Geospatial network decoupling characterization method based on double-cause comparison

Abstract

The invention discloses a geographic space network decoupling characterization method based on double-cause comparison. The method comprises the following steps: acquiring a position and a historical sign-in behavior of a user; extracting user characterization from the historical sign-in behavior sequence of the user; carrying out representation and fusion on the interaction behavior and spatio-temporal information in the interaction behavior by using an embedded layer, and adding position coding information of a sequence; aggregating long and short range behavior representations of the user; separating the coupling characterization into user time behavior characterization and space behavior characterization; characterization decoupling is realized by using a contrast loss implicit mode; quantifying the probability of time transfer from the current sign-in behavior to the next sign-in behavior; modeling the probability of spatial transfer from the current sign-in behavior to the next sign-in behavior by using a global POI transfer graph to realize enhancement of a user-POI matching result; and outputting the predicted people flow data of a certain place at the future moment. According to the method, the uncertainty of time-space transfer of user behaviors is reduced, the problem decision space is reduced, and the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of location-based services, and particularly relates to a geographic space network decoupling characterization method based on double-factor comparison. BACKGROUND

[0002] Human mobility modeling is an important basis for supporting pedestrian trajectory prediction, generation and mining tasks, and is widely used in traffic management and network-based location services. It has guiding significance for discovering group mobility patterns, revealing spatial function area distribution and optimizing traffic and municipal planning decisions. With the development of GPS devices and services and the innovation of Internet business models, mobile phone signaling data and social platform check-in data have also grown. Rich human mobility data provides a good data basis for mining behavior trajectories. Location-based services are a large and dynamic ecosystem, and specific application requirements include navigation, location analysis, urban planning, logistics optimization, social recommendation, etc.

[0003] Unlike group trajectory mining tasks, individual mobility behavior prediction needs to model the diversity and differences of individual behavior and combine spatio-temporal rules and individual differences to generate the final decision. Taking the Point-Of-Interest (POI) recommendation task as an example, the goal is to predict the next visited POI through the user's historical check-in sequence. The essence lies in learning the decision mechanism behind the user's POI interaction behavior in different spatio-temporal scenarios, inferring the user's next possible behavior according to the current spatio-temporal context, and obtaining the recommendation result according to the association between the behavior (user representation) and the POI (POI representation).

[0004] To solve this problem, a series of POI recommendation algorithms mine user preferences and behavior patterns from the interaction history between users and POIs, and explore effective ways to represent POI features, and finally use different matching strategies to recommend POIs to users. Traditional POI recommendation methods use probabilistic models such as Markov chains to model the transition probabilities between POIs, encoding the patterns of user sequential behavior; there are also works based on matrix factorization to decompose the user and POI embeddings from the historical interaction matrix, and to mine the association between users and POIs. This kind of traditional recommendation method regards POI recommendation as a dynamic evolution problem of behavior sequence, and predicts the next POI by modeling the selection and transition mechanism of users on POIs, while adding spatial and temporal prior constraints to model this dynamic mechanism. However, this method mainly considers the regularity of user sequence, and cannot capture the deep representation of the fusion of multiple influencing factors and the high-order pattern of sequence. Considering that the user's check-in behavior is the result of the joint action of multiple influencing factors, such as user preferences, POI spatial distribution, social relationships, and spatio-temporal context, sequence representation methods based on deep learning regard the POI recommendation task as a decision-making problem in a complex coupled space, and fuse multiple sources of information to learn deep representations of users and POIs in this space, and then make recommendations according to the relationship pattern between the two representations. For example, RNN structure and self-attention mechanism are used to construct user behavior sequences, GNN is used to model the spatial representation of POIs and the social representation of users, etc., and finally obtain coupled representations that fuse multiple sources of information. However, this multi-factor coupled representation has the following two problems:

[0005] Weak interpretability: the coupled representation obtained after the fusion of multiple factors cannot be aligned to the influencing factors with clear semantics, making it difficult to provide explanations for the recommendation results. This kind of model learns the coupled representation space by fitting the complex mapping between input features and output results, which encodes the fused representation of multiple cause variables in this space, and cannot explain the specific semantics of each dimension, nor can it quantify the causal effect of each cause on the result.

[0006] Poor generalization: the decision-making mechanism based on coupled representation lacks constraints on individual cause variables, and therefore tends to capture the specific distribution of cause variables in the training data and rely on the learned distribution for decision-making. When the distribution of Z is observed, the originally independent X and Y will exhibit pseudo-correlation in the collider structure. If P train (X,Y|Z)≠P test (X,Y|Z), the model cannot generalize to the test set.

[0007] In general, the current deep representation learning-based methods are the best in the POI recommendation field. In addition to the different representation model structures adopted, the important difference between different methods lies in the understanding of the POI recommendation task, that is, which information needs to be considered, how to obtain the representation of the user and the POI, and how to make a decision on the future interaction between the user and the POI. The hypothesis proposed in the present application is that the user representation should be decoupled into spatial behavior and temporal behavior, and can be implicitly aligned to the location and function representation of the POI, respectively. Therefore, how to learn the decoupled representation of the user and search for the most matched POI combined with the spatio-temporal representation of the next behavior is a key problem to be solved and modeled in reality. SUMMARY

[0008] In order to depict the observation time uncertainty of the geographic spatial network data, the present application is inspired by the uncertainty of human mobility behavior, takes the time uncertainty of individual check-in behavior as the support of the hidden variable time transition mode uncertainty in the indefinite time causality, and selects the interest point prediction as the task scenario. In order to more accurately capture the state transition mode of the hidden variable on the time axis, the present application proposes a causality representation learning method considering the time transition uncertainty, models the transition mode of the hidden variable over time as a causal variable, and separates the information describing the variable from the observation data by using decoupled learning, and finally obtains the representation depicting the time transition mode of the hidden variable. The present application evaluates and compares the representation and prediction ability of the model on two real-world user check-in data sets, and provides further explanation for the proposed strategy through a series of analysis experiments.

[0009] In order to learn the decoupled representation satisfying the above two properties, the present application learns from the positive and negative sample comparison strategy in contrastive learning, takes the decoupled representation obtained from the user historical behavior sequence as the anchor sample, takes the POI representation to be aligned as the positive sample, and takes the decoupled representation to be separated as the negative sample, and constructs a comparison loss to guide the model to learn the decoupled representation with separation and alignment.

[0010] Based on the above two strategies, the present application proposes a BCCDR method based on double-factor comparison for causality decoupled representation, which is used to extract the decoupled user representation from the historical behavior sequence, and further narrows down the decision space of the model by taking the spatio-temporal transition mode of the user behavior as a constraint.

[0011] To achieve the above purpose, the decoupled representation method of geographic spatial network based on double-factor comparison disclosed in the present application comprises the following steps:

[0012] Obtaining the location and historical check-in behavior of the user;

[0013] A Transformer-based behavior sequence encoder is used to extract user representations in user-POI coupling scenarios from historical check-in behavior sequences. An embedding layer is used to represent and fuse interaction behaviors and their spatiotemporal information, while also incorporating positional encoding information from the sequence. A Transformer backbone network is used to aggregate user representations of long- and short-range behaviors.

[0014] The coupled representation is separated into the user's temporal behavior and spatial behavior representations. The category representation and location representation of the target POI are used to guide the decoupled temporal behavior representation and spatial behavior representation to move closer to the POI's function and location representation, respectively. Contrastive loss is used to guide the reduction of similarity between decoupled representations and the enhancement of similarity between aligned representations, implicitly achieving representation decoupling.

[0015] The decoupled temporal behavior is used as the category constraint for the next POI, quantifying the probability of temporal transition from the current check-in behavior to the next one. The global POI transition graph is used to model the probability of spatial transition from the current check-in behavior to the next one, enhancing the user-POI matching results.

[0016] Output the predicted flow of people at a certain place at a certain moment in the future.

[0017] Preferably, the Transformer-based behavior sequence encoder includes a check-in word embedding layer, a Transformer encoder, and a final word representation, which process-wise completes three tasks: representing a single check-in behavior, mining the long- and short-term associations between check-in behavior sequences, and finally obtaining a global representation of the entire behavior sequence.

[0018] Preferably, a single sign-in behavior in the user sign-in sequence is regarded as a word unit, namely a sign-in word unit; each sign-in word unit is a triplet CI i = <u,p,t i >, where u∈U represents the set of all users; p∈P represents the set of all POIs, and each POI is a four-tuple p=<location,category,popularity> , including the location, category and popularity of the POI; t i =<time_in_day,day_in_week> , including the specific time and date when the check-in occurred;

[0019] In summary, the check-in behavior sequence of user u is represented as Seq u =(CI1,CI2,…,CI n ), n is the size of the check-in behavior sequence, and the POI recommendation task for user u is formalized as P(p|Seq u ,t=t n+1), i.e. predicting the probability of each POI point being chosen by the user as the next visiting point, and recommending the result with the highest probability to the user; according to the triplets composed of check-in tokens, each check-in token is formed by the embedding fusion of user, POI and time;

[0020] The user embedding is constructed by using learnable parameters, and the dimension of the user embedding is d u The embedding of the user is represented as follows:

[0021]

[0022] f u (u) is a function for learning the user embedding, and is implemented based on a multi-layer perceptron;

[0023] The number of visits of each POI category within 24 hours is taken as an input feature, and a multi-layer perceptron is used to map to obtain a deep representation, and the category embedding is represented as follows:

[0024]

[0025] pop0, pop1, …, pop 23 are the 0th, 1st and 23rd POI categories respectively, d cat is the dimension of the category embedding, and f cat is a multi-layer perceptron;

[0026] For the location information of the POI, a multi-layer perceptron is used to map the latitude and longitude of the POI to a high-dimensional space with a dimension of d loc , which is used to preserve the spatial information of the POI, and the location embedding is represented as follows:

[0027]

[0028] For the popularity of the POI, the total number of visits constructed from all visit sequences is used as an input to map the POI to a high-dimensional space with a dimension of d pop , which is used to preserve the popularity information of the POI, and the popularity embedding is represented as follows:

[0029]

[0030] f pop is a function for learning the popularity embedding, and is implemented based on a multi-layer perceptron; and popularity is the popularity;

[0031] For each POI, the three types of information are spliced and input to a fusion module to obtain the final embedding d poi is the dimension of the POI embedding, and the fusion process is represented as follows, where || represents feature splicing:

[0032]

[0033] f poi is a fusion module;

[0034] The time embedding layer encodes the specific time of the check-in time in a day and the day index in a week, and the input features are represented by normalized time_in_day and day_in_week respectively; a multilayer perceptron is used to mine deep representations of time information, and the time embedding is represented as follows:

[0035]

[0036] d t is the dimension of the time embedding.

[0037] Preferably, the check-in token embedding layer includes embedding fusion of check-in token triplets and encoding of their positions in the sequence, and combines the two to obtain the final embedding e token input to the representation model; the check-in token embedding layer constructs the input embedding of the i-th check-in token C i in Seq i < u, p, t in > according to the following formula:

[0038]

[0039] d ci is the input dimension of the subsequent encoder, f k is a function for learning check-in token embedding, which is implemented based on a multilayer perceptron;

[0040] Then, a position encoding is added to it according to its position in the entire check-in sequence, which constructs the absolute position of each token in the global sequence while preserving the relative spatial relationship between different tokens, and the representation process is as follows:

[0041]

[0042] j and k are integers, ω in is the angular frequency of the trigonometric function in the position encoding, and i is the i-th check-in token.

[0043] Finally, the embedding of each check-in token is the sum of the input embedding and the position encoding, and the dimension is equal to the input dimension d in of the subsequent encoder, and is represented as follows:

[0044]

[0045] Preferably, the Transformer encoder is used to capture the long-range and short-range associations between different sign-in tokens and perform information aggregation, mapping the sign-in behavior sequence to the dimension d in representation space to obtain a high-level representation of the user behavior sequence; the Transformer encoder includes:

[0046] Transformer encoder is based on the self-attention mechanism of query, key and value subspaces, which embeds the input sequence First, project it to Q, K, V∈R respectively through linear mapping n*D Three subspaces, D is the dimension of the hidden layer of the encoder, and the self-attention calculation is completed through the query operation between Q and K. The calculation process is as follows:

[0047]

[0048] Finally, the representation of each sign-in word in the sequence is obtained through the weighted operation between self-attention and V. The sequence after self-attention transformation of the lth layer is represented by h l ∈R n*D , the transformation process is as follows:

[0049] h l =A·V

[0050] A multi-head attention mechanism is used to capture the dependencies between word units. The operation process is as follows:

[0051]

[0052] Among them, Q,K,V∈R n*D′ ,D′=d / n head ; The transformations within each attention head are independent of each other, and the number of attention heads is n head ;

[0053] The complete Transformer encoding layer consists of attention transformation, residual connection, layer normalization and feedforward network. The operation process is as follows:

[0054]

[0055] LayerNorm is layer normalization, FFN is feedforward network, multi_head is attention transformation, H l―1 is the hidden representation of the previous layer output, is the latent representation output by the attention layer of the lth layer, H l It is the hidden representation output by the lth layer through the feedforward network layer.

[0056] Preferably, the last word representation includes:

[0057] The representation of the last check-in token in the check-in behavior sequence is selected as the proxy of the current behavior representation of the user, and is used to infer the next POI of the user; therefore, the representation of the last token is further represented as the representation U of the current time of the user cur for subsequent prediction; the U cur is a variable representation obtained after behavior sequence encoding, with spatiotemporal dynamic dependence.

[0058] Preferably, the use of contrast loss guides decoupling of representations, reduces similarity between aligned representations, and implicitly realizes decoupling of representations, including:

[0059] First, the coupled representation is mapped to two separate subspaces by two linear projection transformers, obtaining the current temporal behavior and spatial behavior representations of the user, and the transformation process is as follows:

[0060] U t = U cur ·W t

[0061] U s = U cur ·W s

[0062] In order to guide the decoupled representation to adapt to the semantic representation of the POI, the spatial transformation operation involved in the projection is set to

[0063] For the user temporal behavior view, the current temporal behavior representation of the user is taken as the anchor sample, the POI function representation to be aligned is taken as the positive sample, and the spatial behavior representation to be separated is taken as the negative sample; for the user spatial behavior view, the current spatial behavior representation of the user is taken as the anchor sample, the POI location representation to be aligned is taken as the positive sample, and the temporal behavior representation to be separated is taken as the negative sample.

[0064] In summary, the decoupling loss function based on double-factor contrast is constructed as follows:

[0065] loss disen = f con (U t ,p cat ,U s ) + f con (U s ,p loc ,U t )

[0066] f con (anchor, pos, neg) = Softplus(<anchor, neg>―<anchor, pos>)

[0067] anchor is anchor sample, pos is positive sample, neg is negative sample, Softplus is activation function, f con is triplet contrastive loss function, p cat is class representation of POI, p loc is location representation of POI;

[0068] The above formula makes the optimization direction along the loss function decrease, the distance between anchor sample and positive sample gradually decreases, and the distance between anchor sample and negative sample gradually increases.

[0069] Preferably, the decoupled time behavior is used as the class constraint of the next POI, and the probability of the time transition from the current check-in behavior to the next check-in behavior is quantified; the global POI transition graph is used to model the probability of the spatial transition from the current check-in behavior to the next check-in behavior, so as to enhance the user-POI matching result, including:

[0070] A learnable module is used to learn the causal effect proportion of the two decoupled factors and fuse them, and the projection of the next check-in behavior of the user in the POI representation space U nex is obtained according to the following formula:

[0071]

[0072] U nex =σ((U′ s ||U′ t )W fuse +b fuse )

[0073] wherein U′ s and U′ t represent the representations of the user spatial behavior and the time behavior after the decoupling module, σ is an activation function, b fuse is a bias term for calculating the fused representation, and W fuse is a feature transformation term for calculating the fused representation;

[0074] A similarity-based recall matching is used as a downstream prediction head, and a prediction result is generated by a similarity query, and the calculation is as follows:

[0075] y match =U nex ·(e p ) T

[0076] wherein y match ∈R n_poi , and n_poi is the total number of POIs;

[0077] The time transfer prediction is converted into a user intention prediction, and the prediction of the time at which the next check-in behavior occurs is converted into the prediction of the category of the next visited POI. Specifically, the alignment degree between the user time behavior representation and the target POI category is used as a measure of the matching degree, and a matching degree y in the time transfer mode is obtained time ∈R n_poi The calculation process is as follows:

[0078] y time =U t ·(e cat ) T

[0079] In order to capture the global POI transfer mode, a POI space graph in an explicit and implicit perspective is constructed to capture the cross-space transfer mode between POIs, including:

[0080] First, the transfer probability graph G trans between different POIs can be constructed from the training data. The nodes in the graph are a set of all POIs, and the directed edges between POIs are the probabilities calculated from the spatial transfer of all user behaviors. For two POIs p1 and p2, the value corresponding to the edge from p1 to p2 is calculated as follows:

[0081]

[0082] p j is the jth POI, P is the POI, and inalltrajs is all observed trajectory data;

[0083] For p1 and p2, the starting POI is p start , and the transferred POI is p end . They are respectively mapped to a d start -dimensional hidden space through a linear transformation W hidden , and a learnable vector a start , is obtained as the real probability corresponding to the out-node and in-node in the hidden space;

[0084] The transfer relationship captured by the above implicit modeling is further combined with the edges in G trans to calculate the final spatial transfer mode. Specifically, the edge probability described by the normalized Laplacian matrix A lap of the G trans graph is transformed from [0, 1] to [1, 2] by adding a length of 1 offset, and is obtained. Then, the final transfer probability y trans is obtained by bitwise multiplication with the implicitly learned probability, and the calculation process is as follows:

[0085] e implict (p start ,p end )=p start ·W start ·a start +p end ·W end ·a end

[0086]

[0087] p start is the POI at the trajectory starting point, e implict is the implicitly learned probability, is the Laplacian matrix after the shift operation;

[0088] Finally, the matching degree of the transition from the current POI to the next POI is obtained by the following formula:

[0089]

[0090] The probability of the next POI transition of the user is obtained by normalizing the matching degree, and the corresponding prediction loss function is as follows:

[0091]

[0092] y n+1 represents the one-hot vector of the next POI to be predicted, is the predicted next POI probability;

[0093] The overall loss function in the model training process is the combination of the decoupling loss function and the prediction loss function, and is represented as follows:

[0094] loss=loss predict +β*loss disen

[0095] Where β is a hyperparameter, used to control the influence of the decoupling loss in the overall loss function.

[0096] The beneficial effects of the present application are as follows:

[0097] A new POI recommendation task understanding perspective is proposed, which decouples the POI recommendation task into decision problems under the control of user time behavior and spatial behavior respectively, and draws the corresponding causal diagram, which can provide explainability for POI prediction results, and can be migrated to other human movement behavior prediction scenarios.

[0098] A causal decoupling representation based on double-factor comparison is constructed (a method for mining user representation from historical behavior sequence by taking a Transformer as a backbone representation model, and separating the representation under two cause perspectives by a double-factor comparison decoupling module for prediction.

[0099] A POI matching mechanism enhanced by spatiotemporal transfer pattern constraint is constructed, which reduces the uncertainty of user behavior spatiotemporal transfer, narrows down the problem decision space, and improves the prediction accuracy by implicitly capturing the spatiotemporal transfer pattern of user behavior. BRIEF DESCRIPTION OF DRAWINGS

[0100] Figure 1 The overall framework provided by the embodiments of the present application is provided. DETAILED DESCRIPTION

[0101] The present application will be further described below with reference to the drawings, but in no way limits the present application, and any transformation or replacement based on the teaching of the present application belongs to the protection scope of the present application.

[0102] Artificial intelligence is to use digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is the design principle and implementation method of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making.

[0103] The technical solutions provided by the embodiments of the present application relate to machine learning and natural language processing technologies of artificial intelligence, and are specifically introduced and explained as follows.

[0104] Before introducing the embodiments of the present application, some terms related to the present application are explained.

[0105] 1. Traditional recommendation method based on explicit modeling: Traditional POI recommendation methods generally assume that the interaction sequence between users and POIs implies a stable rule of user and POI interaction, so POI recommendation is often regarded as a problem of modeling the explicit sequence rule. This kind of method pays more attention to the transfer probability between adjacent data points in the check-in sequence, or the matching degree between users and POIs captured from the historical sequence, according to the different rule modeling perspectives, which can be divided into methods based on transfer probability modeling and methods based on matrix decomposition.

[0106] The method based on transition probability modeling considers that the transition rules between different POIs can be captured and inferred. Among them, the recommendation method based on Markov chain takes Markov chain as the probability transition mechanism, that is, it is believed that the future state only depends on the current state. This mechanism assumes that the task context of the next POI recommendation is met, but it cannot capture the long-range behavior patterns of users. The method based on matrix factorization assumes that the historical sequence can be decomposed into the implicit embedding of POI and user in the low-dimensional space, which can model the interaction rules between users and POIs.

[0107] 2. Deep learning-based method: Use deep representation models to implicitly learn the interaction patterns between users and POIs in a data-driven manner. In addition, such models provide an effective way to integrate various features such as temporal factors, geographic location factors, and social relationship factors, ultimately obtaining deep and comprehensive representations of users and POIs. According to the different basic model architectures for representing historical sequence data, this type of method can be further divided into sequence representation model-based recommendation methods and graph representation model-based recommendation methods.

[0108] The sequence representation model-based recommendation method can capture long-range and complex dependencies using structures with recurrent hidden units or self-attention mechanisms, and can automatically learn implicit and deep feature associations compared to traditional recommendation models. The graph representation model-based recommendation method enhances the information fusion of POIs in the geographic space and the information fusion of users in the social space from the network perspective. Among them, the graph representation of POI is constructed by establishing a POI space adjacency graph or a POI transition graph, and using graph embedding and graph convolution network and other graph representation models to construct the representation of POI.

[0109] 3. Indeterminate temporal causality: Its characteristic is that the time interval of the observed geographic space network in the process of time evolution is not fixed. This uncertainty of observation time increases the difficulty of learning the temporal causal mechanism between hidden variables. The reason is that the information describing the state of the hidden variable comes from discrete observations at different time stamps. The uncertainty of observation time further increases the uncertainty of the state of the hidden variable, thus introducing uncertainty into the process of modeling the causal relationship between the variables controlled by the causal relationship. In order to solve this problem, it is necessary to identify the transition pattern of the hidden variable state on the time axis, so as to establish deterministic information for the discrete observation time stamp as the basis for subsequent modeling of the causal relationship between hidden variables. In the collision structure of the causal graph, X and Y are cause variables, and Z is a result variable. Taking the causal graph of actor fame as an example, the two cause variables of appearance and acting are originally independent, but when the actor's fame is taken as a condition, the appearance and acting are inversely proportional, and the two variables show a spurious association. The parameters of f(appearance, acting) = actor fame cannot be transferred between the high-appearance actor group and the group of actors from drama schools.

[0110] To solve the above problems, the application proposes a cause and effect decoupling representation method based on double factor comparison, which is used for learning user historical check-in sequence representation and predicting the next POI, mainly including the following two strategies:

[0111] (1) Check-in behavior decoupling under the perspective of double factors

[0112] The application interprets the user check-in behavior as the interaction between the user and the POI, and divides such behavior into time behavior and space behavior. The interaction between the user and the POI can be decomposed into the decision result controlled by the user's time behavior and space behavior. For example, a user visits a restaurant at 12 noon, which is caused by the high matching degree between the user's intention (dining) and the POI function (providing meals) at this time point; and the user chooses a supermarket near home at night, which is more controlled by the high matching degree between the user's site selection (nearby) and the POI location (adjacent). This perspective of decoupling based on time and space behavior double factors not only can provide an explanation for the recommendation result based on the matching degree between the user and the POI, but also can avoid capturing the pseudo correlation between the user's time behavior and space behavior when observing the user-POI interaction result.

[0113] In addition, this decision considering the user's time behavior can reduce the state transition uncertainty of the user's check-in behavior on the time axis, narrow the decision space of the model, and reduce the prediction difficulty. For example, modeling the time interval of the user's check-in behavior can provide important information for understanding the user's next visited POI. Taking user A as an example, the time interval of his visiting POI is different on weekdays and weekends, because the schedule is more tense and regular on weekdays, while on weekends it is more inclined to be rich and leisure. Therefore, if the next POI of user A is predicted at 12:00 on weekdays, the user should be recommended to visit the address point earlier; if it is at 12:00 on weekends, it can be assumed that the user can spend more time to complete this visit, and the interval of the next activity will be longer. The application uses a time transition mode as a constraint POI matching mechanism, after obtaining the representation of the user's time behavior, further uses the representation to identify the user's intention at the current time as a proxy for the next behavior time, implicitly captures the time interval of the user's behavior and helps prediction.

[0114] (2) Implicit alignment of representation based on comparison

[0115] The ideal decoupling representation of user time behavior and space behavior needs to meet the following two properties:

[0116] Separation: As two cause variables forming a collision structure, time behavior and space behavior should have sufficient independence to avoid absorbing pseudo correlation conditioned on observed user behavior. Therefore, the learned decoupled representations should be sufficiently separated.

[0117] Alignment: The temporal behavior and spatial behavior of the user should be able to respectively explain the matching degree between the user's intention and the POI function, and the matching degree between the user's location and the POI location. Therefore, the learned decoupled representation should be able to be aligned with the corresponding representation of the POI.

[0118] Reference Figure 1 The BCCDR method proposed in the present application mainly consists of the following three parts:

[0119] (1) The behavior sequence encoder based on Transformer. This module is mainly used to extract the user representation in the user-POI coupling scenario from the user's historical check-in behavior sequence. The representation describes the user's behavior state at the current time, which is an important source of information for subsequent POI recommendation. In order to preserve the multiple features of the user's historical check-in behavior, the embedding layer is used to represent and fuse the interaction behavior and the spatio-temporal information therein, and the position encoding information of the sequence is added. In order to extract the sequence representation that preserves long-range behavior dependency, the Transformer backbone network is selected to simultaneously realize the aggregation of long and short-range behavior representations of the user.

[0120] (2) Representation decoupling based on double-factor comparison. This module is mainly used to realize the decoupling of the multi-factor coupled representation, and the separation and alignment of the user's temporal behavior and spatial behavior representation. The temporal behavior encoder and the spatial behavior encoder are used to separate the coupled representation into two decoupled representations, the category representation and the location representation of the target POI, which are used to guide the temporal behavior representation and the spatial behavior representation after decoupling to be close to the function and location representation of the POI. In order to realize separation and alignment at the same time, a comparison loss is used as an auxiliary to guide the similarity between decoupled representations to decrease and the similarity between aligned representations to increase, thereby implicitly realizing representation decoupling.

[0121] (3) POI matching mechanism enhanced by spatio-temporal transfer pattern constraint. The matching mechanism between the user representation and the POI representation directly affects the recommendation effect of the model. In order to further narrow down the decision space, this module uses the decoupled temporal behavior as the category constraint of the next POI, quantifies the probability of time transfer from the current check-in behavior to the next check-in behavior, and uses the global POI transfer graph to model the probability of spatial transfer from the current check-in behavior to the next check-in behavior, thereby enhancing the user-POI matching result.

[0122] In one embodiment, the behavior sequence encoder based on Transformer proposed in the present application needs to complete three tasks in a process: representing a single check-in behavior, mining long and short-time correlations between check-in behavior sequences, and finally obtaining the global representation of the entire behavior sequence, corresponding Figure 1The ① part in the middle A module. The three tasks are respectively implemented by a check-in word embedding layer, a Transformer encoder, and a last word representation.

[0123] In one embodiment, the check-in word embedding layer disclosed in the application comprises:

[0124] The word is a basic unit formed after the segmentation of a text sequence in the field of natural language processing, and is also a basic structure input into a representation model when representing a sequence. By analogy with the operation in natural language processing, a single check-in behavior in a user check-in sequence can be regarded as a word, referred to as a check-in word. Each check-in word can be regarded as a three-tuple CI i = <u, p, t i >, where u e U represents a set of all users; p e P represents a set of all POIs, each POI can be regarded as a four-tuple p = <lat, lon, category, popularity>, containing the location, category and popularity of the POI; t i = <time_in_day, day_in_week> contains the specific time and date of the check-in behavior. In summary, the check-in behavior sequence of the user u can be represented as Seq u = (CI1, CI2, …, CI n ), and the POI recommendation task for the user u can be formalized as P(p | Seq u , t = t n+1 ), that is, predicting the probability of each POI point being selected by the user as the next access point, and recommending the result with the highest probability to the user. According to the three-tuple composition of the check-in word, each check-in word can be regarded as a fusion of user, POI and time embeddings.

[0125] User embedding layer. Each user has a unique identity that can be used to construct the user's position in the embedding space. The application directly uses learnable parameters to construct user embeddings, and the dimension of the user embedding is d u , and the user embedding is represented by formula 6-1:

[0126]

[0127] POI embedding layer. The POI embedding needs to fuse the category information, spatial information and personalized representation of the POI. For the category information of the POI, since the category of the POI has a significant time distribution characteristic, for example, the peak period of visiting a restaurant is during meal times, and the visiting crowd of a bar is mainly concentrated at midnight. Therefore, the application uses the number of visits of each category of POI in 24 hours as an input feature, and uses a multi-layer perceptron to map to obtain a deep representation, and the dimension of the category embedding is d cat, the category embedding is represented as Formula 2; for the location information of the POI, a multi-layer perceptron is used to map the longitude and latitude of the POI to a high-dimensional space with dimension d loc , for preserving its spatial information, the location embedding is represented as Formula 3; for its popularity, the total number of visits constructed from all visit sequences is used as input to map it to a high-dimensional space with dimension d pop , for preserving its popularity information, the popularity embedding is represented as Formula 4. For each POI, the three types of information are spliced and input into a fusion module to obtain the final embedding The fusion process is represented as Formula 5, where || represents feature splicing.

[0128]

[0129] The time embedding layer. The check-in time has a high correlation with the time behavior of the user, and to a great extent determines the type of POI visited by the user. Considering the periodicity of user behavior on the day and week scales and other rules, the time embedding layer mainly encodes the specific time of day and the day index in a week for the check-in time, and the input features are represented by normalized time_in_day and day_in_week. This application uses a multi-layer perceptron to mine the deep representation of time information, and the dimension of the time embedding is d t , and the time embedding is represented as Formula 6:

[0130]

[0131] The check-in token embedding layer. The check-in token embedding layer includes embedding fusion of the check-in token triplets and position coding of the same in the sequence, and combines the two to obtain the final embedding e token input to the representation model. For the i-th check-in token C u in Seq i = <u, p, t i >, first construct its input embedding according to Formula 7 Then add a position coding to it according to its position in the entire check-in sequence. The position coding can construct the absolute position of each token in the global sequence while preserving the relative spatial relationship between different tokens, and is often constructed through the combination of sine and cosine functions, and the representation process is as shown in Formula 8. Finally, the embedding of each check-in token is the sum of the input embedding and the position coding, and the dimension is equal to the input dimension d in of the subsequent encoder, and the calculation process is as shown in Formula 9.

[0132]

[0133] In one embodiment, the Transformer encoder disclosed in the application includes:

[0134] The Transformer encoder is used to capture the long- and short-range correlations between different check-in tokens and perform information aggregation, mapping the check-in behavior sequence to a dimension d. in The core working mechanism of the Transformer encoder is the self-attention mechanism based on the query, key, and value subspaces, which embeds the input sequence into First, project it to Q, K, V∈R respectively through linear mapping n*D Three subspaces, D is the dimension of the hidden layer of the encoder, and the self-attention calculation is completed through the query operation between Q and K. The calculation process is shown in Equation 10.

[0135]

[0136]

[0137] Finally, the representation of each sign-in word in the sequence is obtained through the weighted operation between self-attention and V. The sequence after self-attention transformation of the lth layer is represented by h l ∈R n*D , the transformation process is as shown in Equation 11.

[0138] h l =A·V (11)

[0139] In order to capture richer dependencies between word units, this application adopts a multi-head attention mechanism. The transformations in each attention head are independent of each other. The number of attention heads is n. head The calculation process is as shown in formula 12, where Q, K, V∈R n*D′ ,D′=d / n head .

[0140]

[0141] A complete Transformer encoding layer consists of attention transformation, residual connection, layer normalization and feedforward network. The operation process is expressed as Equation 13.

[0142]

[0143] In one embodiment, the final word representation disclosed in this application includes:

[0144] The Transformer-based behavior sequence encoder can mine the correlation between different check-in tokens and obtain a representation that integrates the user's long-term and short-term behavior preferences. It can be considered that this representation encodes the comprehensive representation of the user when the last behavior of the behavior sequence occurs, that is, the coupled representation of the user's spatial behavior and temporal behavior. In the task context of the next POI recommendation, this application selects the representation of the last check-in token in the check-in behavior sequence as a proxy for the user's current behavior representation, and uses it to infer the user's next POI. Therefore, the last token representation It is further represented as the user’s current moment representation U cur , used for subsequent predictions. It should be noted that this representation is different from the user embedding e used in the previous article. u , because U cur It is a variable representation obtained after behavioral sequence encoding, with spatiotemporal dynamic dependence, and e u It can be considered as a static representation, which serves as a learnable identity code for subsequent model learning.

[0145] The representation decoupling module based on dual-factor comparison aims to achieve decoupling between the user's temporal behavior and spatial behavior by ensuring the separation and alignment of the two types of representations. The user's current moment behavior representation U output from the behavior sequence encoder is cur It can be considered as the corresponding representation in the representation space of the coupled temporal and spatial behaviors, and can serve as the representation source to be decoupled. In one embodiment, the present application uses the functional and location representations of the POIs corresponding to the task labels as the samples to be aligned for temporal and spatial behaviors, respectively, and sets a contrast loss function to achieve implicit separation and alignment simultaneously.

[0146] First, two linear projection transformers are used to map the coupled representations into two separate subspaces to obtain the representations of the user’s current temporal behavior and spatial behavior. The transformation process is shown in Equation 14. In order to guide the decoupled representation to adapt to the semantic representation of POI, the spatial transformation operations involved in the projector are set to In this process, the present application does not explicitly constrain the degree of separation of the two subspaces, but guides the separation of the two through a subsequent loss function.

[0147] U t =U cur W t

[0148] U s =U cur W s (14)

[0149] Two contrast views are established from the user's temporal behavior representation and spatial behavior representation respectively. For the user temporal behavior view, the current temporal behavior representation of the user is taken as the anchor sample, the POI function representation to be aligned is taken as the positive sample, and the spatial behavior representation to be separated is taken as the negative sample; for the user spatial behavior view, the current spatial behavior representation of the user is taken as the anchor sample, the POI location representation to be aligned is taken as the positive sample, and the temporal behavior representation to be separated is taken as the negative sample. Based on the above, the decoupling loss function based on double-factor contrast is constructed as formula 15. Along the optimization direction of the loss function reduction, the distance between the anchor sample and the positive sample gradually decreases, and the distance between the anchor sample and the negative sample gradually increases.

[0150] loss disen =f con (U t ,p cat ,U s )+f con (U s ,p loc ,U t )

[0151] f con (anchor,pos,neg)=Softplus(<anchor,neg>―<anchor,pos>) (15)

[0152] The decoupling representation based on double-factor contrast can guide the separation of the user's temporal behavior and spatial behavior, so that the model can make decisions according to the preferences of the user's temporal behavior and spatial behavior respectively, and eliminate the false correlation caused by the collision of the two cause variables. Therefore, in one embodiment, the spatiotemporal transfer mode constraint enhanced POI matching mechanism proposed in the present application respectively realizes the POI matching mechanism with the user behavior temporal transfer mode and spatial transfer mode as constraints, which is used to narrow the decision space of the model recommended POI and enhance the accuracy of the final decision. Specifically, the user's temporal behavior representation is used to improve the accuracy of POI category matching, and the user's spatial behavior representation is used to improve the prediction ability of POI transfer in spatial position. Both constraint enhancements are achieved by operating on the POI recommendation probability.

[0153] (1) Similarity-based recall matching

[0154] The decoupling process of the user's temporal behavior and spatial behavior representation realizes the elimination of the double-factor collision effect, and on this basis, the causal effect of each cause variable on the result variable needs to be estimated and fused to obtain the final decision. The present application uses a learnable module to learn the causal effect proportion of the two decoupling factors and fuse them, and according to formula 16, the projection U nex of the user's next check-in behavior in the POI representation space is obtained.s and U' t respectively represent the representation of user space behavior and time behavior after the decoupling module,

[0155] U nex = σ((U' s || U' t )W fuse + b fuse ) (16)

[0156] Therefore, U nex can be considered as an inference of the next behavior of the user based on the current behavior representation of the user, in order to guide the matching degree of the representation and the target representation in the POI representation space, while enhancing the adaptability of the representation model downstream output label and the existing candidate set, the present application directly uses the similarity-based recall matching as the downstream prediction head, and generates the prediction result by similarity query. Since in the embedding layer, a fused representation e p has been generated for each POI in the candidate set, match ∈ R n_poi .

[0157] y match = U nex ·(e p ) T (17)

[0158] (2) Category matching enhancement under time transfer mode constraint

[0159] y match depicts the decision result under the overall causal effect considering the time behavior and the space behavior, and the time at which the next behavior of the user occurs can be further inferred to narrow down the decision space of the model. Time transfer mode refers to inferring the time point at which the next behavior occurs based on the current time behavior of the user. However, it is difficult to infer the specific time point, and the user check-in behavior has local time similarity, that is, the probability of performing a certain check-in behavior in the same time window is similar. Therefore, the present application converts the time transfer prediction into user intention prediction, and converts the time at which the next check-in behavior occurs into the category of the next visited POI. This way not only narrows down the candidate set of downstream output, but also conforms to the flexibility of human travel rules. Since the alignment trend of the user time behavior representation output by the decoupling module and the target POI category can be directly used as the measurement of the matching degree, the matching degree y time ∈ R n_poi under the time transfer mode is obtained, and the calculation process is as shown in formula 18.

[0160] y time = U t ·(e cat ) T (18)

[0161] (3) Location matching enhancement under spatial transition pattern constraints

[0162] Similar to the effect of time transition pattern constraints, the spatial range of the next behavior can also be effectively narrowed down according to the current spatial location of the user. For example, if a user rarely makes long-distance spatial transitions, the probability of the next access being far away from the current spatial location will be lower. In order to capture this global POI transition pattern, the present application constructs a POI spatial graph from two perspectives of explicit and implicit to capture the cross-space transition pattern between POIs.

[0163] First, a transition probability graph G trans = (V, E) between different POIs can be constructed from the training data, where the nodes of the graph are the set of all POIs, and the directed edges between POIs are the probabilities calculated from the spatial transitions of all user behaviors. Taking two POIs p1 and p2 as an example, the value corresponding to the edge from p1 to p2 is calculated as formula 19:

[0164]

[0165] Since the location information of the POI has been fused into its embedding vector, an adaptive transformation can be used to implicitly capture the transition relationship from the POI embedding. For two POIs, let the starting POI be p start , and the transitioned POI be p end , respectively, map them to a hidden space with dimension d hidden through a linear transformation W start , , and then use a learnable vector a start , to obtain the real probability corresponding to the out-node and in-node in the hidden space. The transition relationship captured by the above implicit modeling is further combined with the edges in G trans to calculate the final spatial transition pattern. Specifically, the edge probability described by the normalized Laplacian matrix A trans of G lap is transformed from [0, 1] to [1, 2] by adding an offset of length 1 , and then multiplied by the implicitly learned probability bit by bit to obtain the final transition probability y trans , the calculation process is as formula 20.

[0166] eimplict (p start ,p end )=p start W start ·a start +p end W end ·a end

[0167]

[0168] Finally, the matching degree y of transferring from the current POI to the next behavior POI is obtained by the three parts, as shown in Equation 21. The matching degree can be further normalized to obtain the probability of the POI to which the user will transfer in the next behavior. The corresponding prediction loss function is shown in Equation 22, y n+1 A one-hot vector representing the next POI to be predicted.

[0169]

[0170] The overall loss function during model training is a combination of the decoupling loss function and the prediction loss function, expressed as Equation 23, where β is a hyperparameter used to control the influence of the decoupling loss in the overall loss function.

[0171] loss=loss predict +β*loss disen (twenty three)

[0172] This application conducts experiments on two public benchmark datasets in the POI recommendation task, namely Foursquare-NYC and OpenStreetMap. Foursquare is a well-known user location service website that allows users to check in at different locations through mobile devices such as mobile phones. The Foursquare dataset contains user check-in data from April 12, 2012 to January 16, 2013, of which the user check-in behavior data includes 227,428 check-in records. The original check-in dataset provides anonymized user ID, point of interest ID, point of interest category, point of interest geographic location (latitude and longitude) and check-in time information. OpenStreetMap is the world's largest crowdsourced free open source map data. It contains rich information such as roads, buildings, water systems, points of interest, and land use.

[0173] This application selected the following baseline methods for comparative experiments:

[0174] MF: Matrix factorization is a classic model widely used in recommendation tasks. It decomposes the interaction representation matrix of users and POIs to obtain the latent representation of users and POIs.

[0175] FPMC: Combining matrix factorization and Markov chain to simultaneously model users' long-range preference sequence behavior.

[0176] LSTM: LSTM is a variant model of RNN that processes long-term and short-term memories through memory gates and forget gates. It is suitable for processing long sequence data.

[0177] PRME: A personalized rating matrix embedding is proposed to capture user preferences and behavior transfer.

[0178] ST-RNN: Combines time and distance transfer matrices with RNN to model the temporal and spatial context of user behavior.

[0179] STGN: A gating mechanism for spatiotemporal information is added to the traditional LSTM model to capture users' behavioral preferences in both time and space.

[0180] STGCN: An improved model of STGN, adding input gate and forget gate on its basis.

[0181] PLSPL: It uses attention mechanism and LSTM to capture users’ long-term and short-term preferences respectively, and combines the two to generate recommendation results.

[0182] STAN: A spatiotemporal attention network that uses the attention mechanism to capture and aggregate spatiotemporal dependencies in user trajectory points, while using a negative sample balanced sampling strategy to optimize recommendation results.

[0183] GETNext: Use the Transformer model to model the check-in behavior sequence and use the trajectory flow graph to assist in estimating the POI transition probability.

[0184] STHGCN: constructs a hypergraph to capture high-level information in user trajectories and uses a hypergraph Transformer layer to model user behavior.

[0185] This application primarily uses prediction accuracy as a metric for evaluating prediction results. Considering that real-world applications often recommend a batch of POIs to users, we use the acc@k metric with k set to 1, 5, 10, and 20. acc@k measures the proportion of correct predictions when recommending k results. The calculation formula is shown in Equation 24, where n represents the number of samples to be predicted. Smaller values ​​for k increase the model's recommendation performance, while larger values ​​for acc@k indicate better model performance.

[0186]

[0187] The model is trained using the Adam stochastic gradient descent optimizer with an initial learning rate of 0.0001 and a weight decay of 0.0001, and the learning rate is adjusted according to the negative log-likelihood loss on the validation set using the ReduceLROnPlateau learning rate scheduler provided by PyTorch. The model training process is completed on an NVIDIA_A100 using 1Core GPU and 64GB memory, and the loss function is negative log-likelihood loss. The maximum number of iterations for training the model is 200, and the tolerance in the early stopping strategy based on the classification accuracy of the validation set is set to 50. Users with less than 10 check-ins and POIs with less than 10 visits are removed from the dataset, and the trajectory data is divided into 24-hour intervals, and the trajectory of each user is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The user and check-in time embedding dimensions in BCCDR are 64 dimensions, the category, location, and popularity embedding dimensions of the POI are 64 dimensions, and the POI representation and hidden layer representation dimensions of the model are both 128 dimensions.

[0188] The experimental results of the present application and the baseline model on the two datasets are shown in Table 1.

[0189] Table 1 Performance Experiment Results

[0190]

[0191]

[0192] The indefinite observation of geospatial network data introduces uncertainty in restoring its evolution mechanism, increasing the learning difficulty of geospatial network representation models in time causal mechanism. Therefore, if we want to accurately model the change of geospatial network over time from observed data, we need to capture the state transition pattern of hidden variables on the time axis and further build a representation model considering the time transition pattern of hidden variables. To this end, the present application proposes a decoupling learning strategy to separate the information related to the time transition pattern of hidden variables from the observed data, and use this information to infer its transition trend. Specifically, the present application proposes a causal representation learning method considering the uncertainty of time transition, which decouples the representation of hidden variables into two types of causal representation, time transition and spatial transition, and constructs two contrastive views to realize the separability and alignment of the representation. In addition, the decoupled causal representation can be used to narrow the state space of the time and spatial representation after transition, respectively, as a constraint to enhance the prediction accuracy of the model. The present application evaluates and compares the representation and prediction ability of the model on two real-world user check-in datasets, and the results show that the present method has better recommendation performance and can eliminate the pseudo-correlation caused by the collision structure through decoupled representation. In addition, this representation method based on double-factor decoupling can provide explanation for the task execution process, which is helpful for further analysis of user behavior.

[0193] The word "preferably" is used herein to mean "desirably, advantageously, or optimally" in a particular context. Any aspect or design described as "preferred" herein is not necessarily to be construed as being more advantageous than other aspects or designs. The use of the words "preferably" or "desired" are not meant to limit the aspect or parameter being described but to provide additional design options for the particular aspect or parameter. The term "or" as used in this application refers to a comprehensive -or, i.e., the term "X employs A or B" means that X employs either A or B, or both A and B. That is, the term "or" as used in this application is intended to be equivalent to the term "inclusive-or" as that term is used in the legal field, rather than being limited to the exclusive-or which is equivalent to the phrase "either A or B but not both."

[0194] Also, although the present disclosure has been described with respect to one or more implementations, those skilled in the art will recognize the equivalence of certain features among the different implementations. The disclosure is intended to cover all such modifications and variations as would be appreciated by those skilled in the art. The disclosure includes all accessions and modifications and is limited only by the scope of the following claims. In particular, with respect to the various functions performed by the above-described components (e.g., elements, etc.), the terms used to describe such components are intended to correspond, unless otherwise indicated, to any component which performs the specified function of the component (e.g., is functionally equivalent thereto) even if not structurally equivalent to the disclosed structure performing the function in the exemplary implementations of the present disclosure shown herein. Moreover, although a particular feature of the present disclosure can have been disclosed with respect to only one of several implementations, such feature can be combined with one or other features of the other implementations as can be desired and advantageous for any given or particular application. Furthermore, to the extent that the terms "including", "includes", "having", "has", "containing", or variants thereof are used in either the detailed description or the claims, such terms are intended to be inclusive in a manner similar to the term "comprising" as an open transition term without precluding any additional or other elements.

[0195] The functional units in the embodiments of the present application can be integrated in one processing module, or each unit can be physically present separately, or a plurality of or more units can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. If the integrated module is realized in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. The above-mentioned devices or systems can execute the storage method in the corresponding method embodiments.

[0196] In summary, the above-mentioned embodiments are one embodiment of the present application, but the embodiments of the present application are not limited by the above-mentioned embodiments, and any changes, modifications, substitutions, combinations and simplifications made without departing from the spirit and principle of the present application are equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A geospatial network decoupling representation method based on dual-factor comparison, characterized by: The following steps are involved: Get the user's location and historical check-in behavior; A Transformer-based behavior sequence encoder is used to extract user representations in user-POI coupling scenarios from historical check-in behavior sequences. An embedding layer is used to represent and fuse interaction behaviors and their spatiotemporal information, and positional encoding information is added to the sequence. Aggregate user long- and short-range behavior representations using the Transformer backbone network; The coupled representation is separated into the user's temporal behavior and spatial behavior representations. The category representation and location representation of the target POI are used to guide the decoupled temporal behavior representation and spatial behavior representation to move closer to the function and location representation of the POI, respectively. Using contrast loss as a guide, the similarity between decoupled representations is reduced, and the similarity between aligned representations is enhanced, implicitly achieving representation decoupling; The decoupled temporal behavior is used as the category constraint for the next POI, and the probability of temporal transfer from the current check-in behavior to the next check-in behavior is quantified. Use the global POI transition graph to model the probability of spatial transition from the current check-in behavior to the next check-in behavior, thereby enhancing the user-POI matching results; Output the predicted flow of people at a certain place at a certain moment in the future.

2. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 1 is characterized in that: The Transformer-based behavior sequence encoder includes a check-in word embedding layer, a Transformer encoder, and a final word representation. It completes three tasks in a process-based manner: representing a single check-in behavior, mining the long-term and short-term associations between check-in behavior sequences, and finally obtaining a global representation of the entire behavior sequence.

3. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 2 is characterized in that: A single sign-in behavior in the user sign-in sequence is regarded as a word unit, namely a sign-in word unit; each sign-in word unit is a triplet CI i = <u,p,t i >, where u∈U represents the set of all users; p∈P represents the set of all POIs, and each POI is a four-tuple p=<location,category,popularity> , including the location, category and popularity of the POI; t i =<time_in_day,day_in_week> , including the specific time and date when the check-in occurred; In summary, the check-in behavior sequence of user u is represented as Seq u =(CI1,CI2,…,CI n ), n is the size of the check-in behavior sequence, and the POI recommendation task for user u is formalized as P(p|Seq u ,t=t n+1 ), that is, predicting the probability of each POI being selected as the next visit point by the user and recommending the result with the highest probability to the user; based on the triple composition of the check-in word, each check-in word is formed by the embedding fusion of the user, POI and time; Use learnable parameters to construct user embeddings with a dimension of d u , the user's embedding is represented as follows: f u (u) is the function for learning user embedding, which is implemented based on a multi-layer perceptron; Taking the number of visitors to each POI category within 24 hours as the input feature, a multi-layer perceptron mapping is used to obtain a deep representation. The category embedding is expressed as follows: pop0,pop1,…,pop 23 They are the 0th, 1st and 23rd POI categories, d cat is the dimension of category embedding, f cat is a multilayer perceptron; For the location information of POI, a multi-layer perceptron is used to map the latitude and longitude of POI to the dimension d loc The high-dimensional space is used to retain the spatial information of POI, and the position embedding is expressed as follows: For POI popularity, we use the total number of visitors constructed from all visit sequences as input and map POIs to a dimension of d pop The high-dimensional space is used to retain the popularity information of POI, and the popularity embedding is expressed as follows: f pop It is a function for learning popularity embedding, implemented based on a multi-layer perceptron; popularity is popularity; For each POI, the three types of information are spliced ​​and input into a fusion module to obtain the final embedding d poi is the dimension of POI embedding, and the fusion process is expressed as follows, where || represents feature concatenation: f poi It is a fusion module; The time embedding layer encodes the check-in time as a specific time of day and a day of the week index. The input features are represented by normalized time_in_day and day_in_week, respectively. A multilayer perceptron is used to mine the deep representation of time information. The time embedding is represented as follows: d t is the dimension of time embedding.

4. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 3 is characterized in that: The sign-in word embedding layer includes the embedding fusion of the sign-in word triple and its position encoding in the sequence, and combines the two to obtain the final embedding input to the representation model. token ; The sign-in word embedding layer is for Seq u The i-th sign-in word CI in i = <u,p,t i >, first construct its input embedding according to the following formula d in is the input dimension of the subsequent encoder, f ci It is a function for learning the embedding of sign-in words, implemented based on a multi-layer perceptron; Then, a positional encoding is added to each word based on its position in the entire sign-in sequence. The positional encoding constructs the absolute position of each word in the global sequence while preserving the relative spatial relationship between different words. The representation process is as follows: j and k are integers, ω k is the angular frequency of the trigonometric function in the position encoding, i is the i-th sign-in word; Finally, the embedding of each sign-in word is the sum of the input embedding and the position encoding, and the dimension is equal to the input dimension d of the subsequent encoder. in , which is expressed as follows:

5. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 4 is characterized in that: The Transformer encoder is used to capture the long- and short-range associations between different sign-in tokens and perform information aggregation, mapping the sign-in behavior sequence to the dimension d in representation space to obtain a high-level representation of user behavior sequences; The Transformer encoder includes: Transformer encoder is based on the self-attention mechanism of query, key and value subspaces, which embeds the input sequence First, project it to Q, K, V∈R respectively through linear mapping n*D Three subspaces, D is the dimension of the hidden layer of the encoder, and the self-attention calculation is completed through the query operation between Q and K. The calculation process is as follows: Finally, the representation of each sign-in word in the sequence is obtained through the weighted operation between self-attention and V. The sequence after self-attention transformation of the lth layer is represented by h l ∈R n*D , the transformation process is as follows: h l =A·V A multi-head attention mechanism is used to capture the dependencies between word units. The operation process is as follows: Among them, Q,K,V∈R n*D′ ,D′=D / n head ; The transformations within each attention head are independent of each other, and the number of attention heads is n head ; The complete Transformer encoding layer consists of attention transformation, residual connection, layer normalization and feedforward network. The operation process is as follows: LayerNorm is layer normalization, FFN is feedforward network, multi_head is attention transformation, H l―1 is the hidden representation of the previous layer output, is the latent representation output by the attention layer of the lth layer, H l It is the hidden representation output by the lth layer through the feedforward network layer.

6. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 5 is characterized in that: The final word representation includes: The representation of the last check-in word in the check-in behavior sequence is selected as the proxy for the user's current behavior representation and used to infer the user's next POI; therefore, the last word representation It is further represented as the user’s current moment representation U cur , used for subsequent prediction; the U cur It is a variable representation obtained after behavioral sequence encoding and has spatiotemporal dynamic dependencies.

7. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 6 is characterized in that: The use of contrast loss as a guide to reduce the similarity between decoupled representations and enhance the similarity between aligned representations, implicitly achieving representation decoupling, includes: First, two linear projection transformers are used to map the coupled representations into two separate subspaces to obtain the representations of the user's current temporal and spatial behaviors. The transformation process is as follows: U t =U cur ·W t U s =U cur ·W s Among them, in order to guide the decoupled representation to adapt to the semantic representation of POI, the spatial transformation operations involved in the projector are set to For the user's temporal behavior view, the user's current temporal behavior representation is used as the anchor sample, the POI function representation to be aligned is used as the positive sample, and the spatial behavior representation to be separated is used as the negative sample. For the user's spatial behavior view, the user's current spatial behavior representation is used as the anchor sample, the POI location representation to be aligned is used as the positive sample, and the temporal behavior representation to be separated is used as the negative sample. In summary, the decoupling loss function based on dual-factor comparison is constructed as follows: loss disen =f con (IN t ,p cat ,IN s )+f con (IN s ,p loc ,IN t ) f con (anchor,pos,neg)=Softplus(<anchor,neg>―<anchor,pos>) Anchor is the anchor sample, pos is the positive sample, neg is the negative sample, Softplus is the activation function, f con is the ternary contrast loss function, p cat is the category representation of POI, p loc It is the location representation of POI; The above formula makes the distance between the anchor sample and the positive sample gradually decrease, while the distance between the anchor sample and the negative sample gradually increase along the optimization direction of the loss function.

8. The geospatial network decoupling characterization method based on dual-factor comparison according to claim 7 is characterized in that: The decoupled time behavior is used as the category constraint of the next POI to quantify the probability of time transfer from the current check-in behavior to the next check-in behavior; The global POI transition graph is used to model the probability of spatial transition from the current check-in behavior to the next check-in behavior, thereby enhancing the user-POI matching results, including: A learnable module is used to learn the causal effect ratio of two decoupling factors and perform integration. Combined, the projection U of the user's next check-in behavior in the POI representation space is obtained according to the following formula: nex : U nex =σ((U′ s ||U′ t )W fuse +b fuse ) where U′ s and U′ t They represent the representation of user spatial behavior and temporal behavior after the decoupling module, σ is the activation function, and b fuse is the bias term for calculating the fusion representation, W fuse It is the feature transformation term for calculating the fusion representation; Using similarity-based recall matching as the downstream prediction head, the prediction results are generated by similarity query, which is calculated as follows: yes match =You nex ·(yes p ) T Among them, y match ∈R n_poi , n_poi is the total number of POIs; The time transfer prediction is converted into user intention prediction, and the predicted time of the next check-in behavior is converted into the prediction of the category of the next visited POI. Specifically, the degree of alignment between the user's time behavior representation and the target POI category is used as a measure of matching, and the matching degree y under the time transfer mode is obtained. time ∈R n_poi , the calculation process is as follows: yes time =You t ·(yes cat ) T In order to capture the global POI transfer pattern, a POI spatial graph from both explicit and implicit perspectives is constructed to capture the cross-spatial transfer pattern between POIs, including: First, we can construct a transition probability graph G between different POIs from the training data. trans , the nodes in the graph are the set of all POIs, and the directed edges between POIs are the probabilities calculated from the spatial transfer of all user behaviors; for two POIs p1 and p2, the value corresponding to the edge from p1 to p2 is calculated as follows: p j is the j-th POI, P is the POI, and inalltrajs is all observed trajectory data; For p1 and p2, the starting POI is p start , the POI after transfer is p end , respectively, by linear transformation W start , Mapped to dimension d hidden The latent space of , and then use a learnable vector a start , Obtain the corresponding real number probability of it as an outgoing node and an incoming node in the latent space; The transfer relations captured by the above implicit modeling are further combined with G trans The edges in G are combined to calculate the final spatial transfer mode; specifically, trans Normalized Laplacian matrix A of the graph lap The edge probability is transformed from [0,1] to [1,2] by adding an offset of length 1, and we get Then multiply it by the implicitly learned probability to get the final transition probability y trans , the calculation process is as follows: e implict (p start ,p end )=p start ·W start ·a start +p end ·W end ·a end p start is the POI of the trajectory starting point, e implict is the implicitly learned probability, is the Laplacian matrix after the offset operation; Finally, the matching degree of the next POI is transferred from the current POI to the next POI. It is obtained from the following formula: The probability of the POI to which the user will transfer in the next behavior is obtained by normalizing the matching degree. The corresponding prediction loss function is as follows: y n+1 Represents the one-hot vector of the next POI to be predicted, is the predicted probability of the next POI; The overall loss function during model training is a combination of the decoupling loss function and the prediction loss function, which is expressed as follows: loss=loss predict +β*loss disen Where β is a hyperparameter used to control the influence of decoupling loss in the overall loss function.

Images