A semantic enhancement-based location recommendation method

Abstract

The application relates to a place recommendation method based on semantic enhancement, which comprises the following steps: acquiring a user trajectory sequence set arranged according to check-in time sequences of a plurality of users, constructing a local hypergraph and a global hypergraph; acquiring global user representation and global place representation, local user representation and local place representation; constructing prompt words and encoding to obtain global user behavior semantic representation and local user behavior semantic representation; calculating final trajectory end representation and final user end representation X f , and X f splicing to obtain sequence representation, performing sequence modeling according to the sequence representation, inputting the final feature obtained by modeling into a neural network V to predict a next check-in time, a check-in place and a check-in category; calculating total loss to train model parameters to be optimal; inputting sequence representation of a complete long trajectory of a user into the trained model to output a next check-in place of the user. Experimental results prove that the method can provide more accurate and personalized recommendation results for users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data, and particularly relates to a place recommendation method based on semantic enhancement. BACKGROUND

[0002] Point of interest recommendation is to predict the next place of interest by modeling the user's historical trajectory, wherein the modeling of the user's trajectory relies on the discrete time sequence of check-in point information of the user, and the user's travel behavior pattern is extracted and mined by modeling the same. The research on point of interest recommendation mainly relies on the user number, place number, place longitude and latitude and check-in time information in the user's check-in data, which are the most easily collected and encoded features. Semantic mining through such context information can effectively help the modeling of user trajectory and preference. Secondly, a small number of service providers have opened source their user comments on places in some open source data sets, which reflect the user's preference and description of the place through a large amount of natural language, and explicitly show the attractiveness of different places and user experience. Such natural language information is an important factor for humans to filter places of interest, but for models, further feature extraction and encoding are needed. In recent years, the rapid development of the field of natural language processing has integrated a large amount of world knowledge into pre-training models, enabling them to effectively understand, encode and generate deep semantic information based on natural language, giving hope to semantic feature extraction in the recommendation task.

[0003] However, it is difficult for the model to extract effective collaborative signals related to user interests from a large amount of natural language text, and the existing excellent pre-training model needs to consume a large amount of computing resources and energy, and the use cost is high. On the data level, the semantic data is still low-quality, and different users have their own personalized preferences and cognition, and the comments on the place have a lot of personal emotional color. On the other hand, the effective comment density in the user comment is low, and in reality, there are a lot of marketing behaviors such as good review return, and the user's comment is not necessarily completely accurate. Therefore, encoding the semantic information of low-quality user trajectory data not only consumes resources, but also is prone to the problem of garbage in garbage out. Therefore, using a large amount of comment text information to encode the deep semantic representation of the place and the user not only consumes resources, but also may not have a significant effect. However, in addition to the comment text, there are also other natural language descriptions in the check-in point data, which are the place categories of the interest point. The place category information is a functional description of the place, which highly condenses the role of the place, is an absolutely accurate, abstract and highly generalized text data label, and is of great help to the cognition of the model to the place. In the past research, the use of place categories mainly converts them into numbers and performs shallow neural network embedding, and cooperates with other temporal and spatial context information to complete the extraction of the shallow semantics of the place. This method does not effectively utilize the explicit natural language semantic information, and the recommendation does not have generalization and universality. SUMMARY

[0004] In view of the above problems existing in the prior art, the technical problem to be solved by the present application is how to model the behavior pattern of the user through the low semantic user check-in data and perform deep semantic enhancement of the user behavior semantics.

[0005] To solve the above technical problems, the present application adopts the following technical solutions: a place recommendation method based on semantic enhancement, comprising the following steps:

[0006] S1: obtaining a user trajectory sequence set S composed of a plurality of users according to the check-in time sequence, and then constructing a global hypergraph G with the complete long trajectory as a hyperedge and the check-in place as a node global The i-th check-in place l u_{i,j} corresponds to a unique tuple, which includes latitude lat, longitude lon, check-in time time, place category cat, place category number and check-in point number;

[0007] Each complete long trajectory is truncated by a time threshold Δt=24h to obtain a plurality of short trajectories, and the j-th short trajectory of the i-th user is denoted as S i_j={l i_{j,1} ,l i_{j,2} ,l i_{j,3} ,...,l i_{j,p}}, where p is the length of the short trajectory, and a local hypergraph G is constructed by using all short trajectories as hyperedges and check-in locations as nodes. local All short trajectories constitute a short sequence set HS.

[0008] S2: G global The input hypergraph convolutional network yields global user representations and global location representations, G. local Inputting a hypergraph convolutional network yields local user representations and local location representations.

[0009] S3: Construct prompt words from the semantic information in the long trajectory, and then use a pre-trained language model to encode the constructed prompt words to obtain a global user behavior semantic representation. Construct prompt words from the semantic information in the short trajectory, and then use a pre-trained language model to encode the constructed prompt words to obtain a local user behavior semantic representation.

[0010] The location category is obtained by one-hot encoding. cat Regarding the user's check-in time, a 24-hour day is divided into several timestamps, and the check-in time is projected onto each timestamp to obtain the time embedding. time ;

[0011] Using existing methods to l cat and l time Embedding is performed to obtain the corresponding time representation h. time and category representation h cat .

[0012] Long trajectories are also called global trajectories, and short trajectories are also called local trajectories. Pre-trained language models use existing models.

[0013] S4: Global user behavior semantic representation and local user behavior semantic representation are obtained by dimensionality reduction through neural network I to obtain the corresponding global semantic representation and local semantic representation;

[0014] The global user representation, global semantic representation, and user ID embedding are concatenated and input into a neural network II for dimensionality reduction. Then, they are concatenated with the global location representation and input into a fusion network to obtain the final trajectory representation.

[0015] Local semantic representations and local location representations are concatenated and input into neural network III for dimensionality reduction to obtain matrix K. Temporal representations and category representations are concatenated and input into neural network IV for dimensionality reduction to obtain matrix Q. Local user representations are used as matrix V. K, Q, and V are processed by linear layers and then enter an attention mechanism to obtain the final user-side representation X. f .

[0016] S5: Will and X f Concatenation yields sequence representation According to the obtained Perform sequence modeling, and input the final features obtained from the modeling into neural network V to predict the next check-in time, check-in location, and check-in category;

[0017] S6: Freeze the parameters of the pre-trained model, calculate the total loss, and backpropagate based on the total loss to update the parameters of all neural networks until the total loss no longer decreases. At this point, the parameters of all neural networks are optimal.

[0018] S7: Process a user's complete long trajectory according to the methods of S1-S4 to obtain the user's sequence representation. During this process, use a neural network with optimal parameters. Input the user's sequence representation into the neural network V with optimal parameters and output the user's next check-in location.

[0019] Furthermore, in S1, G global The construction process is as follows:

[0020] The incidence matrix representing the hypergraph topology The construction method is as follows:

[0021]

[0022] Where e is a hyperedge, v is a node, and e∈ε global , ε global Let be the set of hyperedges and nodes of the global hypergraph. The degree of a hyperedge is the number of nodes it connects, and represents the number of locations covered by the user's trajectory. Therefore, its formula is: The degrees of all hyperedges form the diagonal hyperedge degree matrix D. ge Assign weights W e For each hyperedge, the weights of all hyperedges together form a diagonal matrix. Each node The degree of this node is obtained by multiplying the diagonal matrix and the matrix of hyperedge weights. The degrees of all nodes form the diagonal degree matrix D. gv .

[0023] G in S1 local The construction process is as follows: Local hypergraph G local Correlation matrix The definition is the same as in formula (1), and the weights are assigned to each hyperedge in the same way as in the global hypergraph to form a diagonal matrix. Construct the diagonal hyperedge degree matrix D using a diagonal matrix le and the degree matrix D of the diagonal nodes lv .

[0024] Further, in S2, the steps of obtaining global user representation and global location representation and local user representation and local location representation are:

[0025] By embedding matrix The check-in location sequence with length |L| is embedded into a feature space with dimension d, and the location embedding input into the first layer of the hypergraph neural network is Then, according to the global hypergraph features of this part, the operation mechanism of the global hypergraph convolution network is as follows:

[0026]

[0027] Wherein is the location embedding set vector of the kth layer of the global hypergraph convolution network, and the details of the two-stage information transmission and aggregation are as follows: through to represent the process of the first-stage location representation converging from the node to the hyperedge, and after the aggregation of the global trajectory information of the user on the hyperedge, through transmitted back to the location representation node, and the global location representation is updated using the global information, wherein is a normalization operation.

[0028] The is input into the hypergraph convolution network to obtain the local user representation and the local location representation.

[0029] The operation mechanism of the local hypergraph convolution network is as follows:

[0030]

[0031] Wherein is the location embedding vector set of the kth layer of the local hypergraph convolution network, and the details of the two-stage information transmission and aggregation are the same as those of the global hypergraph convolution network, is the location embedding set vector of the kth layer of the local hypergraph convolution network, and the details of the two-stage information transmission and aggregation are as follows: through to represent the process of the first-stage node representation converging to the hyperedge, and after the aggregation of the local trajectory information of the user on the hyperedge, through transmitted back to the location representation node, and the local location representation is updated using the local information, wherein is a normalization operation.

[0032] Through the hypergraph convolution calculation of N layers, the global location representation set aggregated by the global information and the local location representation set aggregated by the local information

[0033] The user representation is obtained from , and the S i, the global user representation γ is obtained by accumulating and averaging all the location representations on the hyperedge global , which is represented as:

[0034]

[0035] The user representation is obtained from , and the local user representation γ is obtained by accumulating and averaging all the location representations on the hyperedge i_j local , which is represented as:

[0036]

[0037] Further, the steps of obtaining the global user behavior semantic representation and the local user behavior semantic representation in S3 are:

[0038] First, the construction method of the global prompt is as follows: according to the trajectory S i = {l i,1 , l i,2 , l i,3 ,..., l i,k} of the i-th user, the global location category sequence is constructed by the category of each location in the trajectory On the basis of C i , the redundant categories are filtered out, and the category sequence {cat1, cat2, cat3,..., cat c} is a unique value sequence. After obtaining the category sequence of the user, it is spliced with two parts: first, the sentence “This user visited” is spliced in front of the category sequence to describe the encoding task, and then the special symbol [CLS] is spliced at the beginning to facilitate the extraction of the sentence vector. In formula (6), the semicolon represents the splicing of the text, and the global prompt representing the global semantic signal is represented as:

[0039] Prompt global = [[CLS]; This user visited; cat1, cat2, cat3,..., cat c ] (6)

[0040] When constructing the local prompt, the last element of the constructed location category sequence is removed, and then the redundant categories are filtered out on the basis of the local location category sequence, so that the local category sequence {cat1, cat2, cat3,..., cat z ​The sequence is a unique value sequence. First, the statement "This user recently visited" is concatenated before the local category sequence. Then, the special symbol [CLS] is concatenated at the beginning of the sentence. The local cue word representing the local semantic signal is represented as follows:

[0041] Prompt local =[[CLS];This user rencently visited;cat1,cat2,cat3,...,cat c (7)

[0042] The global language extraction model encodes global prompts to obtain a global semantic representation of user behavior, and the global language extraction model encodes local prompts to obtain a local semantic representation of user behavior.

[0043]

[0044]

[0045] Where δ global , This refers to the global user behavior semantic representation and the local user behavior semantic representation corresponding to the [CLS] symbol in the final hidden layer of the global language extraction model and the local language extraction model.

[0046] Furthermore, the final trajectory end representation is obtained in S4. The steps are as follows:

[0047] Using formula (10) to represent the semantics of global user behavior δ global Dimensions to be performed:

[0048]

[0049] in The weight matrix is ​​a learnable matrix. The bias of this linear layer is set, and the LeakyReLU activation function is used.

[0050]

[0051] Where ⊕ represents the vector concatenation operation, To create a learnable weight matrix, the concatenated 4D... τ 2D vector fusion τ The end-user representation of dimension U f In. By using short sequence S i_j Global location representation in By integrating the data, a preliminary representation of the final trajectory from the user's perspective is constructed:

[0052]

[0053] wherein, is the global place representation Z g from which the corresponding place representation, ω se is selected.

[0054] Further, the step S4 of obtaining the final user-end representation X f is performed by:

[0055] using three learnable weight matrices to learn the query vector Q, the key vector K and the value vector V:

[0056] Q = e f W Q (16)

[0057]

[0058] V = g local W V (18)

[0059] wherein, in the process of constructing the Q, K, V vectors, is a learnable weight matrix, is the bias of the linear layer, denotes the local place representation in the short sequence S i_j .

[0060]

[0061] wherein, d k is a scale parameter with the same value as d ε .

[0062] Further, the step of performing sequence modeling on in the S5, and inputting the final features obtained by the modeling into the neural network V to predict the next check-in time, check-in place and check-in category is performed by:

[0063] The modeling of the sequence representation is performed by:

[0064]

[0065] wherein, F is the modeled feature, ω complex is a learnable weight matrix.

[0066]

[0067]

[0068] where ω r1 and ω r2 are learnable weight vectors, with GELU Gaussian Error Linear Unit as their activation function. is the encoding representation of the first layer of the neural network, is the final encoding representation of the neural network.

[0069]

[0070]

[0071]

[0072] where ω poi , ω time , ω poi are learnable parameters, b p , b t , b c are their corresponding biases, and denote the predicted check-in location, predicted check-in time and predicted check-in category, respectively.

[0073] Further, the step of calculating the total loss in S6 is:

[0074] According to the point of interest prediction loss time prediction loss and category prediction loss the total loss function obtained is:

[0075]

[0076] where, and cross-entropy is used to calculate, using the root mean square to calculate.

[0077] Compared with the prior art, the present application has at least the following advantages:

[0078] 1. The present application uses hypergraph neural networks:

[0079] Global hypergraph: used to mine the transfer patterns of user groups within the entire trajectory, capture the association between user behaviors and high-order collaborative signals. Local hypergraph: focuses on short trajectories of specific consecutive check-ins, analyzes the behavior patterns of users in local areas, and understands user behavior at a finer granularity.

[0080] 2. The present application performs semantic enhancement:

[0081] Prompt word engineering: based on the category text information of the user check-in location, explicit semantic templates are constructed, which reflect the functionality and purpose of user behavior. Pre-trained language model: using the prior knowledge in the pre-trained language model, the semantic templates of user behavior are deeply encoded, and rich semantic features are extracted.

[0082] 3. The present application adopts fusion of topological space and semantic space:

[0083] The topological space features captured by the hypergraph neural network are fused with the semantic space features extracted by the pre-trained language model to obtain more comprehensive and in-depth user representation. This fusion representation can consider both the geographical location behavior and semantic intent of the user, providing more accurate information for point of interest recommendation.

[0084] 4. The present application adopts deep semantic enhanced recommendation:

[0085] Using the fused user representation for point of interest recommendation, through deep semantic enhancement, the model can better understand the personalized preferences and potential needs of the user. The recommendation system can therefore provide more personalized and accurate point of interest recommendation results.

[0086] 5. The present application models the user's behavior patterns using low-semantic user check-in data and performs deep semantic enhancement of user behavior semantics. HyperSE effectively alleviates the problem of shallow semantic modeling in low-quality data, improving the accuracy and generalization of point of interest recommendation.

[0087] 6. By mining high-order collaborative signals of user group transfer patterns globally and locally, the influence of low-quality data is alleviated, and the user's behavior patterns are more accurately mined. Using prompt word engineering to model explicit semantic templates of user behavior according to the category text information of the user check-in location, the encoding quality of user behavior semantics is further improved. Using the rich prior knowledge in the pre-trained language model for deep user behavior semantic encoding, the model can better understand the user's behavior intent and preferences. By fusing the user representation in topological space and semantic space and using it for recommendation, deep semantic enhancement of user check-in behavior is achieved, improving the accuracy and generalization of recommendation. Experimental results prove the effectiveness of the HyperSE model in the point of interest recommendation task, which can provide more accurate and personalized recommendation results for users. BRIEF DESCRIPTION OF DRAWINGS

[0088] Figure 1 The HyperSE model structure diagram.

[0089] Figure 2 The fusion recommendation process.

[0090] Figure 3The number of layers of the hypergraph convolution. DETAILED DESCRIPTION

[0091] The application will be further described in detail below.

[0092] A place recommendation method based on semantic enhancement is referred to as HyperSE. From the global and local perspectives, the user's behavior preferences are captured in a high-order manner, and the user's behavior semantics are enhanced through a pre-trained language model. The method first truncates low-quality trajectory sequences through a time interval threshold, so that each user has a relatively continuous short trajectory set. By using the complete trajectory of the user and the short trajectory set, the user group trajectory and the continuous place access trajectory collaborative signal are constructed through the hypergraph neural network from the global and local perspectives. Secondly, the prompt words of the user behavior preferences based on the category text information in the user trajectory data are constructed, which are input into the pre-trained language model for deep encoding of the user behavior semantics. Through the fusion of the topological space and the semantic space, the global and local behavior preferences of the user are mined, and the user behavior semantics are enhanced. Finally, the recommendation is made based on the sequence model.

[0093] Next POI recommendation aims to predict the next check-in location of a user from a set of places. Assuming that the user set is U = {u1, u2, u3,..., u M}, the place set is L = {l1, l2, l3,..., l N}, which indicates that there are M users and N places in the data set. In the trajectory set S = {S1, S2, S3,..., S M}, each user u i ∈ U has his own trajectory sequence S i = {l i,1 ,l i,2 ,l i,3 ,...,l i,k}, where l i,k denotes the kth visited place of the ith user, containing k personal check-in places arranged in chronological order. In the trajectory sequence of each user, the jth check-in place l i,k of the ith user corresponds to a unique tuple (latitude lat, longitude lon, check-in time time, place category cat, place category number, check-in point number), representing the information of the check-in state and the place attribute. Given the check-in time sequence S of the target user u, the goal of next POI recommendation is to output a ranking list showing the places that the user u may be interested in at the next time stamp, and to make recommendations according to the scores of the places.

[0094] The construction process of G global and G local is as follows:

[0095] The advantage of the supergraph structure is that it can overcome the one-to-one relationship between edges and nodes, capture high-order collaborative signals between data through multi-node edge relationships, and has particularly significant help for the next point of interest recommendation task. In this paper, global supergraph G local =(V local ,E local ) and local supergraph G global =(V global ,E global ) are constructed by using global trajectories and local trajectories to mine user group trajectory and continuous point access trajectory collaborative signals. Where V local and represent the node set of the local and global supergraph, |V local |, |V global | represent the number of nodes of the local and global supergraph. And E local and E global represent the superedge set of the local supergraph and the global supergraph, |E local | and |E global | represent the number of superedges of the local supergraph and the global supergraph.

[0096] A semantic enhancement-based location recommendation method includes the following steps:

[0097] S1: Obtain a plurality of users arranged according to the check-in time sequence Complete long trajectory, then take the complete long trajectory as the superedge, and the check-in location as the node to construct the global supergraph G global . Each complete long trajectory is truncated by a time threshold Δt=24h to obtain a plurality of short trajectories, and the jth short trajectory of the ith user is S i_j ={l i_{j,1} ,l i_{j,2} ,l i_{j,3} ,...,l i_{j,p}}, where p is the length of the short trajectory, all short trajectories are taken as superedges, and check-in locations are taken as nodes to construct the local supergraph G local .

[0098] In the construction of the global supergraph G , the complete long trajectory containing all the check-in points of the user is taken as the superedge, and the check-in location is taken as the node of the supergraph. This method can converge the user's personalized location preferences on the superedge, and model the collaborative signals of the user group trajectory through the supergraph structure. The definition of the association matrix representing the supergraph topology is as follows:

[0099]

[0100] In this context, an element equal to 1 in the correlation matrix signifies that the location represented by that row appears on the user's trajectory represented by that column, for the hyperedge e∈ε representing the user's complete trajectory. global Its degree is the number of nodes connected by the hyperedge, representing the number of locations covered by the user's trajectory, therefore its formula is: The degrees of all hyperedges form the diagonal hyperedge degree matrix D. ge Assign weights W e For each hyperedge, the weights of all hyperedges together form a diagonal matrix. Each node The degree of this node can be obtained by matrix multiplication of the diagonal matrix and the weights of the hyperedges. The degrees of all nodes form the diagonal degree matrix D. gv .

[0101] Local hypergraph Previously, this chapter arranged the users' complete sequences according to the check-in time order and truncated the users' uninterrupted sequences with a time threshold Δt = 24h. The result was that each user had multiple short sequences, and the j-th trajectory data of the i-th user was S. i_j ={l i_{j,1} ,l i_{j,2} ,l i_{j,3} ,...,l i_{j,p}}, where p is the length of the trajectory. Then, all short sequences from all users are included in a set. In this model, all short sequences are used to form the hyperedges of a local hypergraph, and the nodes of the hypergraph are also check-in locations. By capturing the coordinated signals of continuous location access trajectories through the processed short sequences, the local hypergraph... Correlation matrix The definition is the same as in Formula 1, local hypergraph. Correlation matrix Definition and Formula (1), where an element of 1 in the correlation matrix actually means that the location represented by that row appears on the short trajectory represented by that column, and weights are assigned to each hyperedge in the same way to form a diagonal matrix. Construct the diagonal hyperedge degree matrix D using a diagonal matrix le and the degree matrix D of the diagonal nodes lv For specific examples of global and local hypergraphs, please refer to [reference needed]. Figure 1 .

[0102] S2: Will Inputting a hypergraph convolutional network yields global user representations and global location representations. Inputting a hypergraph convolutional network yields local user representations and local location representations.

[0103] Using existing hypergraph convolutional network, capture user group trajectory and continuous place access trajectory collaborative signal, to iteratively mine complex, high-order interest point transfer relationship. Through the node to hyperedge, hyperedge to node two-stage information transmission aggregation, hyperedge not only promotes the aggregation of node information within the hyperedge, but also transmits node information across hyperedges. First, this paper first embeds the places in the dataset with a shallow neural network, and the embedding matrix is The check-in place sequence with length |L| is embedded into a feature space with dimension d, and the place embedding of the first layer of the input hypergraph neural network is Then, according to the global hypergraph features of this part, the operation mechanism of the global hypergraph convolutional network is as follows:

[0104]

[0105] Wherein represents the place embedding set vector of the kth layer of the global hypergraph convolutional network, and the details of the two-stage information transmission aggregation are as follows: through to represent the process of the first-stage node feature converging to the hyperedge, and after the complete trajectory information of the hyperedge user is aggregated, through transmit back to the place representation node, update the place representation with global information, and enrich the high-order user group trajectory collaborative signal contained in the node. Wherein is a normalization operation.

[0106] Similarly, the operation mechanism of the local hypergraph convolutional network is as follows:

[0107]

[0108] Wherein is the set of place embedding vectors of the kth layer of the local hypergraph convolutional network, and the details of the two-stage information transmission aggregation are the same as those of the global hypergraph convolutional network, is the place embedding set vector of the kth layer of the local hypergraph convolutional network, and the details of the two-stage information transmission aggregation are as follows: through to represent the process of the first-stage node feature converging to the hyperedge, and after the complete trajectory information of the hyperedge user is aggregated, through transmit back to the place representation node, update the place representation with global information, and enrich the high-order user group trajectory collaborative signal contained in the node. Wherein is a normalization operation. The calculation theory of the two-stage information transmission aggregation is the same, but after the short sequence set information of the hyperedge is aggregated, the effect of the back-transmission information to the place representation node is slightly different. The local hypergraph convolutional network updates the place representation with local information, and enriches the high-order continuous place access trajectory collaborative signal contained in the node.

[0109] To simplify the calculation, the normalization calculation is removed, and the node neural network feature conversion is also not performed through a fully connected layer, because this paper intends to learn global and local high-order collaborative signals through hypergraph, and to serve the sequence construction of this chapter by obtaining the place representation with such signals through the hypergraph structure. The method in this paper is not directly used for recommendation by directly using the node representation and the user representation, but serves the feature modeling of the sequence, so as to save the calculation resources and avoid the overfitting problem, and the fully connected layer is not used in the hypergraph to learn the neural network of the place representation.

[0110] Through the hypergraph convolution calculation of N layers, the model can preliminarily obtain the global place representation aggregated by global information and local information And the local place representation Then it is integrated into the sequence modeling process in the point of interest recommendation part.

[0111] S3: The semantic information in the long trajectory is constructed into prompt words, and then the pre-trained language model is used to encode the constructed prompt words to obtain the global user behavior semantic representation. The semantic information in the short trajectory is constructed into prompt words, and then the pre-trained language model is used to encode the constructed prompt words to obtain the local user behavior semantic representation.

[0112] Firstly, the semantic information in the sequence needs to be constructed into prompt words, because the prompt word engineering can better help the model to understand the input and model the semantic information according to the prompt word description. Secondly, the pre-trained language model is used to encode the constructed prompt words to capture the required semantic information. Because the derived text from different fields and topics may cross different semantic spaces even through the same encoder, the invention encodes each trajectory data in the user trajectory sequence set S and the short sequence set HS according to the specific semantic elements, and the selected semantic elements are place categories. In this low semantic state, the use of prompt word engineering can not only string low semantic data, but also assist semantic modeling and optimization.

[0113] The construction method of the global prompt word based on the global trajectory and the local trajectory is to construct the global place category sequence according to the category text of each place in the trajectory S i of the i-th user i,1 , i,2 , i,3 , i,k ..., i of the i-th user However, the number of user trajectories is long, and individual category places may be visited by a large number of users, so if the category sequence is not processed and the prompt words are directly constructed for language model training, the user interest extraction will fail. Therefore, in the category sequence C ion the basis of, and then filter out redundant categories, so that the category sequence {cat1, cat2, cat3,..., cat c} is a unique value sequence. After obtaining the category sequence of the user, it is spliced with the two parts, first splicing the sentence "This user visited" (the user visited) before the global category sequence to describe the encoding task, and then splicing the special symbol [CLS] at the beginning of the sentence to facilitate the extraction of the sentence vector. In formula (6), the semicolon represents the splicing of the text, and the semicolon represents the splicing of the text, and the global prompt word representing the global semantic signal is represented as:

[0114] Prompt global = [[CLS]; This user visited; cat1, cat2, cat3,..., cat c ] (6)

[0115] The local prompt word and the global prompt word have two differences, the first local prompt word is directed to the trajectory data from the user trajectory data S to HS, and the second local prompt word removes the last element of the constructed place category sequence, because the last element is the category of the place predicted by the text model, and if it is not removed, it will cause serious data leakage. After constructing the local semantic category sequence {cat1, cat2, cat3,..., cat z} in the same way, first splice the sentence "This user recently visited" (the user recently visited), and then splice the special symbol [CLS] at the beginning of the sentence, so the local prompt word representing the local semantic signal is represented as:

[0116] Prompt local = [[CLS]; This user recently visited; cat1, cat2, cat3,..., cat c ] (7)

[0117] During the training of the model, the user's short trajectory is predicted, so the user's local prompt word is encoded by the local semantic, that is, the user's short-term behavior characteristics, so the word "recently" is used in the process of constructing the prompt word to prompt the language model sequence to be the recommended sequence, which is also the user's short-term behavior. Then, the user's global semantic information corresponding to the short trajectory is used to expand the long-term preference and extract the user's historical visit habits. In order to avoid data leakage, all places in the validation and test set are removed in the construction of the user's global prompt word. After constructing the complete local prompt word, the global prompt word is input into the language model l for encoding, and l is an existing semantic extraction model.

[0118] The global language extraction model encodes the global prompt word to obtain a global user behavior semantic representation, and the global language extraction model encodes the local prompt word to obtain a local user behavior semantic representation.

[0119]

[0120]

[0121] where δ global , is the global embedding vector and the local embedding vector corresponding to the [CLS] symbol in the final hidden layer of the global language extraction model and the local language extraction model. Since in the encoding type language model such as Bert, the overall semantic information of the sentence is represented in this bit. However, considering the low semantic nature of the check-in data, its semantic information is relatively similar at the data level. In order to prevent the sentence vector modeling effect of the model from being poor, another strategy is also adopted in this paper, that is, the final hidden layer output is averaged and pooled, so that the representation of the sentence can also be obtained, and the semantic of the user check-in behavior sequence can be more intuitive and explicit.

[0122] S4: The global user behavior semantic representation and the local user behavior semantic representation are dimensionally reduced by the neural network I to obtain corresponding global semantic representation and local semantic representation.

[0123] The global user representation, the global semantic representation and the user number embedding are spliced and input to the neural network II for dimension reduction, and then spliced and input to the fusion network with the global location representation to obtain the final trajectory end representation

[0124] The local semantic representation and the local location representation are spliced and input to the neural network III for dimension reduction to obtain the matrix K, the time representation and the category representation are spliced and input to the neural network IV for dimension reduction to obtain the matrix Q, and the local user representation is taken as the matrix V. After K, Q and V are processed by linear layers respectively, they enter the attention mechanism to obtain the final user end representation X. f .

[0125] Now we have and and δ global , How to fuse features in different representation spaces is crucial to modeling short sequences and recommendation results.

[0126] Before fusion, the trajectory sequence set for training and testing is constructed. First, the entire sequence of the user is truncated with a time domain value Δt = 24h, and the history data of the user training set is randomly selected n check-in locations and their time information is masked to indicate that the model is pseudo data, and the operation is strictly limited in the training process and training set. Then the short sequence is constructed to train the model.

[0127] For the short sequence S i of user u i_j , first embed the user representation according to the user ID through the embedding layer global , and then fuse the global semantic information global semantic representation δ i_j and the global user group trajectory collaborative signal global user representation at the user end. Therefore, first obtain the global user representation from the global place representation obtained from the global hypergraph, and extract the corresponding user global sequence S i . The operation of accumulating and averaging all place representations on the hyperedge is as follows:

[0128]

[0129] Similarly, the hyperedge representation set can be obtained. δ global has been obtained through the language model, so the next step is to fuse the obtained global user representation, the corresponding hyperedge representation of the user, and the global semantic representation δ global . Although the language model has extensive world knowledge, the dimension of the encoded vector is also very large, such as the embedding dimension d W of the Bert model is 768, and the embedding dimension d W of the Bart model with a parameter quantity of more than 100 million is 1024, which is very different from the dimensions of other representations constructed in this invention and does not belong to the same representation space. In addition, the semantic space learned through the similar Bert pre-training model is non-smooth and anisotropic. Based on the above reasons, under the premise of not consuming too many resources, the semantic space is maintained using a linear layer in the recommendation process of this paper, which is not a representation. This is an explanation why fusion is optimized, and also effectively reduces the embedding dimension of the global semantic representation δ global :

[0130]

[0131] wherein, is a learnable weight matrix, is used. After the linear layer, the original d W dimensional high-dimensional vector is reduced to 2d τ , which is more suitable for fusion with other representations and can also alleviate the anisotropy problem of the model. At this point, the model performs the user-end fusion process:

[0132]

[0133] In formula (11), represents the vector splicing operation, The learnable weight matrix fuses the spliced 4d τ dimensional vector into a 2d τ dimensional final user-end representation U f . This fusion mechanism fuses global information with low complexity. By fusing with the global location representation in the short sequence S i_j , the final trajectory-end representation under the user-end perspective is initially constructed.

[0134]

[0135] After obtaining the final trajectory-end representation, the second semantic and topological space fusion will be reflected in the sequence angle. For the short sequence S i of user u i_j , the local semantic information and continuous access local user representation local location representation are fused at the sequence end. Therefore, first, the user representation is obtained from the local location representation obtained from the local hypergraph, and all location representations on the hyperedge are accumulated and averaged for the training sequence S i_j . Its formula can be shown as:

[0136]

[0137] Similarly, it can be obtained from the hyperedge representation set . δ local has been obtained through the language model. Unlike the fusion of global information at the user end, while fusing local information, it is intended to first explore the time patterns possessed by the location category, such as the group consensus information that people will visit bars at night but not in the morning. Because this is a pattern mining operation, the category embedding is directly used for operation in this part. The Time2Vector method is used to first embed the time in the sequence, and then fuse through learnable parameters to explore the pattern between time and category:

[0138] The location category is one-hot encoded to obtain l cat ​Regarding the user's check-in time, a 24-hour day is divided into 48 timestamps, and the check-in time is projected onto each timestamp based on a time scalar value to obtain l. time Therefore, the embedding of time can be expressed as:

[0139]

[0140]

[0141] First, embed the time data in the sequence using the Time2Vector method. emb For embedding in the network.

[0142] The Time2Vector method is used to first embed the time in the sequence, and then the time is fused through learnable parameters to explore the pattern between time and category. See formula (13) for details:

[0143] e f =ω pattern (Time2Vec(l time )⊕f emb (l cat )),l∈S i_j (15)

[0144] in The learnable matrix represents the pattern mining process, ⊕ represents the vector concatenation operation, and f emb This represents an embedding layer based on category ID, ultimately yielding a time-fusion representation.

[0145] The topological and semantic spaces will then be fused using a self-attention mechanism. Three learnable weight matrices will be used. Go and learn about the query vector Q, the key vector K, and the value vector V:

[0146] Q = e f W Q (16)

[0147]

[0148] V = g local W V (18)

[0149] In the process of constructing Q, K, and V vectors This is a learnable weight matrix, and its effect is the same as ω mentioned above. sg Both are representation fusion and dimensionality reduction. This is the bias of the linear layer. Represents short sequence S i_j Local location representation in In the construction of the key vector K, the information embodied in the local trajectory hypergraph and the local semantics is fused, the local place features are fused, and the user topological information of the local hypergraph in the value vector V and the time pattern information of the query vector Q and the local place information of the key vector K are calculated to extract the short-term place transfer pattern and the time pattern access intention relationship under the premise of the user's personalized long-term preference. Thus, the topological and semantic space fusion process at the trajectory end is maintained, which is not only a representation but also an explanation of the last step:

[0150]

[0151] d k is a scale parameter with the same value as d ε . This fusion mechanism not only fuses the local information in the topological space and the semantic space to mine the short sequence pattern, but also mines the relationship between the personalized information and the sequence pattern. Finally, the final user-end representation

[0152] S5: Concatenate and X f to obtain the sequence representation The final feature obtained by modeling is input into the neural network V to predict the next check-in time, check-in place, and check-in category.

[0153] The fusion vectors of the trajectory end and the user end are finally concatenated Each place representation of the trajectory data is calculated according to the above method. During training, the place representations are concatenated according to the time sequence of the trajectory, and finally the required sequence representation is obtained

[0154] After obtaining the sequence representation , frequency domain learning is adopted to model the short sequence, reduce the noise of the short sequence, and learn the transfer pattern. The modeling of the sequence representation can be expressed as:

[0155]

[0156] where F is the modeled feature, ω complex is a learnable weight matrix, which is different from other deep learning learnable weight parameters. The weight parameter of ω complex is a complex number, which enables the network to learn and extract patterns in the frequency domain. For the modeled sequence representation, a simple residual network is used to encode the model, which can avoid the gradient problem. This method has high representation complexity and needs to improve the generalization ability of the model:

[0157]

[0158]

[0159] where ω r1 and ω r2 are learnable weight vectors, using GELU Gaussian Error Linear Unit as its activation function. is the encoding representation of the first layer of the neural network, is the final encoding representation of the neural network.

[0160] When the final feature is obtained after modeling, the next interest point, its time and its category will be predicted by the representation. This method fully utilizes the time and category information, and in the final prediction task, it takes the form of multi-task learning, allowing the model to predict the next interest point from multiple angles, capturing the multi-dimensional information of the user. This method also takes into account that the user's discrete check-in locations do not represent the user's continuous trajectory. By predicting the time and category, it can mine the trajectory patterns represented by different time intervals, and learn the continuity and periodicity of the trajectory through the interval of time:

[0161]

[0162]

[0163]

[0164] where ω poi , ω time , ω poi are learnable parameters, b p , b t , b c are their corresponding biases. and represent the predicted check-in location, the predicted check-in time and the predicted check-in category, respectively.

[0165] S6: Freeze the parameters of the pre-trained model, calculate the total loss, and update all neural network parameters according to the total loss until the total loss no longer decreases, at which point all neural network parameters are optimal.

[0166] They map the three predicted values to the space of each prediction target to complete the final recommendation task. The prediction loss of the next interest point and its category is calculated using cross-entropy, and the time prediction loss is calculated using the root mean square loss. Since the time is processed through standardization, the final total loss is multiplied by ten to make it not be ignored. According to the prediction loss The total loss function obtained is:

[0167]

[0168] The proposed method fuses the topological space and semantic space, and predicts the interest points from the global and local perspective, and multi-angle mining user preferences.

[0169] S7: The complete long trajectory of a user is processed according to the method of S1-S4 to obtain the sequence representation of the user, during which the parameter optimal neural network is used, the sequence representation of the user is input into the parameter optimal neural network V, and the next check-in location of the user is output.

[0170] Experimental design and analysis

[0171] 1. Experimental data set

[0172] The experiment selects the Tokyo (TKY) and New York (NYC) check-in data sets of the Foursquare platform, and the collection time spans from April 2012 to February 2013. Before model training, the data set is subjected to key preprocessing steps: the check-in records of less than 10 places and the users with check-in times less than 5 times are removed. Further, the check-in point data of the user is arranged in time sequence, and the sequence is cut at 24-hour intervals. After extracting multiple subsequences of the user, all the subsequences are randomly sorted, and the sub-data set is divided into 80% training set, 10% validation set and 10% test set. The detailed statistical information of the data set is shown in Table 1.

[0173] Table 1 Details of New York and Tokyo data sets

[0174]

[0175] 2. Evaluation index

[0176] The accuracy (Acc@k) and the average reciprocal rank (MRR) are selected to evaluate the interest point recommendation task. The accuracy (Acc@k) measures the proportion of the top k interest points predicted by the model containing the true value, while the average reciprocal rank (MRR) focuses on the ranking of the true interest point in the predicted list. It calculates the average value of the reciprocal rank of each user's true interest point to evaluate the model. Let n represent the total number of predicted samples, that is, the number of data set trajectories in this task, the top-k accuracy and the average reciprocal rank can be expressed as:

[0177]

[0178]

[0179] where Γ(l true ,E topk ) is the indicator function, l trueE topk represents the top k predicted values of the model. The indicator function is 1 if the true value exists in the top k predicted values, otherwise it is equal to 0. The value of Acc@k ranges from 0 to 1, and the higher the value, the more accurate the model prediction. Different k values are set to evaluate the performance of the model under different recommendation list lengths. Higher accuracy means that the model can more accurately recommend locations to meet the needs of users, and users receive more interesting location push. The average reciprocal rank can reflect whether the user is really interested in the location in the front part of the list, and show the efficiency of the user to find the required information. The two dimensions measure the recommended results of the model comprehensively.

[0180] 3. Experimental setup

[0181] In the experiment, NVIDIA RTX 3060 GPU is used as the hardware platform, and development and testing are carried out based on the PyTorch deep learning framework. In the model training process, the Adam optimizer is used, the learning rate is set to 0.0005, and the weight decay is 7e-4. The dropout rate of the attention module is set to 0.5, and the dropout rate of all other modules is 0.3. The convolutional network layer of the hypergraph is 3, the embedding dimension is d τ =128. The time and space embedding dimension is d ε =32, the user embedding dimension is 128, and finally the sequence recommendation representation dimension ζ=256. The time mask sequence enhancement module hyperparameter is set to 5.

[0182] In the semantic enhancement module, BERT, GPT, RoBERTa, XLNeT, SentenceBERT, BART and BART large pre-trained language models are used for semantic encoding, and the best model BART large is selected as the final semantic enhancement module. Similarly, different pre-trained language models are compared for their encoding ability of user behavior semantics. To ensure the consistency and comparability of the experimental results, the same implementation details are used on two data sets, the batch size is set to 20, the iteration number is 100, and the model performance of each epoch is recorded during training, and the best record is selected as the result.

[0183] 4. Comparative experiment

[0184] Different types of baseline models are introduced, and the performance of the model HyperSE (Global and Local Hypergraph with Semantic Enhancement) proposed by the present application is compared with the baseline models.

[0185] ST-RNN: This model is based on a recurrent neural network, adds time and distance matrices according to the time interval and spatial interval between place transfers, and mines space-time rules to further capture user place transfer patterns.

[0186] SASRec: A powerful sequence recommendation baseline model based on an attention mechanism.

[0187] SGRec: This model uses a graph neural network and proposes a Seq2Graph, a sequence-to-graph enhancement method, to effectively learn sequence information through graph structure

[0188] STAN: This model is based on a self-attention mechanism, uses a time-space matrix and a bidirectional attention framework to aggregate all sign-ins in the trajectory.

[0189] DRAN: This model learns decoupled representations of places using distance and transfer as two influencing factors, and combines a soft attention mechanism with space-time information.

[0190] DSMR: This model constructs prompt words from user short-term trajectory space-time category text and performs semantic mining through a pre-training model, and performs point-of-interest sequence recommendation based on a transformer encoder.

[0191] MSTHN

[68] : This model uses a hypergraph neural network to capture high-order collaborative signals of places in a global view, and introduces a user time preference enhancement strategy.

[0192] The comparison experiment results of the point-of-interest recommendation model HyperSE proposed by the application with the above-mentioned models are shown in Table 2 and Table 3. According to the performance comparison, the following conclusions are drawn: HyperSE is better than other baseline models on each index of the two real data sets, and is significantly better than other models on the Tokyo data set, directly showing the effectiveness of the model algorithm.

[0193] Table 2 Model comparison experiment on Tokyo data set

[0194]

[0195] Table 3 Model comparison experiment on New York data set

[0196]

[0197] In the comparative experiment, it can be seen that the ability of sequence modeling directly affects the recommendation effect, but the graph-based method SGRec, DRAN is better than the model relying only on sequence modeling on all data sets. The MSTHN model using hypergraph for point of interest recommendation achieves excellent ACC@1 and ACC@5 scores, which effectively helps the model predict with higher accuracy by revealing the complex high-order collaborative signal of user group trajectories through hypergraph. Hypergraph can help the model to fuse multi-user information from a global perspective to help the prediction of location transfer behavior under the mining of group wisdom information, but MSTHN only uses hypergraph and ignores the role of sequence modeling, although it can achieve advantages in ACC@1 and ACC@5, but it ignores the sequence of user check-in, so its performance in ACC@10 and ACC@20 is low. And DSMR uses transformer structure for sequence modeling and uses semantic information for user representation enhancement, but it only relies on semantic information fusion recommendation and does not use the excellent ability of topological structure to model user preferences, although it achieves high ACC@10 and ACC@20 scores, but its performance in ACC@1 and ACC@5 is poor. And the model only performs semantic modeling of short-term behavior for user trajectories, plus it uses complex prompt engineering to make the small parameter pre-trained language model perform predictive semantic mining instead of encoding, which to some extent exceeds the semantic ability of the language model used.

[0198] HyperSE performs semantic enhancement based on hypergraph neural network and fuses multi-dimensional information globally and locally, and obtains significant improvement in each evaluation index. And for the more sparse and low-quality data set of Tokyo, the improvement of this method is more significant. From the performance of DSMR on the Tokyo data set, it can be seen that semantic information can effectively use natural language semantics to alleviate noise in low-quality data, and use semantic knowledge to deeply mine user preferences, which is significantly better than other models that use shallow context semantic mining. And using hypergraph neural network to capture group wisdom behavior information can effectively model different regions and different ethnic groups, and can achieve excellent performance on the New York and Tokyo data sets. Therefore, the hypergraph neural network of this chapter can effectively capture user group wisdom behavior information from a global and local perspective, and optimize short sequence modeling and alleviate noise in low-quality data through semantic enhancement.

[0199] For the semantic enhancement module, the application also performs the above comparative test, applies a plurality of pre-trained language models, constructs a strategy through a prompt word, takes the average pooling of the output of the last hidden layer of the model as a semantic enhancement representation. In order to measure the encoding ability of different models for user behavior semantic information. The model selection is: GPT, XLNeT, BERT, SentenceBERT, RoBERTa, BART and BART large. Among them, GPT and XLNeT are decoding pre-training language models, and the rest are encoding models. Among them, SentenceBERT is a document-oriented search language model, which has stronger coding ability for sentences, so the embedding vector of [CLS] is taken as the semantic enhanced representation for this model. In terms of parameter quantity, GPT, XLNeT, BERT, SentenceBERT have a parameter quantity of 110 million, RoBERTa has a parameter quantity of 125 million, BART has a parameter quantity of 139 million, and BART large has the most parameter quantity of 406 million. The performance of the HyperSE model enhanced by the above language models is shown in Table 4.

[0200] Table 4 is a comparison table of the influence of language models on recommendation performance

[0201]

[0202]

[0203] The comparison results on the New York dataset show that different pre-training language models have different abilities to extract user behavior semantics, resulting in differences in model recommendation effects. Among them, the effect of the encoding language model is generally higher than that of the decoding model. Since the encoding model is based on the transformer encoder architecture, it can bidirectionally learn comprehensive user behavior semantic context information and capture semantic relationships in user behavior. In contrast, decoding models such as GPT are based on the transformer decoder architecture. This type of autoregressive language modeling mainly focuses on generation ability, and the ability to encode semantic information is relatively weak. Secondly, models with larger parameter quantities can better enhance semantics, and BART large has the best overall performance. Its higher parameter quantity reflects more rich semantic knowledge, and it can also better capture low-frequency patterns and rich long-tail entity information, enabling semantic representations to be more expressive.

[0204] 5. Parameter experiment

[0205] In order to study the influence of different parameter settings on the recommendation effect of the model, first of all, a comparison experiment was conducted on the number of layers of the hypergraph convolution network. The experimental results are shown in Figure 3

[0206] ​Optimal performance. The number of layers is crucial for information aggregation and representation learning on hypergraphs. Increasing the number of hypergraph convolution layers can improve the model's understanding and characterization of user behavior to some extent, and also capture the high-order collaborative signals that group intelligence behavior brings to the representation of places. More layers mean that the model can consider more extensive global and local behavior information of users, capture more similar behaviors between users, and thus obtain more comprehensive and detailed user and place representations. However, increasing the number of layers to a certain number may cause over-smoothing problems, and the representation of places may be over-averaged to become similar, losing distinguishability and information, resulting in a decline in model performance. Therefore, the number of hypergraph convolution layers is set to 3 in the present application, which allows the model to better learn the high-order interaction patterns between users and between users and places by reasonable setting.

[0207] Through hyperparameter experiments on category loss weights and time loss weights, the present application finds that adjusting these weights has a significant impact on the prediction performance of the model. The loss function of the model can not only mine the sequential nature of user trajectories by predicting place, time and place category information, but also supervise the training of time and category context information, allowing the model to learn more general and robust representations rather than focusing too much on specific patterns of a single task. Through multi-task learning, the potential features or associations between the three types of information are mined, and these shared information is captured to improve overall performance.

[0208] Table 5 Category loss weight hyperparameter experiment

[0209]

[0210] Referring to Tables 5 and 6, the experimental results show that appropriate setting of time and category loss weights helps the model to better balance the learning of context information, thereby improving the performance of the model. However, too high or too low weight settings will cause the performance of the model to decline. Because this may cause the model to pay too much attention to a particular information, while ignoring the prediction of other information. For example, when the time loss weight is 100, the model pays too much attention to predicting the time of the next visit place, ignoring the category and sequence information, which greatly damages the performance of the model. Therefore, the default weight of the model is set to: the category loss weight is 1, and the time loss weight is 10 through experiments.

[0211] 6. Ablation experiment

[0212] A number of ablation experiments were conducted to evaluate the effectiveness of each component proposed in HyperSE. The ablation experiments took HyperSE as the base model and recorded its performance. Then in subsequent ablation experiments, specific components were removed: global, local, and overall module ablation experiments were conducted for the semantic enhancement and hypergraph neural network parts, respectively, while replacing the simple transformer encoder structure as the sequence modeling overall module for sequence modeling ablation, and finally the time and category loss functions in the loss function were removed and experiments were conducted. The experimental results are shown in Table 7:

[0213] Table 7 Ablation experiments

[0214]

[0215]

[0216] For the semantic enhancement module, removing this part will cause the user to fail to access the deep semantic pattern extraction in the sequence, resulting in a decline in model performance. This also proves that the semantic enhancement module can effectively model user behavior semantics in a low semantic environment. Comparing the global and local semantic ablation experiments, removing the local semantics leads to a more serious performance decline, but the ACC@20 value is relatively high. This phenomenon shows that user short-term behavior semantics can effectively help prediction, but too little semantic information leads to neglect of specificity, increasing overfitting to specific short-term semantic patterns. While global semantics have more generalization in user behavior semantic feature extraction through complete user trajectories, but there is a slight sacrifice in accuracy.

[0217] For the hypergraph neural network module, removing this part will cause the user group behavior and continuous location access pattern learning to fail, resulting in a decline in model performance. In contrast to the semantic module, removing the global hypergraph will lead to a more serious performance decline. Therefore, the overall behavior preference of the user group is more conducive to model prediction, and the reason may be that the model hypergraph structure has limited ability to extract location transfer patterns in the user's relatively continuous short trajectory information. While short-term behavior preference can effectively help the model, it is not as good as the global information of the user group.

[0218] Removing the sequence modeling will cause a serious decline in ACC@10 and ACC@20, but the ACC@1 decline is not obvious, confirming the importance of sequence modeling and its impact on the recommendation results as described in the previous section. Therefore, frequency domain learning of relatively continuous short trajectories can effectively model user behavior and filter out certain noise. For the category and time loss, removing the category and time prediction indicators will suffer a certain degree of decline, so the multi-task learning approach based on context prediction can help the model capture user preferences in more detail.

[0219] Finally, it is to be explained that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the purpose and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A semantic-enhanced place recommendation method, characterized in that: Comprising the following steps: S1: Obtain the complete long trajectory of multiple users arranged according to their check-in time sequence to form a user trajectory sequence set. Then, using the complete long trajectory as the hyperedge and the check-in location as the node, a global hypergraph is constructed. , No. The first user's Check-in locations Corresponding to a unique tuple, which includes latitude. ,longitude Sign-in time Location Category Location category number and check-in point number; Each complete long trajectory is defined with a time threshold. The truncation yields multiple corresponding short trajectories. Let the first one be... The first user's The short trajectory is ,in Given the length of the short trajectory, construct a local hypergraph by using all short trajectories as hyperedges and check-in locations as nodes. All short trajectories constitute a short sequence set. ; In the S1 The construction process is: Incidence matrix representing hypergraph topology The construction method is as follows: (1) Where e is a hyperedge, v is a node, , is the set of hyperedges and nodes of the global hypergraph, the degree of a hyperedge is the number of nodes connected by the hyperedge, which represents the number of places covered by the user's trajectory, so its formula is , the degrees of all hyperedges constitute the diagonal hyperedge degree matrix ; Assign a weight to each hyperedge, and all hyperedge weights together constitute a diagonal matrix ; Where each node , the degree of the node is obtained by multiplying the diagonal matrix and the matrix of hyperedge weights , and the degrees of all nodes constitute the diagonal node degree matrix ; The S1 in The construction process is as follows: the local hypergraph The incidence matrix of the local hypergraph The definition is the same as formula (1), and the same way as the global hypergraph is adopted to assign weights to each hyperedge to form a diagonal matrix , and a diagonal hyperedge degree matrix and a diagonal node degree matrix are constructed through the diagonal matrix; S2: obtaining the global user representation and the global location representation by inputting the hypergraph convolution network, inputting the hypergraph convolution network to obtain the local user representation and the local location representation, inputting the hypergraph convolution network to obtain the local user representation and the local location representation, In the S2, the steps of obtaining the global user representation and the global location representation and the local user representation and the local location representation are: By embedding the matrix , the check-in location sequence with length is embedded into a feature space with dimension , and the location embedding input to the first layer of the hypergraph neural network is The operation mechanism of the global hypergraph convolution network is as follows: (2) wherein For the global hypergraph convolution network, the first layer is a node embedding layer The details of the two-stage information transmission and aggregation are as follows: the process of the first-stage place representation converging to the hyperedge is represented by After the global trajectory information of the hyperedge user is aggregated, the place representation node is transmitted back by The global place representation is updated using the global information, wherein is a normalization operation; The steps of inputting the hypergraph convolution network to obtain the local user representation and the local location representation are as follows: The steps of inputting the hypergraph convolution network to obtain the local user representation and the local location representation are as follows: The operation mechanism of the local hypergraph convolution network is as follows: (3) wherein is the set of location embedding vectors for the first layer of the local hypergraph convolutional network, the details of the two-stage information passing aggregation are the same as for the global hypergraph convolutional network, is the set of location embedding vectors for the first layer of the local hypergraph convolutional network, the details of the two-stage information passing aggregation are as follows: the process of the first stage node representation converging to the hyperedge is represented by , and after the local trajectory information of the hyperedge users is aggregated, the local information is passed back to the location representation nodes by , which updates the local location representation with the local information, wherein is a normalization operation; By The global place representation set aggregated by global information and the local place representation set aggregated by local information are obtained through the hypergraph convolution calculation of the layer ;​ from Obtain user representation through The global user representation is obtained by averaging the representations of all locations on the hyperedge. It is represented as: (4) User representation is obtained from User representation is obtained from Local user representation is obtained by accumulating and averaging all location representations on the hyper-edge which is represented as: (5)； S3: The semantic information in the long trajectory is constructed into a prompt word, and then a pre-trained language model is used to encode the constructed prompt word to obtain a global user behavior semantic representation. The semantic information in the short trajectory is constructed into a prompt word, and then a pre-trained language model is used to encode the constructed prompt word to obtain a local user behavior semantic representation; The location category is one-hot encoded to obtain As for the check-in time of the user, 24 hours of a day are divided into a plurality of time stamps, and the check-in time is projected to each time stamp to obtain a time embedding ; Using existing methods to and perform embedding results in corresponding temporal characterizations and categorical characterizations ; S4: The global user behavior semantic representation and the local user behavior semantic representation are dimensionally reduced by the neural network I to obtain corresponding global semantic representations and local semantic representations; The global user representation, the global semantic representation, and the user number embedding are spliced into a neural network II for dimension reduction, and then spliced into a fusion network with the global location representation to obtain a final trajectory end representation ; The local semantic representation and the local location representation are spliced into a neural network III to obtain a matrix K, the time representation and the category representation are spliced into a neural network IV to obtain a matrix Q, and the local user representation is a matrix V, K, Q and V are processed by linear layers respectively, and then enter an attention mechanism to obtain a final user terminal representation ; S5: obtaining the sequence representation by concatenating and obtaining the sequence representation by concatenating , according to the obtained performing sequence modeling, and inputting the final features obtained by modeling into a neural network V to predict the next check-in time, check-in location and check-in category; S6: The parameters of the pre-trained model are frozen, the total loss is calculated, and all neural network parameters are updated according to the total loss by back propagation until the total loss no longer decreases, at which time all neural network parameters are optimal; S7: The complete long trajectory of a user is processed according to the method of S1-S4 to obtain the sequence representation of the user, during which the sequence representation of the user is input into the neural network V with optimal parameters, and the next check-in location of the user is output.

2. The method of claim 1, wherein: The steps of obtaining the global user behavior semantic representation and the local user behavior semantic representation in the S3 are: Firstly, the construction method of the global cue is to construct the global place category sequence according to the trajectory of the i th user The category text of each place in the trajectory constructs the global place category sequence On the basis of , the redundant categories are filtered out, so that the category sequence is a unique value sequence; after obtaining the category sequence of the user, it is spliced with the two parts, first splicing the sentence "This user visited" before the category sequence to describe the encoding task, and then splicing the special symbol at the beginning of the sentence to facilitate the extraction of the sentence vector, in formula (6), the splicing of the text is represented by a semicolon, the splicing of the text is represented by a semicolon, and the global cue representing the global semantic signal is represented as: (6) When constructing the local cue word, the last element of the constructed place category sequence is removed, and the redundant categories are filtered out based on the local place category sequence, so that the local category sequence is a unique value sequence. Then, the sentence "This user recently visited" is concatenated in front of the local category sequence, and the special symbol is concatenated in front of the local cue word representing the local semantic signal at the beginning of the sentence. (7) The global language extraction model encodes the global prompt word to obtain the global user behavior semantic representation, and the global language extraction model encodes the local prompt word to obtain the local user behavior semantic representation: (8) (9) in , For the final hidden layer of the global language extraction model and the local language extraction model, the relationship between... The symbols correspond to global user behavior semantic representation and local user behavior semantic representation.

3. The method of claim 2, wherein: The final trajectory end characterization in S4 is obtained The steps are: using equation (10) to the global user behavior semantic representation dimensioning: (10) wherein is a learnable weight matrix, is a bias for the linear layer, and an activation function; (11) in This represents the operation of concatenating vectors. To create a learnable weight matrix, the concatenated weights will be... dimensional vector fusion End-user representation of dimensionality In; through short sequences Global location representation in By integrating the data, a preliminary representation of the final trajectory from the user's perspective is constructed: (12) wherein, is a global place representation Z g corresponding place representation is selected from the global place representation, is a learnable weight matrix.

4. The method of claim 3, wherein: The S4 gives the final user terminal characterization The steps are: Using three learnable weight matrices , , Unlearning query vectors , key vectors and value vectors : (16) (17) (18) wherein, in building , , vectors, is a learnable weight matrix, is a bias for the linear layer, denotes a local place representation in a short sequence ; (19) wherein is a scale parameter having the same value as is a scale parameter having the same value as 5. The method of claim 4, wherein: The S5 in the method The final features obtained by the sequence modeling are input into the neural network V to predict the next check-in time, check-in location and check-in category. Modelling of the sequence representation reveals that: (20) wherein F is a modeled feature, is a learnable weight matrix; (21) (22) wherein, and are learnable weight vectors, using a Gaussian error linear unit as its activation function; is an encoded representation of the first layer of the neural network, is the final encoded representation of the neural network; (23) (24) (25) wherein , , are learnable parameters, , , is its corresponding bias, , and denote the predicted check-in location, the predicted check-in time and the predicted check-in category, respectively.

6. The method of claim 5, wherein: The step of calculating the total loss in the S6 is: According to the interest point prediction loss , the time prediction loss , and the category prediction loss The total loss function is obtained as (26) wherein, and The cross-entropy is taken to compute, The root mean square is used to compute.

Images