A spatiotemporal multi-scale point of interest recommendation method

Abstract

The present application relates to the technical field of data mining and recommendation system, and particularly relates to a spatiotemporal multi-scale interest point recommendation method; the present application converts learning of user behavior into learning of potential state, and introduces spatial information in a manner of combining distance information, effectively captures the moving characteristics of the user, and the spatiotemporal multi-scale recommendation mode increases the diversity of interest point recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining and recommendation system, and more particularly, to a spatiotemporal multi-scale point-of-interest recommendation method. BACKGROUND

[0002] In recent years, point-of-interest (POI) recommendation has developed rapidly. By mining user check-in data and modeling user behavior patterns, it provides precise services and personalized recommendations for users, increasing their stickiness to the platform. POI recommendation technology is a technology that matches POIs that meet user preferences or habits through big data intelligence and recommends POIs to demanders. Existing POI recommendation technology mainly recommends POIs to users by analyzing user-generated content (UGC) or recommends POIs to users based on user information and POI information. The technology of recommending POIs by analyzing UGC has the problems of high data operation cost and inaccurate POI recommendation caused by the subjectivity of UGC. The technology of recommending POIs based on user information and POI information has the problem of inaccurate POI recommendation caused by sparse user information.

[0003] Traditional POI recommendation methods based on collaborative filtering usually regard user behavior preferences as fixed, but user behavior preferences often change over time, and check-in behavior often shows periodicity. POI recommendation based on spatiotemporal information is the mainstream trend of POI recommendation, but it still faces challenges such as sparse check-in data, difficulty in extracting spatiotemporal sequence features, and difficulty in capturing user individual differences. In existing technology, POI recommendation is mainly carried out through the following three schemes:

[0004] POI recommendation based on spatial information: This method uses spatial information such as user location and POI location to establish spatial relationships between users and POIs, thereby improving the accuracy and diversity of recommendations. This scheme usually regards user behavior preferences as fixed, but user behavior preferences often change over time and mobile behavior has dependence and periodicity.

[0005] POI recommendation based on temporal information: This method uses temporal information such as user check-in time and POI visit period to establish temporal relationships between users and POIs, thereby improving the accuracy and diversity of recommendations. This scheme can mine the periodicity of user mobile behavior to establish temporal relationships between users and POIs, but it ignores important geographic location information. Users often tend to check-in at nearby POIs and there is a kind of spatial clustering phenomenon.

[0006] Point-of-interest recommendation based on spatio-temporal information: This kind of method is to use the check-in time and location of the user, and the spatio-temporal information such as the access period and geographical distribution of the point of interest, to establish the spatio-temporal relationship between the user and the point of interest, so as to improve the accuracy and interpretability of the recommendation. This kind of scheme is the mainstream trend of point-of-interest recommendation at present, but still faces the challenges of sparse check-in data, difficulty in extracting spatio-temporal sequence features, and difficulty in capturing user individual differences. SUMMARY

[0007] The present application is to overcome the above-mentioned problems of the prior art, such as the fact that the user's behavior preferences often change over time, the fact that the mobile behavior has dependency and periodicity, and the fact that the important information of geographical location is ignored. The present application provides a spatio-temporal multi-scale point-of-interest recommendation method and device.

[0008] To solve the above technical problems, the technical scheme of the present application is as follows:

[0009] A spatio-temporal multi-scale point-of-interest recommendation method, comprising the following steps:

[0010] S1: Preprocessing the user's point-of-interest check-in sequence of each period to obtain the potential state vector and distance bias vector of each period in the next period;

[0011] S2: Inputting the potential state vector and distance bias vector of each period in the next period into a decoding multi-layer perception machine to generate the point of interest that meets the user's preferences in each period of the next period;

[0012] S3: According to the time range and space range of the user, further filtering the point of interest that meets the user's preferences obtained in step S2, and returning the filtered point of interest to the user.

[0013] Further, in step S1, the preprocessing of the user's point-of-interest check-in sequence of each period to obtain the potential state vector and distance bias vector of each period in the next period is specifically:

[0014] S1.1: Acquiring the check-in sequence sent by the user, and encoding the user's check-in sequence into a Multi-hot vector;

[0015] S1.2: Calculating the Multi-hot vector of each period in the current period of the user to obtain the distance bias vector of each period in the next period;

[0016] S1.3: Calculating the Multi-hot vector of each period in the current period of the user to obtain the potential state vector of each period in the current period of the user;

[0017] S1.4: Calculate the potential state vector of each period in the current cycle to obtain the state transition amount of each period in the current cycle;

[0018] S1.5: Add the potential state vector of each period in the current cycle to the corresponding state transition amount to obtain the potential state vector of each period in the next cycle.

[0019] Further, in step S1.2, the Multi-hot vector of each period in the current cycle of the user is calculated to obtain the distance bias vector of each period in the next cycle:

[0020] Step S1.21: Calculate the radial basis function value between each pair of interest points in advance using the radial basis function kernel to obtain the radial basis function matrix between the interest points:

[0021] K(poi A ,poi B )=exp(-γ||poi A -poi B || 2 )

[0022] Where poi A represents the latitude and longitude of interest point A, poi B represents the latitude and longitude of interest point B, γ represents the hyperparameter of the interest point relevance level, and K represents the radial basis function matrix.

[0023] Step S1.22: Calculate the importance score of the check-in interest point of each period in the previous cycle using the attention mechanism;

[0024] Step S1.23: Calculate the distance bias vector of each period in the next cycle by the radial basis function matrix and the importance score:

[0025]

[0026] Where W (4) is the POI embedding matrix, is the set of POIs checked in by the user in the jth period of the t-1 cycle, is the embedding matrix of the user in the jth period of the t-1 cycle, w α is a parameter.

[0027] Further, the importance score of the check-in interest point of each period in the previous cycle is calculated, which is:

[0028]

[0029] is the importance score.

[0030] Further, in step S1.4, the potential state vector of each period of the current period is calculated to obtain the state transition amount of each period of the current period, specifically:

[0031] S1.41: modeling the behavior state of the user, obtaining each state node in the model, each state node representing a potential state vector of the user in each period, a user long-term preference vector;

[0032] S1.42: using a fully connected layer to obtain the dependence influence amount of the jth period of the previous period;

[0033] S1.43: inputting the potential state vector of each period, the user long-term preference vector, and the dependence influence amount obtained in step S1.42 into a fully connected network to obtain the state transition amount of each period in the period.

[0034] Further, in step S1.42, a fully connected layer is used to obtain the dependence influence amount of the jth period of the previous period, specifically:

[0035]

[0036] f inf is a fully connected layer, is the dependence influence amount of the jth period of the previous period, is the potential state vector of the ith period of the t-1 period, is the potential state vector of the jth period of the t-1 period.

[0037] Further, in step S1.43, the potential state vector of each period, the user long-term preference vector, and the dependence influence amount obtained in step S3.2 are input into a fully connected network to obtain the state transition amount of each period in the period

[0038]

[0039] where σ is an activation function, and is a learnable parameter, f shift is a nonlinear fully connected network, and the state transition amount is the dependence influence amount of the jth period of the previous period, is the potential state vector of the jth period of the t-1 period, e u is a user long-term preference vector.

[0040] Further, in step S1.5, the potential state vector of each period is added to the corresponding state transition amount to obtain the potential state vector of each period of the next period, wherein:

[0041]

[0042] Further, in step S2, the potential state vector and the distance bias of each time period of the next cycle are input into the decoding multi-layer perception to obtain the predicted value of each time period of the next cycle, and the interest point of each time period of the next cycle is generated according to the predicted value, wherein:

[0043] S2.1: Calculate the distance bias vector of each time period of the next cycle obtained in step S1.2 and the potential state vector of each time period of the next cycle obtained in step S1.5 to obtain the predicted value of each time period of the next cycle;

[0044] S2.2: Sort the probabilities of the predicted value obtained in step S2.1 in descending order to generate the interest point of each time period of the next cycle.

[0045] Further, in step S2.1, the distance bias vector of each time period of the next cycle obtained in step S2 and the potential state vector of each time period of the next cycle obtained in step S1.5 are calculated to obtain the predicted value of each time period of the next cycle:

[0046]

[0047] wherein f θ (·) is a decoding multi-layer perception, is a distance bias vector, is a potential state vector of the jth time period of the tth cycle.

[0048] Compared with the prior art, the beneficial effects of the technical scheme of the present application are:

[0049] In order to solve the above problems, a spatio-temporal multi-scale interest point recommendation method is proposed, which converts the learning of user behavior into the learning of potential state, and introduces spatial information in a way combining distance information, effectively capturing the moving characteristics of users and improving the accuracy of interest point recommendation. The recommendation method aims to predict the check-in interest point of the user in the next cycle by giving the user-interest point interaction matrix of the user in the previous cycle. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of the spatio-temporal multi-scale interest point recommendation method of the present application. DETAILED DESCRIPTION

[0051] The drawings are only used for illustrative description and cannot be understood as a limitation of the patent;

[0052] In order to better illustrate the present embodiment, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0053] It is understood by those skilled in the art that certain well-known structures and their descriptions in the drawings can be omitted.

[0054] The technical solutions of the present application will be further described below in combination with the drawings and examples.

[0055] Example 1

[0056] This embodiment provides a spatiotemporal multi-scale point of interest recommendation method, as shown in the figure, comprising the following steps: Figure 1

[0057] S1: Preprocessing the user's point of interest check-in sequence in each period to obtain the potential state vector and distance bias vector of each period in the next cycle;

[0058] S2: Inputting the potential state vector and distance bias vector of each period in the next cycle into the decoding multilayer perceptron to generate the point of interest in each period in the next cycle that meets the user's preference;

[0059] S3: According to the user's time range and space range, further screening the point of interest that meets the user's preference obtained in step S2, and returning the screened point of interest to the user.

[0060] In the specific implementation process, the user's point of interest check-in sequence in each period of the previous week is input; the user's check-in sequence is encoded into a Multi-hot vector recognizable by the model, and for each user, if the user has check-in behavior in the current cycle and the next cycle, the set of check-in records is retained, and the set of check-in records is encoded, taking the previous week as an example, the encoded check-in matrix of the current cycle is defined as X t-1 , where τ represents the number of periods in a cycle; the encoded check-in matrix of the next cycle is defined as X t , The Multi-hot vector of each period of the user is input into the neighborhood perception bias calculator to obtain the distance bias of each period in the next cycle; the Multi-hot vector of each period of the user is input into the time series state encoder to obtain the potential state vector of each period of the user in the cycle; the potential state vector of each period is input into the state dependent network to obtain the state transition amount of each period; the potential state vector of each period is added to the corresponding state transition amount to obtain the potential state vector of each period in the next cycle; the potential state vector of each period in the next cycle and the distance bias are input into the decoding multilayer perceptron to generate the point of interest in each period in the next cycle that meets the user's preference; according to the user's selected time range and space range, the point of interest is further divided and screened, and the screened point of interest to be recommended is sent to the user end. ​

[0061] The application converts the learning problem of user time preference into the dynamic change and dependency learning problem of time state in terms of time, and uses similar technology to model the influence of the signed POI on the candidate POI in terms of space, and combines the Gaussian radial basis function kernel and the attention mechanism to introduce the spatial information in a distance information combined manner.

[0062] In step S1, the interest point check-in sequence of each period of the user is preprocessed to obtain the latent state vector and the distance bias vector of each period of the next period, specifically:

[0063] S1.1: Obtain the check-in sequence sent by the user, and encode the check-in sequence of the user into a Multi-hot vector;

[0064] S1.2: Calculate the Multi-hot vector of each period of the current period of the user to obtain the distance bias vector of each period of the next period;

[0065] S1.3: Calculate the Multi-hot vector of each period of the current period of the user to obtain the latent state vector of each period of the current period of the user;

[0066] S1.4: Calculate the latent state vector of each period of the current period to obtain the state transition amount of each period of the current period;

[0067] S1.5: Add the latent state vector of each period of the current period and the corresponding state transition amount to obtain the latent state vector of each period of the next period.

[0068] In step S1.2, the Multi-hot vector of each period of the current period of the user is calculated to obtain the distance bias vector of each period of the next period:

[0069] Step S1.21: Use the radial basis function kernel to calculate the radial basis function value between each pair of interest points in advance to obtain the radial basis function matrix between the interest points:

[0070] K(poi A ,poi B )=exp(-γ||poi A -poi B || 2 )

[0071] Wherein, poi A represents the latitude and longitude of the interest point A, and poi Bγ represents the latitude and longitude of point of interest B, γ represents the hyperparameter of the correlation level of point of interest, and K represents the radial basis function matrix;

[0072] Step S1.22: Use the attention mechanism to calculate the importance score of the check-in interest points for each time period in the previous cycle;

[0073] Step S1.23: Calculate the distance bias vector for each time period of the next cycle using the radial basis function matrix and importance score:

[0074]

[0075] Among them, W (1) For the POI embedding matrix, Let be the set of POIs that the user checked in during the j-th time period of period t-1. For the user's embedding matrix in the j-th time period of period t-1, w α For parameters.

[0076] The importance score of check-in interest points for each time period in the previous cycle is calculated as follows:

[0077]

[0078] The importance score is calculated based on the score.

[0079] In step S1.4, the potential state vector for each time period of the current cycle is calculated to obtain the state transition amount for each time period of the current cycle, specifically:

[0080] S1.41: In order to model the complex dependencies of user behavior in different time periods, it is assumed that the user's behavior state can be represented by a graph with τ state nodes. Each node represents the potential state of the user in each time period, and there are dependency edges and time offsets between nodes.

[0081] S1.42: To aggregate the neighbor dependencies of each state node, a fully connected layer f is used. inf To learn the dependency impact of each neighboring edge, we obtain the dependency impact of the j-th time period in the previous cycle.

[0082] S1.43: The potential states at each time period User long-term preference vector e u Dependency impact obtained The inputs are fed into the nonlinear fully connected network f. shift In this process, the state transition amount for each time period within the cycle is obtained.

[0083] In step S1.42, a fully connected layer is used to obtain the dependence influence quantity of the jth period of the previous cycle, specifically:

[0084]

[0085] f inf is a fully connected layer, is the dependence influence quantity of the jth period of the previous cycle, is the latent state vector of the ith period of the t-1 cycle, is the latent state vector of the jth period of the t-1 cycle.

[0086] In step S1.43, the latent state vector of each period, the user long-term preference vector, and the dependence influence quantity obtained in step S3.2 are input into a fully connected network to obtain the state transition quantity of each period in the cycle

[0087]

[0088] where σ is an activation function, and are learnable parameters, f shift is a nonlinear fully connected network, and the state transition quantity is the dependence influence quantity of the jth period of the previous cycle, is the latent state vector of the jth period of the t-1 cycle, e u is the user long-term preference vector, and a parameter is added to facilitate the method to learn the period dependence of the user's check-in in order to better capture the pattern of user state changes and long-term preferences.

[0089] In step S1.5, the latent state vector of each period is added to the corresponding state transition quantity to obtain the latent state vector of each period of the next cycle, where:

[0090]

[0091] In step S2, the latent state vector of each period of the next cycle and the distance bias vector are input into a decoding multilayer perceptron to generate the interest points of each period of the next cycle that meet the user's preferences, where:

[0092] S2.1: Calculate the distance bias vector of each period of the next cycle obtained in step S1.2 and the latent state vector of each period of the next cycle obtained in step S1.5 to obtain the predicted value of each period of the next cycle;

[0093] S2.2: Sort the probabilities of the predicted values obtained in step S2.1 in descending order to generate the interest points of each period of the next cycle.

[0094] In step S2.1, the distance bias vector of each time period of the next cycle obtained in step S2 and the potential state vector of each time period of the next cycle obtained in step S1.5 are calculated to obtain a predicted value of each time period of the next cycle:

[0095]

[0096] where f θ is a decoding multi-layer perception, is a distance bias vector, is a potential state vector of the jth time period of the tth cycle.

[0097] The same or similar reference numerals correspond to the same or similar components;

[0098] The terms describing the positional relationship in the drawings are only used for illustrative description, and should not be understood as a limitation on the patent;

[0099] Obviously, the above embodiments of the present application are merely exemplary for clearly illustrating the present application, and are not intended to limit the implementation manner of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and also impossible to enumerate all the implementation manners. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1.A spatio-temporal multi-scale point of interest recommendation method, characterized in that, The method comprises the following steps: S1: preprocessing the user's interest point check-in sequence of each period to obtain a potential state vector and a distance bias vector of each period in the next cycle; The step S1 comprises the following steps: S1.1: obtaining the check-in sequence sent by the user, and encoding the check-in sequence of the user into a Multi-hot vector; S1.2: calculating the Multi-hot vector of each period in the current cycle of the user to obtain a distance bias vector of each period in the next cycle; The step S1.2 comprises the following steps: Step S1.21: using a radial basis function kernel to calculate the radial basis function value between each pair of interest points in advance to obtain a radial basis function matrix between the interest points: K(poi A , poi B ) = exp(-γ||poi A -poi B || 2 ) wherein, poi A the longitude and latitude of the point of interest A, poi B the longitude and latitude of the point of interest B, γ represents a hyperparameter of the point of interest relevance level, and K represents a radial basis function matrix; Step S1.22: using an attention mechanism to calculate the importance score of the check-in interest point of each period in the previous cycle; Step S1.23: calculating the distance bias vector of each period in the next cycle through the radial basis function matrix and the importance score: wherein W (4) is a POI embedding matrix, is a set of POIs that the user checked-in at the jth time period of the t-1 period, is an embedding matrix of the user at the jth time period of the t-1 period, w α is a parameter is a distance bias vector, is a radial basis function matrix of the user at the jth time period of the t-1 period; S1.3: calculating the Multi-hot vector of each period in the current cycle of the user to obtain a potential state vector of each period in the current cycle of the user; S1.4: calculating the potential state vector of each period in the current cycle to obtain a state transition amount of each period in the current cycle; S1.5: adding the potential state vector of each period in the current cycle to the corresponding state transition amount to obtain a potential state vector of each period in the next cycle; S2: inputting the potential state vector and the distance bias vector of each period in the next cycle into a decoding multi-layer perception to generate an interest point of each period in the next cycle that meets the user's preference; The step S2 comprises the following steps: S2.1: Calculate the distance bias vector for each time period of the next cycle from the distance bias vector for each time period of the current cycle obtained in step S1.2 and the potential state vector for each time period of the next cycle obtained in step S1.5 to obtain a prediction for each time period of the next cycle; S2.2: sorting the probabilities of the predicted values obtained in step S2.1 in descending order to generate an interest point of each period in the next cycle; S3: according to the time range and the space range of the user, further screening the interest point that meets the user's preference obtained in step S2, and returning the screened interest point to the user. 2.The method of claim 1, wherein, The importance score of the check-in interest point of each period in the previous cycle is calculated, and specifically: Importance score. 3.The method of claim 1, wherein, In step S1.4, the potential state vector of each period in the current cycle is calculated to obtain a state transition amount of each period in the current cycle, and specifically: S1.41: modeling the behavior state of the user, obtaining each state node in the model, and each state node representing a potential state vector of the user in each period and a long-term preference vector of the user; S1.42: using a fully connected layer to obtain a dependence influence amount of the jth period in the previous cycle; S1.43: inputting the potential state vector of each period, the long-term preference vector of the user, and the dependence influence amount obtained in step S1.42 into a fully connected network to obtain a state transition amount of each period in the cycle. 4.The method of claim 3, wherein, In step S1.42, a fully connected layer is used to obtain a dependence influence amount of the jth period in the previous cycle, and specifically: f inf is a fully connected layer, is a dependent influence quantity of the jth period of the previous cycle, is a latent state vector of the ith period of the t-1 cycle, is a latent state vector of the jth period of the t-1 cycle. 5.The method of claim 4, wherein, In step S1.43, the potential state vector of each period, the long-term preference vector of the user, and the dependence influence amount obtained in step S3.2 are input into a fully connected network to obtain a state transition amount of each period in the cycle: where σ is an activation function, and are learnable parameters, f shift is a nonlinear fully connected network, is a state transition amount, is a dependent influence amount of the jth period of the previous cycle, is a latent state vector of the jth period of the t-1 cycle, e u is a user long-term preference vector, is a state transition amount of each period within a cycle. 6.The method of claim 1, wherein, In step S1.5, the potential state vector of each time period is added with the corresponding state transition amount to obtain the potential state vector of each time period in the next cycle, wherein: 7.The method of claim 1, wherein, In step S2.1, the distance bias vector of each time period in the next cycle obtained in step S2 and the potential state vector of each time period in the next cycle obtained in step S1.5 are calculated to obtain the predicted value of each time period in the next cycle: where f θ (·) is a decoded multi-layer perceptron, is a distance bias vector, is a latent state vector for the jth time bin of the tth period.

Images