Method and device for recommending next interest point based on fusion of interactive features and associated features

Abstract

The invention provides a next interest point recommendation method and device based on interaction feature and association feature fusion, and the method comprises the steps: obtaining the similarity between users through employing a matrix decomposition model, and obtaining a user feature vector; constructing directed graphs of all users by using the user track information, and inputting the directed graphs into the graph attention network model to obtain point-of-interest feature vectors; fusing the user feature vector and the interest point feature vector into a track interaction feature vector; obtaining an interest point type and time information in the user track information, encoding the interest point type and the time information, converting an encoding result into vector dimension information, inputting the vector dimension information into the time sequence model, and obtaining an associated feature vector of the interest point type changing along with a time period; and training a deep learning network model by using the trajectory interaction feature vector and the association feature vector to obtain a next interest point recommendation model. According to the method, collaborative behaviors and time characteristics among users are effectively considered, and the accuracy of recommendation results is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of feature fusion technology, and in particular to a next point of interest recommendation method and device by fusing interactive features with associated features. Background Art

[0002] Point of Interest (POI) recommendation is one of the important tasks in location-based services, which aims to find user preferences through users' historical visit information and make personalized recommendations for the next point of interest to users.

[0003] In the related technologies, currently, recommendations based on the next point of interest are mainly made based on location information, social network and other service information. Location information can only reflect the user's current location or places frequently visited, and social networks such as other users' check-ins can only provide some references. Users are independent of each other. Existing technical means pay less attention to the collaborative information between users, that is, the correlation information generated by the interactive behavior between users, resulting in inaccurate recommendation results.

[0004] Based on the above analysis of the development status of this technical field, the existing technology lacks a solution that takes into account the similarity between users, the interaction information of interest points between users and other text information as comprehensive features to build an interest point recommendation model. Summary of the invention

[0005] The purpose of the present invention is to provide a method and device for recommending the next point of interest by fusing interactive features with associated features, aiming to solve the above-mentioned problems in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, a method for recommending a next point of interest by fusing an interactive feature with a correlation feature is provided, comprising:

[0007] Use the matrix decomposition model to obtain the similarity between users and obtain the user feature vector; use the user trajectory information to build a directed graph of all users, input the directed graph into the graph attention network model, and obtain the point of interest feature vector;

[0008] The user feature vector and the interest point feature vector are merged into the trajectory interaction feature vector;

[0009] Obtain the type and time information of the points of interest in the user trajectory information, encode the type and time information of the points of interest, convert the encoding result into vector dimension information and input it into the time series model to obtain the associated feature vector of the type of the points of interest changing with the time period;

[0010] The trajectory interaction feature vector and the correlation feature vector are used to train a deep learning network that couples importance scoring and sparse attention mechanism to obtain the next point of interest recommendation model.

[0011] According to a second aspect of an embodiment of the present invention, there is provided a device for recommending a next point of interest by fusing an interactive feature with a correlation feature, comprising:

[0012] The initial modeling module is used to obtain the similarity between users using the matrix decomposition model to obtain the user feature vector; the user trajectory information is used to build a directed graph of all users, and the directed graph is input into the graph attention network model to obtain the point of interest feature vector;

[0013] An interactive fusion module, used to fuse the user feature vector and the interest point feature vector into a trajectory interactive feature vector;

[0014] The auxiliary association module is used to obtain the type and time information of the interest points in the user trajectory information, encode the type and time information of the interest points, convert the encoding result into vector dimension information and input it into the time series model to obtain the associated feature vector of the interest point type changing with the time period;

[0015] The model training module is used to train a deep learning network that couples importance scoring and sparse attention mechanism using trajectory interaction feature vectors and associated feature vectors to obtain the next point of interest recommendation model.

[0016] According to a third aspect of an embodiment of the present invention, there is provided an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the next point of interest recommendation method by fusing interactive features with associated features as provided in the first aspect of the present disclosure are implemented.

[0017] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps of the next point of interest recommendation method for fusing interactive features and associated features provided in the first aspect of the present disclosure are implemented.

[0018] The technical solution provided by the embodiment of the present invention includes the following beneficial effects: obtaining the similarity between users through a matrix decomposition model, integrating the user feature vector and the point of interest embedding interaction into trajectory interaction features; and considering auxiliary related information such as point of interest type and check-in time as auxiliary related features; using the above two types of features to train a deep learning network model, effectively considering the collaborative behavior between users, and greatly improving the accuracy of the recommendation results.

[0019] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] Figure 1 is a flow chart of a method for recommending a next point of interest by fusing interactive features with associated features according to an embodiment of the present invention;

[0022] Figure 2 is a schematic diagram of interest point trajectory information according to an embodiment of the present invention;

[0023] Figure 3 is a schematic diagram of a device for recommending a next point of interest by fusing interactive features with associated features according to an embodiment of the present invention;

[0024] Figure 4 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0026] Method Embodiment

[0027] According to an embodiment of the present invention, a method for recommending a next point of interest by fusing interactive features with associated features is provided. Figure 1 is a flowchart of a method for recommending a point of interest by fusing interactive features with associated features according to an embodiment of the present invention. Figure 1 As shown, the next point of interest recommendation method based on the fusion of interactive features and associated features according to an embodiment of the present invention specifically includes:

[0028] In step S110, the matrix decomposition model is used to obtain the similarity between users and obtain the user feature vector; the user trajectory information is used to construct a directed graph of all users, and the directed graph is input into the graph attention network model to obtain the point of interest feature vector, which specifically includes:

[0029] Users are an important part of POI recommendation. To consider the collaboration between users, we need to consider the similarity between users, which is implemented based on the matrix decomposition model. Construct the user's 0 / 1 check-in matrix U m×n , which is a matrix including m users and n points of interest; the check-in matrix is ​​decomposed using the bias term, which is expressed as Among them, b i represents the bias term of the i-th interest point, b u represents the bias term of user u, p u represents the feature vector of user u, The feature vector representing the i-th interest point is transposed;

[0030] Use Formula 1 to establish the objective function that minimizes the difference between the bias evaluation value and the actual check-in:

[0031]

[0032] Among them, w ui represents the weight of the access frequency, γ ui represents the actual check-in status of user u at point of interest i (0 or 1), is the check-in probability or score of user u at point of interest i predicted by the model, λ 1 and λ 2 represents the regularization parameter, and Represented as user feature vector P u and the interest point feature vector q i The Frobenius norm of can prevent overfitting, b u and b i is the bias term for users and points of interest, further improving the generalization ability;

[0033] The objective function is solved using an adaptive gradient algorithm. Preferably, in the adaptive gradient algorithm, adjustments are made based on the historical gradient of each parameter, and the historical gradient square sum G of each parameter is initialized. 0 = 0, after t iterations, the gradient is calculated as Update the historical sum of squared gradients If θ is used to represent all model parameters, the parameter update after the tth iteration is The parameters that need to be updated include p u ,q i 、b u and b i , where η represents the learning rate and ε represents a small constant used to avoid division by zero errors, which is set to 10 in the embodiment of the present invention. -8 ;

[0034] Use formula 2 to express the weight wui :

[0035]

[0036] Among them, γ(F u,i ) represents a relative access frequency F u,i A monotonically increasing function, which is set to 0.1 in the embodiment of the present invention;

[0037] Finally, we get the user feature vector p u , the preliminary solution for user u is In order to be able to interact and merge with the subsequent interest point embedding, the solution result is linearly transformed Among them, o represents the target dimension, so that it can be mapped to the dimension fused with the interest point embedding, and the user feature vector e of each user is obtained. u =Wp u ∈R Ω , where Ω=o×m.

[0038] Assume that the user set is U = {u 1 ,u 2 ,…u M}, the set of interest points is P = {p 1 ,p 2 ,…,p N}, the timestamp is T = {t 1 ,t 2 ,…,t I}, where M and N represent intervals, and I represents a short time interval. In the embodiment of the present invention, 24 hours are divided into 48 intervals;

[0039] Define a check-in tuple q including user information, POI information and timestamp =<u,p,t> ∈U×P×T, which means that user u visits the point of interest p at timestamp t, where the point of interest information includes longitude, latitude, point of interest type and number of registrations. Each point of interest p∈P is defined as auxiliary information p=<lat,lon,type,freq> ,The parameters in the tuple represent the longitude, latitude, POI type and registration times respectively;

[0040] The user action point is represented by the check-in tuple, and all the action points corresponding to the user in time are connected to form the individual trajectory information J corresponding to the user. u =(q 1 ,q 2 ,…,q N ), the individual trajectory information of all users is integrated to obtain the complete user trajectory information Figure 2 is a schematic diagram of the trajectory information of the points of interest according to an embodiment of the present invention. Figure 2As shown, the separate trajectory information of two users is displayed;

[0041] Gathering all user trajectory information can effectively solve the cold start problem in POI recommendation;

[0042] Use the trajectory information of all users to construct a directed graph G = (V, E, f p ,w p ), the directed graph includes graph structure and node features, the nodes are points of interest, and the edge weight information is the number of visits between all users between points of interest, where V represents the node information of the points of interest, E represents the edge information between the points of interest, and f p represents the auxiliary information of p∈P, w p represents the weight, i.e., the number of times the same trajectory segment appears;

[0043] Based on the directed graph, the graph attention network is used to obtain the vectorized representation of the topological information of the POI point of interest. The graph structure and node features are input into the graph attention network model. Let A∈R N×N is the adjacency matrix of graph G, and the attention coefficient is calculated to capture the relationship between nodes. Let H(0) = X∈R N×C is the input node feature matrix, where X is a matrix including the longitude, latitude, time and type features of each point of interest. The attention coefficient method is used to update the features of each node in the directed graph. Formula 3 is used to calculate the attention coefficient of the directed graph node and its neighboring nodes:

[0044] α ij =Leaky ReLU(a T |W n H(i)||W n H(j)|) Formula 3;

[0045] Among them, Leaky ReLU is an improved activation function that can solve the "dead neurons" problem that may occur in the traditional ReLU function. n represents the linear transformation matrix, a represents the learning vector of the attention weight, represents the splicing operation, and in the embodiment of the present invention, the attention coefficient needs to be normalized Where N(i) represents the neighbor set of node i, then the propagation rule of the lth layer can be defined as: Among them, H (l-1) is the input feature of the previous layer, W n (l) is the weight matrix of the lth layer, b n (l) is the corresponding bias and σ is the activation function.

[0046] Stack *A graph attention mechanism layer is added to enhance the expressiveness of the model. Before the last layer, the dropout technique is used to prevent overfitting. The output is expressed as in Ω is the output dimension.

[0047] In step S120, the user feature vector and the interest point feature vector are merged into a trajectory interaction feature vector, specifically including:

[0048] The user feature vector and the point of interest feature vector are concatenated to obtain the cascade feature, which represents the interactive features between the user's visit trajectory location features and the user's similar features, effectively capturing the interactive information between users' trajectories;

[0049] The cascade features are input into the fully connected layer for fine-tuning. By learning a set of adjustment weights, the cascade features are transformed and deeply mapped to improve their expression and information capture capabilities, and the trajectory interaction feature vector is output. The output is represented as e u,p =σ(w u,p [e u ;e p ]+b u,p )∈R Ο×2 , where W p,u and b u,p denote the weight vector and bias respectively, [;] denotes the connection operation, and Ω×2 denotes the dimension of embedding.

[0050] In step S130, the interest point type and time information in the user trajectory information are obtained, the interest point type and time information are encoded, and the encoding result is converted into vector dimension information and input into the time series model to obtain the associated feature vector of the interest point type changing with the time period, specifically including:

[0051] In the recommendation system, users often show dependence on a certain type of preference, which means that when users choose places of interest, they will be influenced by past behaviors and preferences. For example, the check-in rate of catering points of interest is usually higher during lunch and dinner hours. In addition, seasonal changes in user check-in time may also affect the choice of point of interest type. Therefore, type information can be trained as an embedding layer to capture the regularity and behavior patterns of users on different types of points of interest.

[0052] Extract the POI type from the POI information in the check-in tuple. Because in the process of forming the trajectory interaction feature, the user's POI already includes information such as location, POI type, and time, which can be directly obtained through the access record and encoded as e t =f embed (type)∈R Ψ , where Ψ represents the embedding dimension;

[0053] The user's access behavior is naturally time-dependent. Different types of POIs will have different peaks and valleys in a day. Therefore, it is necessary to effectively capture the periodicity and structural correlation characteristics of time. The Time2Vec model is used to obtain the periodic features of POIs through Formula 4. v [i]:

[0054]

[0055] Among them, w, represents a learnable parameter, the sin activation function is used to capture periodic patterns, and e v The vector dimension of [i] is k+1;

[0056] In order to integrate more time features, features based on historical event information are proposed, that is, events related to the check-in of the POI that occurred before the time scale period. The Chronos model is used to obtain the historical event features at different time scales, and the periodic features and historical event features are spliced ​​to obtain the time coding of the POI. The feature vector C(t)=|d is output to capture the relevant time. week ||m mouth ||h hour |, where d week Indicates the week code, m month Indicates the month code, h hour represents hour code, || represents concatenation;

[0057] Time encoding is represented by the fusion result of Time2Vec and Chronos output. T =concat(e v ,C(t))∈R μ , where R μ Dimensions representing type embeddings;

[0058] Concatenate the interest point type code and the interest point time code and convert them into vector dimension information;

[0059] The type code and time code are concatenated to obtain the concatenated features, forming a common representation of category and time. The type feature of the interest point is embedded as e t =f embed (type)∈R Ψ , the temporal feature is embedded as Concatenate the two embedding vectors to get a new vector f combined =|e t ||e T |∈R Ψ+μ , as the vector dimension information, || represents the concatenation operation;

[0060] The vector dimension information is input into the time series model of the fully connected layer, and the associated feature vector of the interest point type in the vector dimension information that changes with the time period is learned through the learning mechanism; after inputting into the fully connected layer, the activation of the fully connected layer changes to f end =ReLU(W h ×f combined +b j ), where W h ∈R m×(Ψ+μ) represents the learnable weight matrix, b j ∈R m represents the bias term, ReLU represents the activation function;

[0061] The nonlinear relationship between time and interest point type in the splicing feature is learned through the learning mechanism, and the process is f interaction =ReLU(W interaction ×f combined +b interaction ), where W interaction and b interaction Represents the new learning parameters and outputs the associated feature vector.

[0062] In step S140, the trajectory interaction feature vector and the associated feature vector are used to train a deep learning network coupled with importance scoring and sparse attention mechanism to obtain the next point of interest recommendation model, which specifically includes:

[0063] Concatenate the trajectory interaction feature vector and the correlation feature vector q =[e u,p ;e t,T ] as a fusion feature, each input trajectory (q 1 ,q 2, …,q N ) will be represented as a series of interest point embeddings (e q1 ,e q2 ,…,e qN );

[0064] The fusion features are used to train a deep learning network, and all the POIs to be selected are selected as candidate POIs in the deep learning network according to the importance scoring mechanism;

[0065] Dynamically select importance through importance scoring mechanism i =σ(X[l]W j ), where W jrepresents the linear transformation weights of different trajectories, σ represents the activation function, and makes the importance score in the range of [0,1]. The importance score is used to determine whether to perform the next step of calculation. For the input X[l], a selective mask M is defined. If the importance score is higher than γ, the selective mask value is 1, otherwise it is 0. The current points of interest that can be recommended as alternatives can be filtered, and the output is filtered as filtered_output=e qi ⊙M, where e qi is the checkpoint embedding sequence mentioned above, ⊙ represents element-wise multiplication;

[0066] In order to reduce complexity and unnecessary information processing, a query system for candidate points of interest is established based on the sparse attention mechanism, that is, a query system is established based on the above filtering results, and the query vector Q, key vector K, and value vector V of the feature vector X[l] are defined, where Q represents the entry for establishing the query system for the current position, K represents the position feature used to match the query, and V represents the information of the actual feature;

[0067] Generate Q = X[l]W through the learned linear change q , K=X[l]W k , V=X[l]W v , where W q , W k , W V Represents the weight matrix. Only the most relevant m keys will be selected to calculate the matrix score. The sparse attention score S i,j By calculating the correlation between Q and K, only the m elements with the largest correlation are retained.

[0068] Perform softmax normalization on S to obtain the attention weight Here, the softmax operation ensures that the weight of each row sums to 1, and finally the output is obtained by weighted summation of the attention weight and the value V. Aggregate the output to form the output tensor H h , specifically H h ∈R k×d ; Based on the query system, an output tensor of interest points is formed.

[0069] The output tensor is input into the multi-layer perceptron to obtain the recommendation probability of the point of interest. First, linear changes and activations are performed, which is represented by Z (1) =ReLU(W x (1) filtered_output+b x (1) ), Z (2) =ReLU(W x (2) Z(1) +b x (3) ), taking the three-layer perception layer as an example, the probability distribution of the final output is calculated by softmax to obtain P(q k+1 )=softmax(W x 3 Z( 2) +b x (3) ), where W x (3) and b x (3) They represent the weights and biases of the last layer of the multi-layer perception layer respectively, and the output shape is the probability distribution of k categories, that is, the probability scores of each point of interest to be recommended, and the point of interest with the highest probability is selected as the next point of interest;

[0070] Iteratively train the deep learning network that couples importance scoring and sparse attention mechanism to obtain the next point of interest recommendation model.

[0071] Preferably, in an embodiment of the present invention, a convolutional neural network is used to extract text features such as user comments and POI comments to assist in the training of a POI recommendation model.

[0072] The method further comprises:

[0073] In step S150, the trajectory sequence of the user to be recommended is input into the interest point recommendation model, and the next interest point is output, which specifically includes:

[0074] Input the trajectory sequence of the user u to be recommended The goal is to predict the next point of interest Compute each candidate checkpoint embedding to construct the input tensor, denoted as X|0|∈R k×d , where the embedding dimension is d = 2×(Ω+Ψ).

[0075] To sum up, in response to the existing problems, the next point of interest recommendation method that integrates interactive features and associated features is invented in this invention. The similarity between users is obtained through a matrix decomposition model, and the point of interest embedding is obtained based on the complete point of interest trajectory information, which solves the cold start problem in point of interest recommendation, and fuses the user feature vector and the point of interest feature vector into a trajectory interactive feature vector; and considers auxiliary associated information such as point of interest type and check-in time as an associated feature vector, explores the relationship between the user's check-in behavior and time at different types of points of interest, and reveals the user's preferences, behavior patterns and potential consumption habits, so as to better understand the user's behavior pattern; uses the above two types of features to train a deep learning network model, effectively taking into account the collaborative behavior between users, and greatly improving the accuracy of the recommendation results; in the application process of the point of interest recommendation model, the importance scoring mechanism is used to screen out alternative points of interest, and a query system is established, and then the probability scores of the alternative points of interest are output through a multi-layer perception layer, reducing unnecessary information processing.

[0076] Device Embodiment

[0077] According to an embodiment of the present invention, a device for recommending a next point of interest by fusing interactive features with associated features is provided. Figure 3 is a schematic diagram of a device for recommending next points of interest by fusing interactive features with associated features according to an embodiment of the present invention. Figure 3 As shown, the next point of interest recommendation device for fusing interactive features and associated features according to an embodiment of the present invention specifically includes:

[0078] The initial modeling module 30 is used to obtain the similarity between users using a matrix decomposition model to obtain a user feature vector; use the user trajectory information to construct a directed graph of all users, input the directed graph into the graph attention network model, and obtain an interest point feature vector, which is specifically used for:

[0079] Construct the user's sign-in matrix, use the bias term to decompose the sign-in matrix, establish the objective function that minimizes the difference between the bias evaluation value and the actual sign-in, and use the adaptive gradient algorithm to solve the objective function;

[0080] Perform linear transformation on the solution result to obtain the user feature vector.

[0081] Define a check-in tuple including user information, POI information and timestamp, where POI information includes longitude, latitude, POI type and registration times. Use the check-in tuple to represent the user's action point, connect all the action points of the corresponding user in time to form the individual trajectory information of the user, and merge the individual trajectory information of all users to obtain the complete user trajectory information;

[0082] A directed graph is constructed using all user trajectory information, where the nodes of the directed graph are points of interest, and the edge weight information is the number of visits between all user points of interest. The attention coefficient method is used to update the features of each node in the directed graph, and the stacked graph attention network model is used to output the point of interest feature vector.

[0083] The interactive fusion module 32 is used to fuse the user feature vector and the interest point feature vector into a trajectory interactive feature vector, specifically for:

[0084] The user feature vector and the interest point feature vector are concatenated to obtain the cascade feature;

[0085] The cascade features are input into the fully connected layer, and the cascade features are transformed and mapped by adjusting the weights to output the trajectory interaction feature vector.

[0086] The auxiliary association module 34 is used to obtain the type and time information of the interest point in the user trajectory information, encode the type and time information of the interest point, convert the encoding result into vector dimension information and input it into the time series model to obtain the associated feature vector of the interest point type changing with the time period, which is specifically used for:

[0087] Use the Time2Vec model to obtain the periodic features of points of interest, use the Chronos model to obtain the features of historical events at different time scales, and combine the periodic features and historical event features to obtain the time coding of points of interest;

[0088] Concatenate the interest point type code and the interest point time code and convert them into vector dimension information;

[0089] The vector dimension information is input into the time series model of the fully connected layer, and the associated feature vectors of the interest point types in the vector dimension information that change with the time period are learned through the learning mechanism.

[0090] The model training module 36 is used to train a deep learning network coupling importance scoring and sparse attention mechanism using the trajectory interaction feature vector and the associated feature vector to obtain a next point of interest recommendation model, which is specifically used for:

[0091] The trajectory interaction feature vector and the correlation feature vector are concatenated as fusion features;

[0092] The fusion features are used to train a deep learning network, and all the POIs to be selected are selected as candidate POIs in the deep learning network according to the importance scoring mechanism; a query system for candidate POIs is established according to the sparse attention mechanism;

[0093] Based on the query system, an output tensor of interest points is formed, and the output tensor is input into a multi-layer perceptron to obtain the recommendation probability of the interest point, and the next interest point recommendation model is obtained through iterative training.

[0094] The device further comprises:

[0095] The model application module 38 inputs the trajectory sequence of the user to be recommended into the interest point recommendation model and outputs the next interest point.

[0096] To sum up, in response to the existing problems, the next point of interest recommendation device that integrates interactive features and associated features of this invention obtains the similarity between users through a matrix decomposition model, and the point of interest embedding is obtained based on the complete point of interest trajectory information, which solves the cold start problem in point of interest recommendation, and fuses the user feature vector and the point of interest feature vector into a trajectory interactive feature vector; and considers auxiliary associated information such as point of interest type and check-in time as an associated feature vector, explores the relationship between the user's check-in behavior and time at different types of points of interest, and reveals the user's preferences, behavior patterns and potential consumption habits, so as to better understand the user's behavior pattern; uses the above two types of features to train a deep learning network model, effectively taking into account the collaborative behavior between users, and greatly improving the accuracy of the recommendation results; in the application process of the point of interest recommendation model, the importance scoring mechanism is used to screen out alternative points of interest, and a query system is established, and then the probability scores of the alternative points of interest are output through a multi-layer perception layer, reducing unnecessary information processing.

[0097] Electronic device embodiment

[0098] Figure 4 4 is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device 400 may include at least one processor 410 and a memory 420. The processor 410 may execute instructions stored in the memory 420. The processor 410 is connected to the memory 420 through a data bus. In addition to the memory 420, the processor 410 may also be connected to an input device 430, an output device 440, and a communication device 450 through a data bus.

[0099] The processor 410 may be any conventional processor, such as a commercially available CPU. The processor may also include a graphics processor (Graphic Process Unit, GPU), a field programmable gate array (Field Programmable Gate Array, FPGA), a system on chip (System on Chip, SOC), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC) or a combination thereof.

[0100] Memory 420 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0101] In the embodiment of the present disclosure, executable instructions are stored in the memory 420, and the processor 410 can read the executable instructions from the memory 420 and execute the instructions to implement all or part of the steps of the next point of interest recommendation method for fusing interactive features and associated features in any of the above exemplary embodiments.

[0102] Computer Readable Storage Medium Embodiments

[0103] In addition to the above-mentioned methods and devices, the exemplary embodiments of the present disclosure may also be a computer program product or a computer-readable storage medium storing the computer program product. The computer product includes computer program instructions, which can be executed by a processor to implement all or part of the steps described in the next interest point recommendation method for fusing interactive features and associated features in any of the above-mentioned exemplary embodiments.

[0104] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages ​​such as Java, C++, etc., and also conventional procedural programming languages ​​such as "C" language or similar programming languages ​​and scripting languages ​​(e.g., Python). The program code may be executed entirely on the user computing device, partially on the user computing device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0105] Computer readable storage media can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples of readable storage media include: a static random access memory (SRAM) with one or more wires electrically connected, an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk, or any suitable combination of the above.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A next point of interest recommendation method by fusing interactive features and associated features, characterized in that: include: A matrix decomposition model is used to obtain the similarity between users and obtain a user feature vector; a directed graph of all users is constructed using user trajectory information, and the directed graph is input into a graph attention network model to obtain a point of interest feature vector; Merging the user feature vector and the interest point feature vector into a trajectory interaction feature vector; Obtaining the type and time information of the point of interest in the user trajectory information, encoding the type and time information of the point of interest, converting the encoding result into vector dimension information and inputting it into the time series model to obtain the associated feature vector of the type of the point of interest changing with the time period; The trajectory interaction feature vector and the association feature vector are used to train a deep learning network that couples importance scoring and sparse attention mechanism to obtain a next point of interest recommendation model.

2. The method according to claim 1, characterized in that The method further comprises: The trajectory sequence of the user to be recommended is input into the interest point recommendation model, and the next interest point is output.

3. The method according to claim 1, characterized in that The method of using the matrix decomposition model to obtain the similarity between users and obtain the user feature vector specifically includes: Constructing a user's sign-in matrix, decomposing the sign-in matrix using a bias term, establishing an objective function that minimizes the difference between the bias evaluation value and the actual sign-in, and solving the objective function using an adaptive gradient algorithm; Perform linear transformation on the solution result to obtain the user feature vector.

4. The method according to claim 1, characterized in that: The method of using user trajectory information to construct a directed graph of all users and inputting the directed graph into a graph attention network model to obtain a feature vector of a point of interest specifically includes: Define a check-in tuple including user information, point of interest information and timestamp, wherein the point of interest information includes longitude, latitude, point of interest type and registration times, the user action point is represented by the check-in tuple, all action points of the corresponding user in time are connected to form the individual trajectory information corresponding to the user, and the individual trajectory information of all users is merged to obtain the complete user trajectory information; A directed graph is constructed using the user trajectory information, wherein the nodes of the directed graph are points of interest, the weight information of the edges is the number of visits by all users between the points of interest, the attention coefficient method is used to update the features of each node in the directed graph, and the stacked graph attention network model is used to output the point of interest feature vector.

5. The method according to claim 1, characterized in that: The step of fusing the user feature vector and the interest point feature vector into a trajectory interaction feature vector specifically includes: Concatenating the user feature vector and the interest point feature vector to obtain a cascade feature; The cascade features are input into a fully connected layer, the cascade features are transformed and mapped by adjusting weights, and a trajectory interaction feature vector is output.

6. The method according to claim 1, characterized in that The interest point type and the time information are encoded, and the encoding result is converted into vector dimension information and input into the time series model to obtain the associated feature vector of the interest point type changing with the time period, which specifically includes: The Time2Vec model is used to obtain the periodic features of the points of interest, the Chronos model is used to obtain the features of historical events at different time scales, and the periodic features and the historical event features are concatenated to obtain the time coding of the points of interest; Concatenate the interest point type code and the interest point time code and convert them into vector dimension information; The vector dimension information is input into the time series model of the fully connected layer, and the associated feature vectors of the interest point types in the vector dimension information that change with the time period are learned through a learning mechanism.

7. The method according to claim 1, characterized in that The step of using the trajectory interaction feature vector and the association feature vector to train a deep learning network coupled with importance scoring and sparse attention mechanism to obtain a next point of interest recommendation model specifically includes: splicing the trajectory interaction feature vector and the correlation feature vector as a fusion feature; Using the fusion features to train a deep learning network, selecting all the points of interest to be selected in the deep learning network according to an importance scoring mechanism as candidate points of interest; establishing a query system for the candidate points of interest according to a sparse attention mechanism; An output tensor of interest points is formed based on the query system, and the output tensor is input into a multi-layer perceptron to obtain the recommendation probability of interest points, and the next interest point recommendation model is obtained through iterative training.

8. A device for recommending the next point of interest by fusing interactive features with associated features, characterized in that: include: The initial modeling module is used to obtain the similarity between users using the matrix decomposition model and obtain the user feature vector; Use user trajectory information to construct a directed graph of all users, input the directed graph into a graph attention network model, and obtain a feature vector of interest points; An interactive fusion module, used to fuse the user feature vector and the interest point feature vector into a trajectory interactive feature vector; The auxiliary association module is used to obtain the interest point type and time information in the user trajectory information, encode the interest point type and the time information, convert the encoding result into vector dimension information and input it into the time series model to obtain the associated feature vector of the interest point type changing with the time period; The model training module is used to train a deep learning network coupling importance scoring and sparse attention mechanism using the trajectory interaction feature vector and the association feature vector to obtain a next point of interest recommendation model.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the next point of interest recommendation method by fusing interactive features with associated features as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by the processor, the steps of the next point of interest recommendation method for fusing interactive features with associated features as described in any one of claims 1 to 7 are implemented.

Images