Point of interest recommendation method based on multivariate Hawkes spatiotemporal point process embedded attention

Abstract

The present invention discloses a method for recommending points of interest (POIs) based on multivariate Hawkes spatiotemporal point process embedded attention, comprising: a user-POI interaction sequence module, a representation learning embedding layer, multivariate Hawkes spatiotemporal point process sequence modeling, and an attention mechanism for capturing and recommending user preferences. The method first processes user-POI interaction data to obtain a user-POI interaction sequence; the interaction sequence and user information are input into the representation learning embedding layer to construct a time series and user feature representation; further, the user interaction sequence feature representation is fused through an attention mechanism to obtain the user's long-term and short-term preference features; finally, the multivariate Hawkes spatiotemporal point process model is combined with the user's long-term and short-term preference features to recommend POIs for the user; and by calculating the loss function to optimize the parameters, the model is able to dynamically and adaptively integrate the target user's long-term static preferences and short-term dynamic preferences, performing sequential recommendations to meet the target user's real-time needs.

Description

Technical Field

[0001] The present invention relates to the technical field of interest points of recommendation systems, and in particular to a method for recommending interest points based on multivariate Hawkes spatiotemporal point process embedded attention. Background Art

[0002] In recent years, the number of mobile smart devices equipped with GPS (Global Positioning System) modules has surged, and spatiotemporal location-based services have become a significant development trend. Location-based social network (LBSN) services such as Foursquare, Gowalla, and Yelp have experienced rapid growth. People check in to record points of interest (POIs). POI recommendation uses user check-in data to analyze historical interactions and further calculate preferences, helping recommendation systems more accurately make POI recommendations.

[0003] Currently, recommendation methods, including collaborative filtering (CF), content-based (CB) methods, context-aware methods and their hybrid methods, have been successfully applied in many fields, such as movie / video recommendation, and music recommendation. There have been many works on point of interest recommendation.

[0004] Existing POI recommendation methods still have shortcomings: Many LBSN services and applications only consider sequence information and ignore spatial coordinates. Some programs consider both temporal and spatial coordinates simultaneously, ignoring the temporal and spatial correlations, which can lead to the loss of critical information. Integrating temporal and spatial information into recurrent neural networks (RNNs) also suffers from strong assumptions, leading to interdependencies between adjacent check-ins in a check-in sequence. This can easily lead to erroneous dependencies and affect recommendation results. Summary of the Invention

[0005] Aiming at the shortcomings of existing point of interest recommendation, this paper proposes a point of interest recommendation method based on multivariate Hawkes spatiotemporal point process embedded attention. A multivariate Hawkes process-based method is designed to simulate the complex sequential relationship between user behavior sequences and the corresponding functions of user behavior sequences, and recommend appropriate points of interest according to the long-term / short-term preferences of target users to meet their real-time needs.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] The present invention adopts the multivariable Hawkes time-space point process. The Hawkes process is mainly a self-motivated process. The most essential feature of the Hawkes process is the point process: events that occurred in the past will definitely increase the probability of future events. The Hawkes time-space point process is an extension of the Hawkes process. When the first event does not occur, there is a μu,i The intensity function (base intensity) is the most basic probability of an event occurring.

[0008] In this paper, the similarity product of the user's feature representation and the feature representation of the recommended POI is used as the basis strength. Because different users have different access patterns and preferences for POIs, integrating user information and POI information into the Hawkes spatiotemporal point process enables personalized modeling and differentiates different user behavior patterns. Incorporating user and POI information allows the model to learn each user's unique visit patterns.

[0009] The present invention first obtains a relevant data set for implementing the task of recommending points of interest, wherein the data set contains a user set U = {u1, u2, ..., u |U|}、User visit interest point dataset I={i1,i2,...,i |I| ) and the time-space information sequence corresponding to the user's visit to the point of interest is

[0010] S u =[(u,i1,t1,s1),(u,i2,t2,s2),...,(u,i |Su|-1 , t |Su|-1 , s |Su|-1 )], (u, i, t, s)∈U×I×T×S, (u,i,t,s)∈S u , S u ∈S,

[0011] Where (u, i, t, s) means that user u visits point of interest i at time t, i is at location s, T is the timestamp set, S is the set of all points of interest, S u is the access sequence of user u. In the present invention, the spatial position s is the latitude and longitude coordinates (latitude, longitude).

[0012] Specifically, the following steps are included:

[0013] S1. Extract valid data from the original data and construct user-interest point interaction data. S u is the access sequence of user u, (u, i, t, s)∈S u , S u ∈S, where S is the access sequence of all users, u∈U, i∈I, where U is the set of all users, I is the set of all points of interest, u is the user number, i is the point of interest number, t is the time when user u visits point of interest i, and s is the location of point of interest i. u The features of user u and point of interest i are constructed by the embedding module.

[0014] S2. Construct the basic strength of the probability of user u visiting point of interest i The feature representation vector of user u and the feature representation vector of user u visiting point of interest i at time t Input the model. Perform element-wise multiplication and sum along the first dimension (batch dimension). Get v u and v u,t,i Correlation score μ u,i The formula is:

[0015] μ u,i =(v u ) T v u,t,i

[0016] S3. Construct the initial excitation degree vector of the historical sequence h in the user u behavior sequence for the current interest point i Feature representation matrix for user u's historical visit to interest point h The feature representation vector v of user u visiting point of interest i at time t u,t,i Perform element-by-element multiplication and sum the last dimension to obtain v u,h With v u，t，i The correlation score α between u，h，i The formula is:

[0017] α u，h，i =v u，h v u，t，i

[0018] S4. Calculate the attention weight of user u to historical behavior.

[0019] S41. The feature representation matrix v of the user u's historical visit to the point of interest h u，h Input linear layer (v u，h W h +b h ),in is the trainable model parameter, and then passes through the nonlinear activation function relu() to obtain the historical access sequence latent feature representation matrix The formula is:

[0020] Hvec u，h =relu(v u，h W h +b h )

[0021] S42. Represent the feature vector v of user u u and the historical access sequence latent feature representation matrix Hvec u，hMultiply by Softmax() function to obtain the user u's attention vector for different historical interest points The formula is:

[0022] att u，h =Softmax(Hvec u，h v u )

[0023] S5. Utilize and integrate user preferences through an attention-based adaptive approach. Use the attention model to determine the dynamic and static weights of users' long-term and short-term preferences.

[0024] S51. Represent the feature vector v of user u u and average historical visit interest point feature representation vector Splice together to build a long-term and short-term feature representation matrix Obtain a feature representation that has both global and local information, The formula definition is as follows:

[0025]

[0026] where v u，h is the feature representation matrix of user u’s historical visits to interest points h; is the sum of the feature representations of the historically visited points of interest h by user u at time t; |S u，t | is the length of the historical access sequence of user u at time t; S u，t is the historical access sequence of user u at time t.

[0027] S52. Construct a trainable long-term and short-term preference dynamic and static weight matrix Where W [l，s] It can be decomposed into the long-term preference static weight w l and short-term preference dynamic weight w s ,in The initialization method is used to randomly initialize the trainable long-term and short-term preference dynamic and static weight matrix W [l，s] .

[0028] S53. Representation of long-term and short-term characteristics and the trainable long-term and short-term preference weight matrix W [l，s] Perform matrix multiplication and add the trainable bias vector to the result Use the nonlinear activation function relu(), sum the last dimension, and perform the Softmax() operation on the result to normalize the result into a probability distribution. The formula is:

[0029]

[0030] in is the dynamic and static weight vector of the long-term and short-term preferences for user u. is the static weight of the long-term preference for user u. is the dynamic weight of the short-term preference for user u.

[0031] S6. Construct an exponential kernel function for user u, which calculates the exponential decay of the historical impact of past event occurrences on current event occurrences over time and space K u (t - t h , s - s h ).

[0032] Introduce user-related parameters δ u > 0, γ u > 0 representing the decay rate of the historical impact, controlling the impact of historical events on current event occurrences, that is, the impact of previous events decays to 0 over time, approaching their basic rate μ u，i , and the formula is:

[0033] K u (t - t h , s - s h ) = exp(-δ u |t - t h | - γ u |s - s h |)

[0034] where δ u controls the degree of influence of past events on the intensity of current event occurrences over time; γ u controls the degree of influence of past events on the intensity of current event occurrences over distance; K(t - t h , s - s h ) is the exponential kernel function of user u with respect to |t - t h | and |s - s h |, which calculates the exponential decay of the historical impact over time; |t - t h | is the time difference between the current time t and the historical time t h ; |s - s h | is the distance difference between the location s accessed at the current time t and the historical access location s h .

[0035] S7. Use a multivariate Hawkes spatio-temporal point process to model the user behavior sequence and predict and recommend the next point of interest based on the historical sequence. Substitute the above results into the model, and the formula is:

[0036]

[0037] where is the probability that user u prefers the point of interest i at location s at time t; μu，i is the basic strength of the probability of user u visiting point of interest i; is the static weight of user u’s long-term preference; is the short-term preference dynamic weight of user u; att u，h is the attention vector of user u to different historical points of interest; α u，h，i It is the initial excitation degree vector of the historical sequence h in the user u behavior sequence to the current interest point i.

[0038] S8. Calculate the probability that user u is interested in point of interest i′ at time t. For each point of interest i′∈I, the conditional distribution p i′|u (s, t) is on the entire set of interest points I. The formula is:

[0039]

[0040] S9. Sample a negative sample for user u at time t. The negative sample is a feature representation of the point of interest (different from the feature representation of the user's historically visited points of interest at time t). Replace the above i' feature representation with the feature representation of the negative sample point of interest k':

[0041] S10. Calculate the loss and optimize the Hawkes time-space point process parameters. In order to improve the accuracy of the recommendation, the present invention adopts the loss formula as follows:

[0042]

[0043] Where σ(x) is the Sigmoid function, n is the number of negative samples, and k′ is based on P I Interest points sampled from the interest point set, P I is the degree distribution P I ∝d u 3 / 4 , d u is the degree of user u, and samples negative samples that do not appear in the historical access sequence.

[0044] As a preferred method, step S1 extracts user data (u,i,t,s) from the original data, u∈U,i∈I, where U is the user set, I is the interest point set, u is the user number, i is the interest point number, t is the time when user u visits interest point i, and s is the location of interest point i. Group by user number and filter out the historical visit sequence of user u S u ∈S, where S is the access sequence of all users, S u is the access sequence of user u. u The features of user u and point of interest i are constructed by the embedding module.

[0045] Preferably, in step S2, the similarity between the user feature representation and the interest points visited at each moment is used as the basic intensity value of the Hawkes spatiotemporal point process.

[0046] Preferably, step S4 calculates the attention weights of the user and the historical visit sequence to capture the user's attention to different visited interest points in the historical visit sequence.

[0047] As a preference, step S6 calculates the time difference between time t and the historical access sequence |tt h | and the distance difference between the points of interest visited by the user at time t and the points of interest visited in the historical sequence |ss h |Used to capture the changes in user access sequences during the Hawkes space-time point process.

[0048] Preferably, step S7. calculates the user's preference for visiting points of interest as a Hawkes space-time point process model.

[0049] As a preference, step S10. through the degree distribution P I ∝d u 3 / 4 , where d u is the degree of user u. Sampling negative samples that do not appear in the historical visit sequence helps avoid summing the entire location sequence set and saves computing time.

[0050] The present invention has the following characteristics and beneficial effects:

[0051] Using the above technical solution, an embedding layer is added to the Hawkes spatiotemporal point process model to alleviate data sparsity. Directly using IDs for modeling will result in an overly sparse user-POI interaction matrix. Mapping to a low-dimensional dense embedding space improves model expressiveness and captures implicit features. Because IDs inherently carry limited information, mapping to the embedding space allows implicit feature representations of users and POIs to be learned through model training. Direct computation in the large-scale user-POI space is computationally complex, but mapping to a low-dimensional embedding space significantly accelerates computation. Finally, users and POIs are unified. Mapping different ID types to the same space facilitates the model's unified learning of user-POI correlations.

[0052] In addition, the similarity product of the user's feature representation and the feature representation of the recommended POI is used as the basis strength. Because different users have different access patterns and preferences for POIs, integrating user and POI information into the Hawkes spatiotemporal point process enables personalized modeling and differentiates different user behavior patterns. Incorporating user and POI information allows the model to learn each user's unique visit patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 Architecture diagram of the point of interest recommendation model with multivariate Hawkes spatiotemporal point process embedded attention. DETAILED DESCRIPTION

[0055] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0056] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0057] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0058] The present invention provides a point of interest recommendation method based on multivariate Hawkes spatiotemporal point process embedded attention, such as Figure 1As shown in the figure, the overall architecture consists of a user-POI interaction sequence module, a representation learning embedding layer, a multivariate Hawkes spatiotemporal point process sequence modeling, and an attention mechanism for capturing user preferences and recommendations. First, a low-dimensional, dense feature representation (embedding) of users and POIs is learned, which is more informative and effective than user / POI IDs or one-hot representations. Then, a multivariate Hawkes spatiotemporal point process simulates user behavior sequences, learning temporal dependencies and event correlations in a continuous time space. An attention mechanism is then employed to leverage users' long-term and short-term preferences to improve model accuracy. Finally, the next POI is recommended based on the user's historical behavior sequence.

[0059] The present invention first obtains a relevant data set for implementing the task of recommending points of interest, wherein the data set contains a user set U = {u1, u2, ..., u |U|}、User visit interest point dataset I={i1,i2,...,i |I|} and the time-space information sequence corresponding to the user's visit to the point of interest is S u =[(u,i1,t1,s1),(u,i2,t2,s2),...,(u,i |Su|-1 , t |Su|-1 , s |Su|-1) ], (u, i, t, s)∈U×I×T×S, (u,i,t,s)∈S u , S u ∈S, where (u, i, t, s) means user u visits point of interest i at time t, i is at location s, T is the timestamp set, S is the set of all points of interest, S u is the access sequence of user u. In the present invention, the spatial position s is the latitude and longitude coordinates.

[0060] Next, the implementation steps will be described in detail with reference to the accompanying drawings.

[0061] Step (1). Extract valid data (u, i, t, s) from the original data, u∈U, i∈I, where U is the user set, I is the interest point set, u is the user number, i is the interest point number, t is the time when user u visits interest point i, and s is the location of interest point i. Group by user number and filter out the historical access sequence of user u S u ∈S, where S is the access sequence of all users, S u is the access sequence of user u. u The data is fed into the embedding module to learn a low-dimensional dense feature representation vector (embedding) with rich information about users and points of interest, capturing the characteristics and relationships between users and points of interest.

[0062] Step (2). Build the basic strength of the probability of user u visiting point of interest i The feature representation vector of user u and the feature representation vector of user u visiting point of interest i at time t Input the model. Perform element-wise multiplication and sum along the first dimension (batch dimension). Get v u and v u，t，i Correlation score μ u，i The formula is:

[0063] μ u,i =(v u ) T v u,t,i

[0064] Step (3). Construct the initial excitation degree vector of the historical sequence h in the user u behavior sequence to the current interest point i Feature representation matrix for user u's historical visit to interest point h The feature representation vector v of user u visiting point of interest i at time t u,t,i Perform element-by-element multiplication and sum the last dimension to obtain v u,h With v u，t，i The correlation score α between u，h，i The formula is:

[0065] α u，h，i =v u，h v u，t，i

[0066] Step (4). Calculate the attention weight of user u to historical behavior.

[0067] Step (4.1). The feature representation matrix v of the user u's historical visit interest point h u，h Input linear layer (v u，h W h +b h ),in is the trainable model parameter, and then passes through the nonlinear activation function relu() to obtain the historical access sequence latent feature representation matrix The formula is:

[0068] Hvec u,h =relu(v u,h W h +b h )

[0069] Step (4.2). Represent the feature vector v of user u u and the historical access sequence latent feature representation matrix Hvec u，hMultiply by Softmax() function to obtain the user u's attention vector for different historical interest points The formula is:

[0070] att u，h =Softmax(Hvec u，h v u )

[0071] Step (5): Utilize and integrate user preferences through an attention-based adaptive method. Use the attention model to determine the dynamic and static weights of users' long-term and short-term preferences.

[0072] Step (5.1). Represent the feature vector v of user u u and average historical visit interest point feature representation vector Splice together to build a long-term and short-term feature representation matrix Obtain a feature representation that has both global and local information, The formula definition is as follows:

[0073]

[0074] where v u,h is the feature representation matrix of user u’s historical visits to interest points h; is the sum of the feature representations of the historically visited points of interest h by user u at time t; |S u,t | is the length of the historical access sequence of user u at time t; S u,t is the historical access sequence of user u at time t.

[0075] Step (5.2). Construct a trainable long-term and short-term preference dynamic and static weight matrix Where W [l，s] It can be decomposed into the long-term preference static weight w l and short-term preference dynamic weight w s ,in The initialization method is used to randomly initialize the trainable long-term and short-term preference dynamic and static weight matrix W [l，s] .

[0076] Step (5.3). Representation of long-term and short-term features and the trainable long-term and short-term preference weight matrix W [l，s] Perform matrix multiplication and add the trainable bias vector [b l , b s ] Use the nonlinear activation function relu(), sum the last dimension, and perform the Softmax() operation on the result to normalize the result into a probability distribution. The formula is:

[0077]

[0078] Among them is the long-term / short-term preference dynamic / static weight vector of user u is the long-term preference static weight of user u is the short-term preference dynamic weight of user u

[0079] Step (6). Construct an exponential kernel function for user u, which calculates the exponential decay of the historical impact of past event occurrences on current event occurrences over time and space K u (t - t h , s - s h ).

[0080] Step (6.1). Introduce user-related parameters δ u > 0, γ u > 0 represents the decay rate of historical impact, controlling the impact of historical events on current occurrences, that is, the impact of previous events decays to 0 over time, approaching their basic rate μ u，i , and the formula is:

[0081] K u (t - t h , s - s h ) = exp(-δ u |t - t h | - γ u |s - s h |)

[0082] Among them, δ u controls the degree of influence of past events on the intensity of current events over time; γ u controls the degree of influence of past events on the intensity of current events over distance; K(t - t h , s - s h ) is the exponential kernel function of user u with respect to |t - t h | and |s - s h |, which calculates the exponential decay of historical impact over time; |t - t h | is the time difference between the current time t and the historical time t h ; |s - s h | is the distance difference between the location s accessed at the current time t and the historical access location s h .

[0083] Step (7). Use a multivariate Hawkes spatio-temporal point process to model the user behavior sequence, and predict and recommend the next point of interest based on the historical sequence. Substitute the above results into the model, and the formula is:

[0084]

[0085] in is the probability that user u prefers point of interest i at position s at time t; μ u,i is the basic strength of the probability of user u visiting point of interest i; is the static weight of user u’s long-term preference; is the short-term preference dynamic weight of user u; att u，h is the attention vector of user u to different historical points of interest; α u，h，i It is the initial excitation degree vector of the historical sequence h in the user u behavior sequence to the current interest point i.

[0086] Step (8). Calculate the probability that user u is interested in point of interest i′ at time t. For each point of interest i′∈I, the conditional distribution p i′|u (s, t) is on the entire set of interest points I. The formula is:

[0087]

[0088] Step (9). Sample a negative sample for user u at time t. The negative sample is the feature representation of the interest point (different from the feature representation of the user's historical visit interest point at time t). Replace the above i' feature representation with the feature representation of the negative sample interest point k'

[0089] Step (10). Calculate the loss and optimize the Hawkes time-space point process parameters. In order to improve the accuracy of recommendation, the present invention adopts the loss formula as follows:

[0090]

[0091] Where σ(x) is the Sigmoid function, n is the number of negative samples, and k′ is based on P I Interest points sampled from the interest point set, P I is the degree distribution P I ∝d u 3 / 4 , d u is the degree of user u, and samples negative samples that do not appear in the historical access sequence.

[0092] This method was experimented on the Gowalla dataset. The Gowalla dataset was processed according to the following requirements, extracting data to generate five subsets: the top 2,000 users who frequently visited points of interest and the top 10,000 popular points of interest; the top 2,000 users who frequently visited points of interest and the top 15,000 popular points of interest; the top 2,000 users who frequently visited points of interest and the top 20,000 popular points of interest; the top 2,000 users who frequently visited points of interest and the top 25,000 popular points of interest; and the top 2,000 users who frequently visited points of interest and the top 30,000 popular points of interest. Table 1 shows that compared to current mainstream recommendation methods, this method achieves significant improvements in recall at R@1, R@5, R@10, and R@20. This demonstrates that this method can achieve high-quality and efficient recommendation tasks. "top_n_item" in the table represents the number of extracted points of interest. Recall is an important evaluation metric in recommendation systems. MHSTPP-a is the name of this method.

[0093] Table 1 Recall experiment results on the Gowalla dataset

[0094]

[0095] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A point of interest recommendation method based on multivariate Hawkes spatiotemporal point process embedding attention, characterized by: S1. Extract valid data from the original data and construct user-POI interaction data; S2. Based on the user-POI interaction data, construct the basic strength of the probability of the user visiting the POI event through element-wise multiplication operation; and construct the initial excitation degree vector of the historical sequence in the user behavior sequence for the current POI through element-by-element multiplication; S3. Calculate the user's attention weight on historical behavior S3-1. The feature representation matrix v of the user u's historical visit to the point of interest h u，h Input linear layer (v u,h W h +b h ),in is the trainable model parameter, and then passes through the nonlinear activation function relu() to obtain the historical access sequence latent feature representation matrix The formula is: Hvec u,h =relu(v u，h W h +b h ); S3-2. Represent the feature vector v of user u u and the historical access sequence latent feature representation matrix Hvec u,h Multiply by Softmax() function to obtain the user u's attention vector for different historical interest points The formula is: to u，h =Softmaxx((Hvec u，h v u ); S4. Utilize and integrate user preferences based on attention-based adaptive methods to determine the dynamic and static weights of users' long-term and short-term preferences, the static weights of long-term preferences, and the dynamic weights of short-term preferences; S5. Construct an exponential kernel function for the user, which calculates the exponential decay of the historical impact of past event occurrences on the current event occurrence over time and space, K u (t - t h , s - s h ), where K(t - t h , s - s h ) is the exponential kernel function of user u with respect to |t - t h | and |s - s h |, which calculates the exponential decay of the historical impact over time; |t - t h | is the time difference between the current time t and the historical time t h ; |s - s h | is the distance difference between the location s accessed at the current time t and the historical access location s h . S6. Use the multivariate Hawkes spatiotemporal point process to model the user behavior sequence. The constructed model expression is as follows: Predict and recommend the next point of interest based on historical sequences, and bring the results into the model. in is the probability that user u prefers point of interest i at position s at time t; μ u，i is the basic strength of the probability of user u visiting point of interest i; is the static weight of user u’s long-term preference; is the short-term preference dynamic weight of user u; att u，h is the attention vector of user u to different historical points of interest; α u，h，i is the initial excitation degree vector of the historical sequence h in the user u behavior sequence to the current interest point i; S u，t is the historical access sequence of user u at time t; S7. Calculate the probability that user u is interested in point of interest i′ at time t. For each point of interest i′∈I, the conditional distribution p i′|u (s, t) on the entire interest point set I, the formula is: S8. Sample a negative sample for user u at time t. The negative sample is the feature representation of the point of interest. Replace the above i′ feature representation with the feature representation of the negative sample point of interest k′: S9. Calculate the loss and optimize the Hawkes space-time point process parameters.

2. The method for recommending points of interest based on multivariate Hawkes spatiotemporal point process embedded attention according to claim 1, characterized in that: The specific method of step S1 is: make S u is the access sequence of user u, (u, i, t, s)∈S u , S u ∈S, where S is the access sequence of all users, u∈U, i∈I, where U is the set of all users, I is the set of all points of interest, u is the user number, i is the point of interest number, t is the time when user u visits point of interest i, s is the location of point of interest i, and the user access sequence S is converted to u The interaction sequence is fed into the embedding module to construct feature representations of user u and interest point i respectively.

3. The method for recommending points of interest based on multivariate Hawkes spatiotemporal point process embedded attention according to claim 2, characterized in that: Method for obtaining the basic strength in step S2: Extracting the feature representation vector of user u from user-POI interaction data and the feature representation vector of user u visiting point of interest i at time t And perform element-wise multiplication and sum along the first batch dimension to obtain v u and v u，t，i Correlation score μ u，i , that is, the basic strength of the probability of user u visiting point of interest i 4. The method for recommending points of interest based on multivariate Hawkes spatiotemporal point process embedded attention according to claim 2, characterized in that: The method for obtaining the initial excitation degree vector in step S2 is: Extract the feature representation matrix of user u's historical visited interest points h from user-interest point interaction data The feature representation vector v of user u visiting point of interest i at time t u，t，i , and perform element-by-element multiplication and sum the feature dimensions to obtain v u，h With v u，t，i The correlation score α between u，h，i , that is, the initial excitation degree vector 5. The method for recommending points of interest based on multivariate Hawkes spatiotemporal point process embedded attention according to claim 4, characterized in that: Step S4 includes the following sub-steps: S4-1 represents the feature vector v of user u u and average historical visit interest point feature representation vector Splice together to build a long-term and short-term feature representation matrix Obtain a feature representation that has both global and local information, The formula definition is as follows: where v u，h is the feature representation matrix of user u’s historical visits to interest points h; is the sum of the feature representations of the historically visited points of interest h by user u at time t; |S u，t | is the length of the historical access sequence of user u at time t; S4-2 Constructing a trainable long-term and short-term preference dynamic and static weight matrix Where W [l，s] It can be decomposed into the long-term preference static weight w l and short-term preference dynamic weight w s ,in The initialization method is used to randomly initialize the trainable long-term and short-term preference dynamic and static weight matrix W [l，s] ; S4-3 Representation of long-term and short-term characteristics and the trainable long-term and short-term preference weight matrix W [l，s] Perform matrix multiplication and add the trainable bias vector to the result Use the nonlinear activation function relu(), sum the last dimension, and perform the Softmax() operation on the result to normalize the result into a probability distribution. The formula is: in is the dynamic and static weight vector of user u’s long-term and short-term preferences, is the static weight of user u’s long-term preference, is the dynamic weight of user u’s short-term preference.

6. The method for recommending points of interest based on multivariate Hawkes spatiotemporal point process embedded attention according to claim 5, characterized in that: The specific method of step S5 is: Introducing user-related parameters δ u >0,γ u >0 represents the decay rate of historical influence, which controls the impact of historical events on current events, that is, the impact of previous events decays towards 0 over time, approaching their basic rate μ u，i , the formula is: K u (t-t h ,s-s h )=exp(-δ u |t-t h |-γ u |s-s h |) where δ u Controls the influence of past events on the intensity of current events over time; γ u Controls how far away a past event is from the present.

7. The method for recommending points of interest based on multivariate Hawkes spatiotemporal point process embedded attention according to claim 6, characterized in that: In step S9, the loss is used for optimization, and the expression is as follows: Where σ(x) is the Sigmoid function, n is the number of negative samples, and k′ is based on P I Interest points sampled from the interest point set, P I is the degree distribution P I ∝d u 3 / 4 , d u is the degree of user u, and samples negative samples that do not appear in the historical access sequence.

Images