A point of interest recommendation method

Abstract

The application discloses a point of interest recommendation method, comprising the following steps: acquiring a user set, a location point of interest set and a user check-in sequence set; inputting an improved space-time fusion embedding module to obtain fused space-time embedding information; obtaining user basic trajectory embedding information by using the user check-in sequence set; adding an absolute position embedding matrix to the user basic trajectory embedding information to obtain user trajectory embedding information; and inputting the fused space-time embedding information and the user trajectory embedding information into a Transformer coding module at the same time to obtain a prediction probability of a next location point of interest. The application selects a suitable embedding dimension to fuse time interval information and space distance information, simultaneously codes the relationship between user trajectory and space-time information, deeply mines the diversity of user interest and improves the overall interest distribution of the user. The application solves the problem that the individual diversity of the user is insufficient in the existing recommendation system and the more accurate and diversified location point of interest recommendation cannot be realized.

Description

Technical Field

[0001] This invention relates to a point-of-interest recommendation method, belonging to the field of data processing technology. Background Technology

[0002] In personalized recommendation systems, the diversity of user preferences is one of the key indicators for evaluating recommendation quality. With the advent of the information age, users' preferences for goods, services, or locations are becoming increasingly diversified, requiring recommendation systems to accurately capture and respond to these diverse needs. For example, in e-commerce, users may be interested in fashion apparel, technology products, and health foods simultaneously; in tourism, users may enjoy both historical and cultural cities and natural landscapes. Therefore, a successful personalized recommendation system needs to understand and utilize the diversity of user preferences to provide users with rich and varied recommendations to meet their different needs and interests.

[0003] In location-of-interest (POI) recommendations, in addition to users' historical browsing history and preferences, spatial distance is also an important factor. Users tend to prefer visiting locations that are closer to them because short-distance travel is more convenient and time-saving. For example, when planning a day's itinerary, users may prioritize nearby restaurants, cafes, or attractions that match their interests.

[0004] To address the aforementioned needs, existing methods consider the impact of temporal factors on location-based interest (PPI) recommendations, inputting temporal information from the user's check-in sequence into a Transformer model to predict the user's next likely PPI visit. However, these methods often only consider the impact of temporal factors on the diversity of PPI recommendations, neglecting spatial distance factors, which also significantly influence user travel. Furthermore, existing methods employ traditional dot-product matching functions when predicting the user's next likely PPI visit, a method limited by the length of the input vector and lacking flexibility. This leads to insufficient research on the individual diversity of user preferences, increased computational complexity, and neglect of the importance of unvisited PPIs.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to integrate time interval and spatial distance information, optimize the spatiotemporal diversity and individual diversity of user preferences, and achieve more accurate and diversified location interest point recommendations.

[0007] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0008] This invention provides a method for recommending points of interest, comprising:

[0009] Get the user set, the location interest point set and the user check-in sequence set, input them into the improved spatiotemporal fusion embedding module, and get the fused spatiotemporal embedding information;

[0010] After arranging the user check-in sequence sets in chronological order, the user trajectory embedding layer of the spatiotemporal fusion embedding module uses the user check-in sequence set to obtain the user's basic trajectory embedding information, and then embeds the absolute position into the matrix. Add to the user's basic trajectory embedding information to obtain the user's trajectory embedding information;

[0011] The fused spatiotemporal embedding information and user trajectory embedding information are simultaneously input into the Transformer encoding module to obtain the predicted probability of the next location point of interest.

[0012] The improved method for obtaining the spatiotemporal fusion embedding module includes:

[0013] Divide the time interval between two points of interest accessed by a user into multiple time slots;

[0014] The time slot interval between two points of interest is used as a time embedding element. The Sin activation function is used to encode the time slot interval between all points of interest to obtain the final time embedding information.

[0015] The spatial distance between any two adjacent embedded elements between two points of interest is taken as a spatial embedded element. The spatial distance between all adjacent embedded elements is calculated using the semi-sine formula to obtain the final spatial embedding information.

[0016] By using the Poisson distribution to predict the probability of a user moving from one point of interest to another, we obtain the expression for the probability prediction of the point of interest location.

[0017] Based on the probability prediction expression of point of interest location, the parameters representing spatial distance information in the probability prediction expression of point of interest location are replaced with the final spatial embedding information, and the parameters representing time interval information in the probability prediction expression of point of interest location are replaced with the final temporal embedding information, thus obtaining fused spatiotemporal embedding information.

[0018] Furthermore, the acquired user set, location interest point set, and user check-in sequence set are used as initial data and input into the improved spatiotemporal fusion embedding module in batches, and the user check-in sequence is arranged in chronological order.

[0019] Furthermore, methods for dividing the time interval between two locations of interest accessed by a user into multiple time slots include:

[0020] The time interval between two locations of interest visited by a user is divided into... In each time slot, the time value of any scalar can be projected onto... In any of the time slots.

[0021] Furthermore, the time slot interval between two points of interest is used as a temporal embedding element. The Sin activation function is then used to encode the time slot intervals between all points of interest to obtain the final temporal embedding information. Here, the first... A time-embedded element, represented as:

[0022] ;

[0023] in, Indicates the first A time-embedded element, Indicates the first A weight vector, Indicates the first A bias vector. and These represent the time slots to which the user's check-in time belongs at location i and location j, respectively; This indicates the length of the user's check-in sequence. k has no actual meaning and represents a count.

[0024] Furthermore, the spatial distance between any two adjacent embedded elements between two points of interest is taken as a spatial embedded element. Using the semi-sine formula, the spatial distance between all adjacent embedded elements is calculated to obtain the final spatial embedding information. Here, the th... A spatially embedded element is represented as:

[0025] ;

[0026] in, Indicates the first A space-embedded element; Indicates the first The embedded element and the first Spatial distance between embedded elements.

[0027] Furthermore, by using the Poisson distribution to predict the probability of a user moving from one location point of interest to another, we obtain the expression for predicting the location probability of a point of interest, which is expressed as:

[0028] ;

[0029] in, Indicates user Visit two points of interest and The probability of; Indicates learnable parameters, Represents the exponential factor in the Poisson distribution; This indicates that the user visited two points of interest. and The time interval information between them, where log represents taking the logarithm.

[0030] Furthermore, based on the probability prediction expression for the location of points of interest, the parameters representing spatial distance information in the expression are replaced with the final spatial embedding information, and the parameters representing time interval information are replaced with the final temporal embedding information, resulting in fused spatiotemporal embedding information, which is expressed as:

[0031] ;

[0032] in, This represents the spatiotemporal embedded information of the fusion. This represents the final spatial embedding information. This represents the final spatial embedding information. This represents the exponential factor in the Poisson distribution.

[0033] Furthermore, the user trajectory embedding information is represented as follows:

[0034] ;

[0035] Where E represents user trajectory embedding information, Represents the absolute position embedding matrix. This indicates the embedded information of the user's basic trajectory.

[0036] Furthermore, the Transformer encoding module includes a two-layer location feedforward network and an equalization sampler based on an improved cross-entropy loss function; the fused spatiotemporal embedding information and user trajectory embedding information are simultaneously input into the Transformer encoding module to obtain the predicted probability of the next location interest point, including:

[0037] A symmetric masking matrix is ​​constructed, embedding user trajectory information. The fused spatiotemporal embedding information and the masking matrix are input to the head self-attention formula, and operations are performed with the parameter matrix to obtain the result. The result is then linearly projected onto... Parallel applications in a subspace of time slots Each attention function generates an output representation, which is then concatenated and projected again to obtain a connection sequence.

[0038] Input the connection sequence into a two-layer position feedforward network to obtain each element of the connection sequence;

[0039] For each element in the connection sequence, apply residual connection, layer normalization, and dropout regularization to obtain the elements after residual connection, layer normalization, and dropout regularization. Arrange all the elements after residual connection, layer normalization, and dropout regularization in order to obtain the final output sequence.

[0040] The final output sequence is converted into location interest point probabilities by using linear transformation and the Softmax function instead of the traditional dot product matching function.

[0041] The probability of the point of interest is input into an equalization sampler based on the cross-entropy loss function. The unvisited locations with negative point of interest probabilities are weighted to obtain the predicted probability of the point of interest at the next location.

[0042] Furthermore, the interest point probability is input into an equalization sampler based on the cross-entropy loss function. A weighted operation is then performed on unvisited locations with negative interest point probabilities to obtain the predicted probability of the interest point at the next location. The cross-entropy loss function is expressed as:

[0043] ;

[0044] in, express user Points of interest and points of interest All of these are initial data. , Indicates to the user Visit points of interest The probability is taken as logarithm. Indicates the length of the user's check-in sequence. Indicates that for a given user Under the movement trajectory, position The probability of being a negative sample.

[0045] Among them, in a given user Under the movement trajectory, position Probability of being a negative sample , is represented as:

[0046] ;

[0047] in, Indicates that for a given user Under the movement trajectory, position The probability of being a negative sample; This represents the temperature coefficient, used to control for deviations in the probability distribution. This indicates the exponentiation operation. and Indicate points of interest, and These represent points of interest that have not been visited.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0049] This invention, by fusing time interval and spatial distance information and selecting appropriate embedding dimensions, enhances the spatiotemporal diversity of user preferences. Simultaneously, it employs an improved Transformer encoding model to encode the relationship between the trajectories and spatiotemporal information of individual users within the same location subspace, thereby deeply exploring the individual diversity of user interests. This invention uses an improved cross-entropy loss function to combine the spatiotemporal and individual diversity of users, comprehensively improving the distribution of user interests. It solves the problem in existing recommendation systems where the lack of individual user diversity prevents more accurate and diverse location-based interest point recommendations. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the structure of an interest point recommendation method provided in an embodiment of the present invention;

[0051] Figure 2 This is a structural diagram of the improved Transformer encoding module provided in the embodiments of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0053] Example 1

[0054] This embodiment provides a method, including the following steps:

[0055] Get the user set, the location interest point set and the user check-in sequence set, input them into the improved spatiotemporal fusion embedding module, and get the fused spatiotemporal embedding information;

[0056] An improved spatiotemporal fusion embedding module is used to fuse temporal and spatial information into a unified embedding vector. By considering both temporal and spatial factors, the ability to describe user movement behavior is improved, resulting in more accurate probability prediction.

[0057] After arranging the user check-in sequence sets in chronological order, the user trajectory embedding layer of the spatiotemporal fusion embedding module uses the user check-in sequence set to obtain the user's basic trajectory embedding information, and then embeds the absolute position into the matrix. Add to the user's basic trajectory embedding information to obtain the user's trajectory embedding information;

[0058] The user trajectory embedding layer transforms a user's check-in sequence into an embedding vector, reflecting the user's movement trajectory. Basic user trajectory embedding information captures long-term movement patterns and preferences. By adding an absolute location embedding matrix, the basic user trajectory embedding information is combined with specific geographic location information. The user trajectory embedding information not only includes the user's movement trajectory but also specific geographic location information, enabling more accurate prediction of locations the user might visit in the future.

[0059] The fused spatiotemporal embedding information and user trajectory embedding information are simultaneously input into the Transformer encoding module to obtain the predicted probability of the next location point of interest.

[0060] By utilizing the Transformer encoding module and combining fused spatiotemporal embedding information with user trajectory embedding information, we can predict the next location of interest that a user may visit. By comprehensively considering the user's movement trajectory, spatiotemporal information, and specific geographic location information, we can more accurately predict the user's future movement behavior and provide users with more personalized recommendations and services.

[0061] The improved method for obtaining the spatiotemporal fusion embedding module includes:

[0062] Divide the time interval between two points of interest accessed by a user into multiple time slots;

[0063] The time slot interval between two points of interest is used as a time embedding element. The Sin activation function is used to encode the time slot interval between all points of interest to obtain the final time embedding information.

[0064] The time interval is divided into multiple time slots, and these time slot intervals are encoded to obtain time embedding information. Time embedding information can reflect the user's movement frequency and preferences in different time periods, which helps to predict the user's movement behavior in a future time period.

[0065] The spatial distance between any two adjacent embedded elements between two points of interest is taken as a spatial embedded element. The spatial distance between all adjacent embedded elements is calculated using the semi-sine formula to obtain the final spatial embedding information.

[0066] The spatial distances between points of interest are calculated and encoded to obtain spatial embedding information. This spatial embedding information reflects users' movement patterns and preferences between different locations, helping to predict where users might visit in the future.

[0067] By using the Poisson distribution to predict the probability of a user moving from one point of interest to another, we obtain the expression for the probability prediction of the point of interest location.

[0068] Based on the probability prediction expression of point of interest location, the parameters representing spatial distance information in the probability prediction expression of point of interest location are replaced with the final spatial embedding information, and the parameters representing time interval information in the probability prediction expression of point of interest location are replaced with the final temporal embedding information, thus obtaining fused spatiotemporal embedding information.

[0069] By combining temporal and spatial embedding information, and using the Poisson distribution to predict the probability of a user moving from one location point of interest to another, we obtain fused spatiotemporal embedding information. By fusing temporal and spatial information, we obtain a more comprehensive and accurate description of user movement behavior.

[0070] Example 2

[0071] like Figure 1 As shown in the figure, this embodiment introduces a point-of-interest recommendation method, including the following steps:

[0072] Step 1: Obtain the user set Location of Interest Set and user check-in sequence set As initial data.

[0073] For the acquired user set Location of Interest Set and user check-in sequence set Preprocessing is performed to filter out users and locations of interest with fewer than 10 check-ins.

[0074] Step 2: Input the initial data in batches into the improved spatiotemporal fusion embedding module to construct the spatiotemporal fusion embedding and arrange the users' check-in sequence in chronological order.

[0075] Step 3: In the spatiotemporal fusion embedding module, a time interval scaling strategy is adopted to divide the time interval into multiple time slots. At the same time, Poisson distribution is used to model the user's check-in behavior at points of interest to obtain the fused spatiotemporal embedding information. ,

[0076] The specific steps for constructing the improved spatiotemporal fusion embedding module are as follows:

[0077] Step 3.1: Divide the time interval between a user's visits to two points of interest into... Each time slot and In this context, the time value of any scalar can be projected onto... In any time slot of the time slot.

[0078] Step 3.2: Using the time slot interval between two points of interest as a temporal embedding element, the Sin activation function is used to encode the time slot intervals between all points of interest to obtain the final temporal embedding information, including:

[0079] Users from location points of interest Departure to Location Point of Interest The The formula for calculating each time-embedded element is expressed as:

[0080] (1);

[0081] in, Indicates the first A time-embedded element, Indicates the first A weight vector, Indicates the first A bias vector. and These represent the time slots to which the user's check-in time belongs at location i and location j, respectively; This indicates the length of the user's check-in sequence. k has no actual meaning and represents a count.

[0082] According to the The calculation formula for each time embedding element is used to calculate the time embedding elements when k takes values ​​from 1 to t, and the time slot interval between all adjacent embedding elements is encoded to obtain the final time embedding information. The time-embedded information , Indicates the embedding dimension. Indicates the length of the user's check-in sequence. Represents a spatial vector.

[0083] Step 3.3: Using the time slot interval between two points of interest as a temporal embedding element, the Sin activation function is used to encode the time slot intervals between all points of interest to obtain the final temporal embedding information, including:

[0084] Users from location points of interest Departure to Location Point of Interest Get points of interest latitude and longitude and points of interest latitude and longitude Inputting the data into the semi-versus formula yields the interest point. and points of interest The formula for calculating spatial distance is expressed as:

[0085] (2);

[0086] in, , , Indicates the cosine operation; Representing points of interest respectively and indicating points of interest Longitude; Representing points of interest respectively and points of interest Latitude.

[0087] Based on points of interest and points of interest The formula for calculating spatial distance is used to obtain the first... The embedded element and the first The spatial distance between the embedded elements is used as the first... The space-embedded element, the first The formula for calculating the spatially embedded elements is expressed as:

[0088] (3);

[0089] in, Indicates the first A space-embedded element; Indicates the first The embedded element and the first Spatial distance between embedded elements.

[0090] According to the The formula for calculating each spatial embedding element is used to calculate the spatial embedding elements when k takes values ​​from 1 to t, and the spatial distance between all adjacent embedding elements is obtained to obtain the final spatial embedding information. ,in, , Indicates the embedding dimension. Indicates the length of the user's check-in sequence. Represents a spatial vector.

[0091] Step 3.6: Predict users using Poisson distribution From location points of interest Departure to Location Point of Interest The probability is expressed as:

[0092] (4);

[0093] in, Indicates user Visit two points of interest and The probability of; Indicates learnable parameters, The exponential factor in the Poisson distribution can be viewed as a trade-off between controlling for temporal and spatial effects. This indicates the time interval between a user's visits to two points of interest.

[0094] To optimize the algorithm complexity, the logarithm of both sides of equation (4) is taken to obtain the probability prediction formula for the location of interest points, which is expressed as:

[0095] (5);

[0096] in, Indicates user Visit two points of interest and The probability of; Indicates learnable parameters, Represents the exponential factor in the Poisson distribution; This indicates that the user visited two points of interest. and The time interval information between them, where log represents taking the logarithm.

[0097] Step 3.7: To fully consider spatial distance factors, the learnable parameters in the probability prediction formula for point of interest locations are... Replace with spatial distance information to obtain the user after replacing with learnable parameters. From location points of interest Departure to Location Point of Interest The probability follows a Poisson distribution, expressed as:

[0098] (6);

[0099] in, This indicates that the user visited two points of interest. and Spatial distance information between them; The exponential factor in the Poisson distribution can be seen as a trade-off between controlling for temporal and spatial effects; log represents taking the logarithm. This indicates that the user visited two points of interest. and Information on the time interval between them.

[0100] Step 3.7: Based on the probability prediction expression for the location of the point of interest, replace the parameters representing spatial distance information in the expression with the final spatial embedding information, and replace the parameters representing time interval information in the expression with the final temporal embedding information to obtain the fused spatiotemporal embedding information:

[0101] The formula for predicting the probability of user visits to two points of interest locations... and Spatial distance information between Replace with the final spatial embedding information ;

[0102] The formula for predicting the probability of user visits to two points of interest locations... and Time interval information between Replace with the final time-embedded information To obtain the fused spatiotemporal embedded information ;

[0103] The fused spatiotemporal embedded information Represented as:

[0104] (7);

[0105] in, This represents the spatiotemporal embedded information of the fusion. This represents the final spatial embedding information. This represents the final spatial embedding information. The exponential factor in the Poisson distribution can be seen as a trade-off between controlling for time and space effects.

[0106] Step 4: Obtain user trajectory embedding information based on the user check-in sequence arranged in chronological order:

[0107] Step 4.1: Based on the user The check-in sequence is input into the user trajectory embedding layer to obtain the user's basic trajectory embedding information.

[0108] User A single check-in is input into the user trajectory embedding layer, and the output of the user trajectory embedding layer is obtained.

[0109] The output of the user trajectory embedding layer is represented as and .

[0110] User Sign-in sequence The input is fed into the user trajectory embedding layer to obtain the output sequence of the user trajectory embedding layer. The obtained output sequence of the user trajectory embedding layer is used as the user's basic trajectory embedding information.

[0111] The user's basic trajectory embedding information is represented as follows: and ,

[0112] Among them, e r Indicates user trajectory embedding, e u Indicates user embedding, e l This represents the embedding of interest points, where d represents the embedding dimension. Represents a spatial vector.

[0113] Step 4.2: Embed the absolute position into the matrix Add to user base trajectory embedding information The user trajectory embedding information E obtained is as follows:

[0114] (8);

[0115] in, E represents user trajectory embedded information. Represents the absolute position embedding matrix. This indicates the embedded information of the user's basic trajectory.

[0116] Step 5: Incorporate the fused spatiotemporal embedding information The user trajectory embedding information E is simultaneously input into the Transformer encoding module to obtain the predicted probability of the next location interest point. The structure of the Transformer encoding module is as follows: Figure 2 As shown, the specific steps include:

[0117] Step 5.1: Construct a symmetric masking matrix This involves masking the embedded location information and focusing on the relationship between the current point of interest itself and the unmasked location information, where t represents the length of the user's check-in sequence. Represents a spatial vector.

[0118] Step 5.2: Embedding user trajectory information fused spatiotemporal embedded information and masking matrix The input head self-attention formula is used to perform operations with the parameter matrix to obtain the result, which is then linearly projected onto... Parallel applications in each subspace Each attention function produces an output representation, which is then concatenated and projected again to obtain a final connection sequence. Where t is the length of the user's check-in sequence, Indicates the embedding dimension. Represents spatial vectors, connecting sequences Represented as:

[0119] (9);

[0120] in, Represents a learnable parameter vector. Indicates the embedding dimension. Represents a spatial vector. Indicates the number of subspaces. This represents h attention functions, and Concat means concatenation.

[0121] in, yes A self-attention head is represented as:

[0122] (10);

[0123] in, Let Q represent the query for point of interest i, K represent the key for point of interest i, and V represent the value matrix for point of interest i, respectively. Let M represent the fused spatiotemporal embedding information, M represent the symmetric masking matrix, and Att() represent the masking attention function. Denotes the embedding dimension, d h The embedding dimension of the subspace is represented. Represents a spatial vector.

[0124] Step 5.3: Connect the sequences Input a two-layer position feedforward network to obtain the connection sequence. Each element in the array; the output of the two-layer position feedforward network is represented as:

[0125] (11);

[0126] Among them, FFN( ) represents the output of the two-layer position feedforward network. Represents a connection sequence The first in One element; , These represent the weights and biases of the inner layer of the position feedforward network, respectively. , represents the weights and biases of the outer layer of the position feedforward network, respectively, and ReLU represents the activation function.

[0127] Step 5.4: Connect the sequences Each element in the matrix is ​​subjected to residual join, layer normalization, and dropout regularization to obtain the elements after residual join, layer normalization, and dropout regularization, represented as:

[0128] (12);

[0129] in, This indicates that regularization should be discarded; This represents a two-layer position feedforward network; Representation layer normalization, Represents a connection sequence The first in One element;

[0130] Representation layer normalization, specifically represented as:

[0131] (13);

[0132] in, Representation layer normalization, Represents a connection sequence The first in One element, They represent The mean and variance; These represent the learned scaling factor and bias term, respectively; This indicates element-wise multiplication. This represents the minimum value and is used to prevent the denominator from being 0.

[0133] Arrange all residual connected, layer normalized, and discarded regularized elements in order to obtain the final output sequence. .

[0134] Step 5.5: Replace the traditional dot product matching function with a linear transformation and the Softmax function to transform the final output sequence. Converted to location interest point probabilities, it is expressed as:

[0135] (14);

[0136] in, Indicates user Predicted scores of accessible candidate location points of interest; Indicates the first A positional weight matrix, and , Indicates the embedding dimension. Indicates the length of the user's check-in sequence. Represents a spatial vector. This represents the output sequence, and softmax represents the activation function.

[0137] Step 6: Input the location interest point probability into the equalization sampler based on the cross-entropy loss function. Perform a weighted operation on unvisited locations with negative interest point probabilities to improve the cross-entropy loss function and obtain the predicted probability of the next location interest point.

[0138] The improved cross-entropy loss function is expressed as:

[0139] (15);

[0140] in, Indicates user Points of interest and points of interest All of these are initial data. , Indicates to the user Visit points of interest The probability is taken as logarithm. Indicates the length of the user's check-in sequence. Indicates that for a given user Under the movement trajectory, position The probability of being a negative sample.

[0141] Among them, in a given user Under the movement trajectory, position Probability of being a negative sample , is represented as:

[0142] (16);

[0143] in, Indicates that for a given user Under the movement trajectory, position The probability of being a negative sample; This represents the temperature coefficient, used to control for deviations in the probability distribution. This indicates the exponentiation operation. and Indicate points of interest, and These represent points of interest that have not yet been visited. This not only avoids overlooking the importance of unvisited points of interest but also emphasizes the location-specific points of interest that contain more information.

[0144] Businesses and organizations can better understand user needs, optimize resource allocation, and provide better personalized services by predicting the probability of the next location's point of interest.

[0145] Example 3

[0146] A computer-readable storage medium storing computer instructions thereon, characterized in that the computer instructions, when executed by a processor, implement the steps of the method described in Embodiment 1 or 2 above.

[0147] Example 4

[0148] The present invention also provides a computer program product, including computer instructions, characterized in that the computer instructions, when executed by a processor, implement the steps of the methods described in Embodiment 1 or 2 above.

[0149] In summary, this invention, by fusing time interval and spatial distance information and selecting an appropriate embedding dimension, enhances the spatiotemporal diversity of user preferences. Simultaneously, it employs an improved Transformer encoding model to encode the relationship between the trajectories and spatiotemporal information of individual users within the same location subspace, thereby deeply exploring the individual diversity of user interests. This invention uses an improved cross-entropy loss function to combine the spatiotemporal and individual diversity of users, comprehensively improving the distribution of user interests. It solves the problem in existing recommendation systems where the lack of individual user diversity prevents more accurate and diverse location-based interest point recommendations.

[0150] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0154] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. An interest point recommendation method, characterized in that, include: Get the user set, the location interest point set and the user check-in sequence set, input them into the improved spatiotemporal fusion embedding module, and get the fused spatiotemporal embedding information; After arranging the user check-in sequence sets in chronological order, the user trajectory embedding layer of the spatiotemporal fusion embedding module uses the user check-in sequence set to obtain the user's basic trajectory embedding information, and then embeds the absolute position into the matrix. Add to the user's basic trajectory embedding information to obtain the user's trajectory embedding information; The fused spatiotemporal embedding information and user trajectory embedding information are simultaneously input into the Transformer encoding module to obtain the predicted probability of the next location point of interest. The improved method for obtaining the spatiotemporal fusion embedding module includes: Divide the time interval between two points of interest accessed by a user into multiple time slots; The time slot interval between two points of interest is used as a time embedding element. The Sin activation function is used to encode the time slot interval between all points of interest to obtain the final time embedding information. The spatial distance between any two adjacent embedded elements between two points of interest is taken as a spatial embedded element. The spatial distance between all adjacent embedded elements is calculated using the semi-sine formula to obtain the final spatial embedding information. By using the Poisson distribution to predict the probability of a user moving from one point of interest to another, we obtain the expression for the probability prediction of the point of interest location. Based on the probability prediction expression of point of interest location, the parameters representing spatial distance information in the probability prediction expression of point of interest location are replaced with the final spatial embedding information, and the parameters representing time interval information in the probability prediction expression of point of interest location are replaced with the final temporal embedding information, thus obtaining fused spatiotemporal embedding information. The Transformer encoding module includes a two-layer location feedforward network and an equalization sampler based on an improved cross-entropy loss function; the fused spatiotemporal embedding information and user trajectory embedding information are simultaneously input into the Transformer encoding module to obtain the predicted probability of the next location interest point, including: A symmetric masking matrix is ​​constructed, embedding user trajectory information. The fused spatiotemporal embedding information and the masking matrix are input to the head self-attention formula, and operations are performed with the parameter matrix to obtain the result. The result is then linearly projected onto... Parallel applications in a subspace of time slots Each attention function generates an output representation, which is then concatenated and projected again to obtain a connection sequence. Input the connection sequence into a two-layer position feedforward network to obtain each element of the connection sequence; For each element in the connection sequence, apply residual connection, layer normalization, and dropout regularization to obtain the elements after residual connection, layer normalization, and dropout regularization. Arrange all the elements after residual connection, layer normalization, and dropout regularization in order to obtain the final output sequence. The final output sequence is converted into location interest point probabilities by using linear transformation and the Softmax function instead of the traditional dot product matching function. The probability of the point of interest is input into an equalization sampler based on the cross-entropy loss function. Unvisited locations with negative point of interest probabilities are weighted to improve the cross-entropy loss function and obtain the predicted probability of the point of interest at the next location.

2. The point-of-interest recommendation method according to claim 1, characterized in that, The acquired user set, location interest point set, and user check-in sequence set are used as initial data and input into the improved spatiotemporal fusion embedding module in batches.

3. The point-of-interest recommendation method according to claim 1, characterized in that, Methods for dividing the time interval between two locations of interest accessed by a user into multiple time slots include: Divide the time interval between two locations of interest visited by the user into In each time slot, the time value of any scalar can be projected onto... In any of the time slots.

4. The point-of-interest recommendation method according to claim 1, characterized in that, Using the time slot interval between two points of interest as a temporal embedding element, the Sin activation function is used to encode the time slot intervals between all points of interest, obtaining the final temporal embedding information, where the th A time-embedded element, represented as: ； in, Indicates the first A time-embedded element, Indicates the first A weight vector, Indicates the first Bias vectors and These represent the time slots to which the user's check-in time belongs at location i and location j, respectively; This indicates the length of the user's check-in sequence. k has no actual meaning and represents a count.

5. The point-of-interest recommendation method according to claim 1, characterized in that, The spatial distance between any two adjacent embedded elements between two points of interest is taken as a spatial embedded element. Using the semi-versus formula, the spatial distances between all adjacent embedded elements are calculated to obtain the final spatial embedding information. Here, the th... A spatially embedded element is represented as: ； in, Indicates the first A space-embedded element; Indicates the first The embedded element and the first Spatial distance between embedded elements.

6. The point-of-interest recommendation method according to claim 1, characterized in that, Using the Poisson distribution to predict the probability of a user moving from one point of interest to another, we obtain the expression for the probability prediction of point of interest locations, which is expressed as follows: ； in, Indicates user Visit two points of interest and The probability of; Indicates learnable parameters, Represents the exponential factor in the Poisson distribution; This indicates that the user visited two points of interest. and The time interval information between them, where log represents taking the logarithm.

7. The point-of-interest recommendation method according to claim 6, characterized in that, Based on the probability prediction expression for point-of-interest (POI) locations, the parameters representing spatial distance information in the expression are replaced with the final spatial embedding information, and the parameters representing time interval information are replaced with the final temporal embedding information, resulting in fused spatiotemporal embedding information. This fused spatiotemporal embedding information is expressed as follows: ； in, This represents the spatiotemporal embedded information of the fusion. This represents the final spatial embedding information. This indicates the final embedded time information. This represents the exponential factor in the Poisson distribution.

8. The point-of-interest recommendation method according to claim 1, characterized in that, The user trajectory embedding information is represented as follows: ； Where E represents user trajectory embedding information, Represents the absolute position embedding matrix. This indicates the embedded information of the user's basic trajectory.

9. The point-of-interest recommendation method according to claim 8, characterized in that, The interest point probability is input into an equalization sampler based on the cross-entropy loss function. Unvisited locations with negative interest point probabilities are weighted to improve the cross-entropy loss function, thus obtaining the predicted probability of the interest point at the next location. The improved cross-entropy loss function is expressed as: ; in, Indicates user Points of interest and points of interest All of these are initial data. , Indicates to the user Visit points of interest The probability is taken as logarithm. Indicates the length of the user's check-in sequence. Indicates that for a given user Under the movement trajectory, position The probability of being a negative sample.

Images