted that global residual connections can combine historical interest point trajectory sequences with the decoder output, thereby combining users' long-term dependencies with current personalized needs, further improving the accuracy and personalization of recommendation results. In step S28, global residual connections are used to combine historical interest point trajectory sequences with the decoder output to form a more complex data structure. This processing method can capture users' long-term behavioral patterns and preferences, while also taking into account users' personalized needs and changes. Through global residual connection processing, recommendation results can be more accurate and personalized, improving user satisfaction.
[0179] S209. Use average pooling to retain all user information;
[0180] The formula for average pooling is:
[0181]
[0182] Among them, t n This represents the current trajectory sequence of the user's points of interest;
[0183] This represents the result after the current point of interest trajectory has been processed sequentially through the multi-head attention mechanism, decoder, and global residual connection.
[0184] |T n | indicates taking the modulus of the current trajectory sequence of points of interest;
[0185] like Figure 2 As shown, it should be noted that the purpose of average pooling is to retain all user information in order to better reflect the user's real-time behavioral dependencies. Specifically, average pooling averages the results of the user's current interest point trajectory sequence after passing through a multi-head attention mechanism, a decoder, and a global residual connection to obtain the user's long-term dependencies.
[0186] S210. Utilize non-localized operations to reduce the impact of users' real-time location information on spatial dimensions, and calculate users' long-term dependencies:
[0187] The non-localized operation formula is:
[0188]
[0189]
[0190]
[0191] Where, d n,h Indicates the location of the point of interest (l) t-1 With interest point trajectory t h The distance between them;
[0192] Er represents the Earth's radius;
[0193] Both b and a represent intermediate variables;
[0194] Lon indicates longitude;
[0195] lat represents latitude;
[0196] Longitude indicating a point of historical interest;
[0197] Longitude indicating the location of the point of interest;
[0198] The latitude representing the location of the point of interest;
[0199] Dimensions representing points of historical interest;
[0200] The formula for calculating long-term user dependency is:
[0201]
[0202] in, This indicates long-term user dependence;
[0203] W h This represents the trainable projected weight matrix;
[0204] n-1 represents the length of the user's historical points of interest trajectory;
[0205] Indicates matrix transpose;
[0206] t h t represents the user's historical points of interest trajectory. h ∈{t1,t2,...,t n-1}
[0207] It should be noted that, in order to consider the historical interest point trajectory T of each user h ∈{T1,T2,...,T n-1 For the current point of interest trajectory T n To mitigate the impact of real-time user location information on spatial dimensions, this invention utilizes geographic non-local operations. n,hThe median coordinates can be calculated using the Havesing algorithm. To consider typical user access patterns, this invention employs a mean-based approach to define the median coordinates. and This is used to calculate long-term user dependency.
[0208] As an optional embodiment of the present invention, the step of capturing the user's short-term dependence using a skip learning algorithm based on the current point of interest trajectory sequence includes:
[0209] S301. Establish a distance matrix between each pair of interest points in the current interest point trajectory sequence described by the user, and use the distance matrix to determine the distance that the skip learning algorithm needs to skip.
[0210] It's worth noting that by using a distance matrix to determine the distances that the skip learning algorithm needs to skip, it's possible to better capture users' short-term dependencies and avoid unnecessary redundancy and complexity in data processing. It also helps determine which data between points of interest can be skipped and which require focused attention. Furthermore, the distance matrix can be used to analyze and model user behavior, such as analyzing user interest preferences and travel paths. This information helps businesses better understand user needs and behaviors, thereby providing more personalized and precise services and products.
[0211] S302. Process the current interest point trajectory sequence and establish an input set;
[0212] It should be noted that processing the current point of interest trajectory sequence includes sorting the trajectory sequence according to a set standard to obtain an ordered trajectory sequence, and extracting features reflecting the user's point of interest, such as location and time, from the trajectory sequence. Missing data or outliers in the trajectory sequence can be filled or removed to ensure data consistency and integrity. The processed trajectory sequence is used as the input set for subsequent recommendation model prediction. Through the above processing steps, an effective input set can be established. (where 1 < σ < ... < θ < t-1), which provides an accurate data foundation for subsequent user behavior analysis and interest point prediction, and improves the prediction accuracy of the recommendation model.
[0213] The formula for calculating the skip learning algorithm is:
[0214]
[0215]
[0216] Among them, h t-1 By s θ and s t-1 Calculated;
[0217] s θ This represents the point of interest at time step θ.
[0218] s t-1 This represents the point of interest at time step t-1;
[0219] {s θ ,s t-1} represents s θ and s t-1 The shortest distance between them;
[0220] n R Indicates the number of linear residual layers;
[0221] RLL represents the residual linear layer;
[0222] This indicates the user's short-term dependence;
[0223] h θ Indicates the result of the intermediate sequence;
[0224] W q and u q Initialize the matrix randomly;
[0225] A t-1 It is a weighted vector;
[0226] W′,W″,b′,b″ are learnable parameters.
[0227] As an optional embodiment of the present invention, the calculation formula for balancing the user's long-term dependence and the user's short-term dependence using the sample balancer is:
[0228]
[0229] Where F represents the probability set of each point of interest;
[0230] W s This represents a trainable projection matrix of all interest points;
[0231] R represents a trainable matrix, and W s ∈R |P|×2d Each point of interest can be projected;
[0232] P represents the set of current points of interest;
[0233] |P| represents taking the modulus of P;
[0234] d represents the number of columns in the training matrix;
[0235] ∈ indicates an adjustable parameter used to set the weighting ratio between long-term and short-term factors;
[0236] This indicates a join operation.
[0237] It's important to note that the sample balancer is a computational method used to balance long-term and short-term user dependencies. It achieves this by adjusting the probability set of each point of interest. By using the sample balancer, it avoids situations where one class of samples is too large or too small, thus making model training more stable and reliable. In practice, the sample balancer can be used to adjust the probability set of user interest points to balance long-term and short-term dependencies. For example, if a user has strong long-term dependencies, the sample balancer can increase the weight of the user's historical behavior on their current behavior; conversely, if a user has strong short-term dependencies, the sample balancer can decrease the weight of the user's historical behavior on their current behavior. Furthermore, the sample balancer can be customized according to different application scenarios and needs. For example, the parameters of the sample balancer can be adjusted based on different industries, user groups, and time periods to achieve a better balance.
[0238] As an optional embodiment of the present invention, the calculation formula for the final output recommendation data is optionally:
[0239] Next poi =Max(F)
[0240] Among them, Next poi Let F represent the point of interest that the next user is most likely to visit, and Max(F) represent the point of interest with the largest probability in the probability set.
[0241] It should be noted that the formula used to calculate the final recommendation data effectively generates the final recommendation data based on the data from the user's interest probability set, enabling users to find content they are interested in more accurately. Furthermore, this formula effectively reduces computational load, improves computational efficiency, and makes the recommendation system respond faster.
[0242] As an optional embodiment of the present invention, the recommendation model may be evaluated using a loss function.
[0243]
[0244] in, This represents the loss rate, which is the error rate of the model's predictions.
[0245] K represents the total number of training samples, that is, the total number of samples used to train the recommendation model;
[0246] j represents a variable, which can be a parameter in the recommendation model or an independent variable;
[0247] p j This represents the probability that the recommendation model correctly identifies the true point of interest based on j training samples;
[0248] log(p j ) indicates that for p j Take the logarithm.
[0249] It should be noted that, due to W s Let W represent a trainable matrix. s ∈R |L|×2d Each point of interest can be projected, thus maximizing the probability that a target user will access a point of interest at the next time step t. This invention uses p∈F to represent the probability of a true point of interest. The above formula is used to calculate the loss rate of the recommendation model, where the loss rate measures the model's prediction error. Using the objective function allows for a better evaluation of the recommendation model's performance. Furthermore, this invention allows for adjusting the parameters in the above formula as needed to achieve better recommendation results. For example, if the model's loss rate is found to be too high, the total number of training samples K can be increased, or the value of the independent variable j can be adjusted to improve the model's performance. Additionally, if the model's accuracy p... j If the accuracy is low, the structure or parameters of the recommendation model can be adjusted to improve the model's accuracy.
[0250] Although embodiments of the invention have been shown and described, those skilled in the art will understand 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 claims and their equivalents.
Claims
1. An interest point recommendation method based on residual spatiotemporal cooperative networks, characterized in that, The method includes: S100: Obtain the user's historical points of interest and current points of interest, sort the historical points of interest and current points of interest in chronological order, and generate a historical point of interest trajectory sequence and a current point of interest trajectory sequence. S200. Establish a time-cooperation matrix, and use the historical point-of-interest trajectory sequence and the time-cooperation matrix to capture the user's long-term dependency; the step of using the historical point-of-interest trajectory sequence and the time-cooperation matrix to capture the user's long-term dependency includes: S205, Stack and tile the dynamic covariates of the past and future, and then... The mapping formula for the vector at each time step is as follows: ， in, This is the encoded embedding after processing by the residual linear layer. Indicates the number of residual linear layers; S206. Utilize a multi-head attention mechanism to calculate the salience of different interest points in the user's historical interest point trajectory or current interest point trajectory. The calculation formula is as follows: ， in, This represents the output of the multi-head attention mechanism. This represents the embedding of a user's historical point of interest trajectory or the embedding of the current point of interest trajectory encoded by a residual linear layer. Indicates the present or the past; S207. Use a decoder to map the output of the multi-head attention mechanism to the same dimension as the historical interest point trajectory sequence; The calculation formula is: ， in, This represents the output of the decoder. This represents the output of the historical interest point trajectory after processing by the multi-head attention mechanism. S208. Use global residual connections to combine the historical interest point trajectory sequence with the decoder output; The calculation formula is as follows: ， in, This is the output after global residual connection; S209. Use average pooling to retain all user information; The formula for average pooling is: ， in, This represents the current trajectory sequence of the user's points of interest. This represents the result after the current point of interest trajectory has been processed sequentially through the multi-head attention mechanism, decoder, and global residual connection. S210. Utilize non-localized operations to reduce the impact of users' real-time location information on spatial dimensions, and calculate users' long-term dependencies: The non-localized operation formula is: ， in, Indicates the location of the point of interest Trajectory of Points of Interest The distance between them Represents the Earth's radius. Indicates longitude. Indicates latitude; The formula for calculating long-term user dependency is: ， in, This indicates long-term user dependence. This represents the trainable projected weight matrix; S300. Based on the current interest point trajectory sequence, use a skip learning algorithm to capture the user's short-term dependence. S400. Use a sample balancer to balance the user's long-term dependence and the user's short-term dependence, and output the recommendation results.
2. The interest point recommendation method based on residual spatiotemporal cooperative network as described in claim 1, characterized in that, The establishment of the time collaboration matrix includes: S201. Divide the same time period into multiple time steps, calculate the time similarity between different users in the first time step and the second time step, and use the time similarity to recommend the interest points of the second user who is highly similar to the first user to the first user. The formula for calculating the time similarity is: ， in, Indicates the first time step. Second time step Temporal similarity between them Indicates the time step A set of points of interest Indicates the time step A set of points of interest; S202. Calculate the temporal collaboration weight, and measure the similarity of interest points within the first time step and the second time step based on the temporal collaboration weight; The formula for calculating the time collaboration weight is as follows: ， ， in, This is the current representation of the user's historical points of interest trajectory. For the user's historical points of interest trajectory sequence, As a time-based collaboration weight, To decode historical trajectory, This is the time similarity matrix.
3. The interest point recommendation method based on residual spatiotemporal cooperative networks as described in claim 2, characterized in that, The method of capturing users' long-term dependencies using the historical point of interest trajectory sequence and time collaboration matrix includes: S203. Use an encoder to map the historical interest point trajectory sequence of each time step onto a low-dimensional projection. The mapping formula is: ， in, for The result of embedding a sequence of historical interest point trajectories after processing by an encoder. for Embedding of a historical interest point trajectory sequence; S204. Embed the historical point of interest trajectory sequence into the residual linear layer, and use the residual linear layer to obtain the user's long-term dependency.
4. The interest point recommendation method based on residual spatiotemporal cooperative network as described in claim 3, characterized in that, The method of obtaining the user's long-term dependency using the residual linear layer includes: S241. Set up two multilayer perceptrons and use the first perceptron to calculate the dynamic covariates. The calculation formula is: ， in, Indicates time Time trajectory sequence of Dimensional dynamic covariates; S242. Calculate the density of the user's historical interest point trajectory using the activation function and the second multilayer perceptron. The calculation formula is: ， in, Dense expressions representing historical trajectories, It is an intermediate expression of the historical trajectory. This indicates a residual connection.
5. The interest point recommendation method based on residual spatiotemporal cooperative network as described in claim 1, characterized in that, The step of capturing short-term user dependence using a skip learning algorithm based on the current point of interest trajectory sequence includes: S301. Establish a distance matrix between each pair of interest points in the current interest point trajectory sequence described by the user, and use the distance matrix to determine the distance that the skip learning algorithm needs to skip. S302. Process the current interest point trajectory sequence and establish an input set; The formula for calculating the skip learning algorithm is: ， ， in, This indicates a user's short-term dependence.
6. The interest point recommendation method based on residual spatiotemporal cooperative network as described in claim 1, characterized in that, The calculation formula for balancing the user's long-term dependence and short-term dependence using a sample balancer is as follows: ， in, This represents the probability set for each point of interest.
7. The interest point recommendation method based on residual spatiotemporal cooperative network as described in claim 1, characterized in that, The formula for calculating the final recommended output data is as follows: ， in, This indicates the next point of interest that the user is most likely to visit. This represents the point of interest with the highest probability in the probability set.