Popularity deviation relieving method for place recommendation

Abstract

The invention relates to a place recommendation-oriented popularity deviation relieving method. The method comprises the following steps of selecting a public user track data set; a model AGSRec is constructed; inputting the action track of the user into a model AGSRec to obtain an attraction perception graph feature # imgabs0 # and a semantic enhancement feature # imgabs1 #, and respectively performing feature fusion on the two features and the user features to obtain a user attraction perception graph feature and a user semantic enhancement feature; performing position vector embedding on the user attraction perception graph features and the user semantic enhancement features, and performing self-attention mechanism processing on the user attraction perception graph features and the user semantic enhancement features after position embedding to obtain final unbiased user features hb; inputting the hb into a prediction layer module to obtain a recommendation score set Score; and sorting the scores in the Score in a descending order, and selecting the first P places as a recommendation list and outputting the recommendation list. By using the method disclosed by the invention, the popularity deviation appearing in two levels of dominant behavior pattern and implicit semantic association during POI recommendation can be relieved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a popularity deviation mitigation method for location recommendation. Background Art

[0002] POI recommendation is an intelligent service system that recommends potential locations of interest to users based on their historical behavior, location, and contextual information. Its core function is to optimize user decision-making efficiency by accurately matching user preferences with POI attributes. Although existing recommendation systems have made significant progress in modeling the association between user preferences and locations, the recommendation process still suffers from several inherent issues: location interaction data is generally unevenly distributed, and traditional recommendation algorithms have limitations in representing complex user behavior patterns. These issues make recommendation models highly susceptible to bias during training.

[0003] Biases in recommendations manifest themselves in various forms, with common types including popularity bias, consistency bias, and exposure bias. Popularity bias creates a Matthew effect by reinforcing the over-recommendation of popular content, leading to the continued marginalization of long-tail projects. This not only exacerbates the solidification of user information cocoons, but also causes a lack of diversity and fairness crisis in the recommendation ecosystem, posing a great threat.

[0004] Many studies have proposed solutions to popularity bias. Krishnan et al. proposed an adversarial training strategy, which uses an adversarial network to learn the implicit correlation structure in user feedback data, and combines it with a neural collaborative filtering model to reduce the over-recommendation of popular items, thereby alleviating the long-tail problem; Liu et al. used the text matching technology model TASTE to represent long-tail products through full-text modeling, reducing the preference for popular products; Pan et al. used a multi-source information model that fused single-category implicit access information and explicit keyword weights to effectively eliminate the homogenization tendency in academic writing; Bao et al. proposed a news recommendation method CADNN based on causal graph model and conformity perception, which jointly modeled user preferences, news popularity and user conformity traits, and used causal inference to separate false associations. These studies provide diverse solutions for debiasing recommendation systems.

[0005] While existing research has made some progress in mitigating popularity bias, several challenges remain. First, mainstream debiasing methods rely on confounding variables calculated using statistical methods. This quantified-results approach fails to effectively model the explicit behavioral patterns in users' decision-making processes. Dynamic features inherent in user interactions, such as temporal correlations and differences in interaction depth, are abstracted into static statistical metrics, making it difficult for the model to capture the behavioral logic driving users' actual decisions. This representational bias makes it difficult for debiased recommendations to deviate from users' personalized preferences, resulting in reduced recommendation accuracy. Second, the unstructured semantic information implicit in user behavior data has not been fully explored. For example, complex factors such as the spatiotemporal regularities inherent in user travel trajectories, the potential social connections between high-frequency users, and the ecological characteristics of regional venue distribution are difficult to quantify using traditional statistical methods. Existing debiasing frameworks often simplify this information into discrete features, resulting in incomplete data representation and, consequently, poor recommendation effectiveness after bias removal. Summary of the Invention

[0006] In view of the above problems existing in the prior art, the technical problem to be solved by the present invention is: how to remove popularity deviation at the level of explicit interaction mode and implicit semantic association when recommending places.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A method for alleviating popularity bias in location recommendation includes the following steps:

[0009] S100: Select a public user trajectory dataset, which includes a user set U and the action trajectory of each user. The expression of U is as follows:

[0010] U=(u1,u2,…,u N )

[0011] Where N represents the total number of users, and the kth user u k The action trajectory is expressed as:

[0012]

[0013] c i =(u k ,l i ,t i ,ts i )

[0014] Among them, c i is the trajectory information at the check-in location i, specifically represented by u k At time t i At the check-in location i Check-in record, ts iIndicates the specific time period of check-in. Each check-in location has a serial number label, k∈[1,N], i∈[1,M], where M represents the total number of check-in locations.

[0015] S200: Constructing a popularity bias mitigation model AGSRec, which includes a feature extraction module and a POI recommendation module. The feature extraction module includes an attractiveness perception graph bias balancing module AGB and a semantic enhancement bias balancing module SEB. The POI recommendation module includes a feature fusion module, a two-stream self-attention mechanism module, and a prediction layer module. The two-stream self-attention mechanism module includes a multi-head attention mechanism and a feedforward neural network.

[0016] S300: k The action trajectory is input into the feature extraction module to obtain u k Attractiveness perception map features and semantic enhancement features

[0017] S400: c i The user features are obtained by converting them into vectors through multi-layer perceptron. and Input POI recommendation module, user characteristics and After the feature fusion module is spliced, the user's attractiveness perception map features are obtained; user features and User semantic enhancement features are obtained through splicing through feature fusion module;

[0018] S500: Embed the serial number label of the check-in location into the user's attractiveness perception map feature and the user's semantic enhancement feature. At the same time, the user's attractiveness perception map feature and the user's semantic enhancement feature embedded with the serial number label of the check-in location are simultaneously input into the dual-stream self-attention mechanism module to obtain the unbiased user feature h b ;

[0019] S600: Repeat S300-S500 to obtain the unbiased user features of all users in U, and use the negative sampling strategy to train AGSRec;

[0020] Constructing the loss function of AGSRec And according to Update the parameters of AGSRec and stop training when the maximum number of training times is reached or the loss function no longer changes, and obtain the trained popularity deviation mitigation model AGSRec';

[0021] Among them, the calculation formula of the AGSRec model loss function is as follows:

[0022]

[0023] Among them, lpos Indicates the next actual check-in point, l neg Represents the randomly sampled negative samples in the negative sampling strategy;

[0024] S700: For a new user u, obtain u's movement trajectory T u , T u Each trajectory point in T contains a position number. u Input AGSRec' to get the unbiased user feature h' corresponding to the new user b , h′ b Input the prediction layer module to get the recommended score set Score of all check-in locations of u u , the calculation formula is as follows:

[0025] Score s =h′ b ·e s

[0026] Score=(Score1,…,Score s ,…,Score n )

[0027] Among them, Score s represents the recommendation score of the sth historical check-in location, e s Represents the location embedding vector of the s-th historical check-in location, s∈[1,n];

[0028] S800: Arrange the scores in Score in descending order, select the check-in locations corresponding to the first P scores as the POI recommendation list and output it.

[0029] Preferably, in S300, the process of obtaining the attractiveness perception map features is as follows:

[0030] S310: Setting the attraction threshold θ and the distance threshold δ, Extract c i Location attraction Category attractiveness Spatial attractiveness and space-time attraction The calculation expression is as follows:

[0031]

[0032] Among them, CTR(·) represents the function of calculating the click rate of the check-in location, M c Indicates the number of check-in location categories, cat i represents category i, CTR weekend (·) represents the function used to calculate the click rate of a location during the weekend, li represents the check-in location j, I represents the exponential function, d ij It is calculated based on longitude and latitude. i and l j The spatial distance between

[0033] S311: and Summing to get c i Multidimensional attractiveness value when When it is higher than θ, define c i Highly attractive location H ,when When it is lower than θ, define c i For low-attractive locations L ,in The calculation expression is as follows:

[0034]

[0035] S312: Repeat S310-S311 and calculate The multidimensional attractiveness values ​​of all check-in locations in the , and confirm that each check-in location corresponds to l H or l L ;

[0036] S313: u k All check-in locations interact with each other to obtain four types of interaction behaviors, namely hotspot aggregation type H2H, abnormal attraction type L2H, demand-driven H2L and long-tail interaction type L2L;

[0037] S314: Calculate the interaction value set between the two check-in locations i and j The calculation expression is as follows:

[0038]

[0039] Among them, i and j represent the check-in location, w ij represents the attraction interaction value between check-in location i and check-in location j, α L2H , α H2H , α L2L , α H2L Both represent the interaction type weight, and α L2H ,α H2H <1,α L2L ,α H2L >1;

[0040] S315: Calculate u k The attraction interaction value between any two check-in locations in the , sum up all the attraction interaction values ​​to get the attraction interaction value set wint ;

[0041] S316: W int Converted into attraction interaction graph G1 = (V, E, w int ), V represents the user, and E represents the edge set between check-in locations;

[0042] S320: Based on Form an attraction trajectory and calculate the average attraction weight between the two check-in locations i and j The calculation expression is as follows:

[0043]

[0044] Among them, a i and a j Denote the attractiveness of check-in location i and check-in location j, respectively. i With D j Represent the out-degree matrices of check-in location i and check-in location j respectively;

[0045] Calculate the attraction weight between any two check-in locations and obtain the attraction interaction value set w w ;

[0046] S321: w w Converted into attraction weight graph G2 = (V, E, w w );

[0047] S330: G1=(V,E,w int ) The interaction graph feature vector is obtained through the graph convolutional network The calculation expression is as follows:

[0048]

[0049] in, represents the out-degree matrix corresponding to the attraction interaction graph, W (l-1) represents the learnable parameter matrix of the l-1th layer of the attraction interaction graph in the graph convolutional network, σ is the activation function ReLU, Represents the attraction interaction graph with the self-connected adjacency matrix, Represents the interaction graph feature vector of the l-1th layer of the graph convolutional network, where when l=1, is the initialization value;

[0050] G2=(V,E,w w ) The weighted graph feature vector is obtained through the graph convolutional network The calculation expression is as follows:

[0051]

[0052] in, represents the out-degree matrix corresponding to the attraction weight graph, Θ (l-1) Represents the learnable parameter matrix of the attraction weight map in the l-1 layer of the graph convolutional network, Represents the attraction weight graph with self-connected adjacency matrix, Represents the interaction weight feature vector of the l-1th layer of the graph convolutional network, where when l=1, is the initialization value;

[0053] S340: Merger and Get the uth k Features of user's attractiveness perception graph The calculation expression is as follows:

[0054]

[0055] Preferably, in S300, the process of obtaining the semantic enhancement feature is as follows:

[0056] S350: Constructing a spatiotemporal pattern dynamic prompt template M1, a social association dynamic prompt template M2, and a regional ecological dynamic prompt template M3;

[0057] S351: According to M1, M2, M3 Rewrite them into three types of information: spatiotemporal pattern, social connection and regional ecology, and then combine the three types of information to get u k The complete semantic information M of is calculated as follows:

[0058] M=[CLS]+M1+[EOS]+[CLS]+M2+[EOS]+[CLS]+M3+[EOS]

[0059] Among them, [CLS] and [EOS] both represent template identifiers;

[0060] S352: Input M into the PLM model to obtain semantic enhancement features

[0061] Preferably, the PLM model in S352 may specifically be GPT-2.

[0062] Starting from the perspective of explicit user patterns and implicit semantic associations, the proposed method visualizes abstract popularity bias through attraction, and then uses graph-semantics to intervene in attraction to construct a data-level bias balancing framework. The AGSRec model constructs an attraction perception graph based on users' explicit interaction patterns. The interaction graph balances the weights between different user interaction behaviors, and the weight graph balances the attraction information between extreme locations. A graph convolutional network is then used to learn high-order attraction features to capture the complex associations between locations. The AGSRec model performs semantic enhancement based on users' implicit semantic associations. It constructs user spatiotemporal patterns, social connections, and regional ecological semantic information through prompt engineering, and extracts semantic features using a pre-trained language model to mine user preferences and location attraction information that cannot be directly calculated from statistical methods. Finally, a cross-modal attention mechanism is used to achieve a deep fusion of graph structural features and semantic features, eliminating popularity bias while maintaining the accuracy of personalized recommendations.

[0063] Compared with the prior art, the present invention has at least the following advantages:

[0064] 1. By constructing two different types of attraction perception maps, we explore users' explicit interaction patterns to balance the positive and negative effects of bias. In the attraction interaction map, the model divides two locations where users interact continuously into different types based on their attractiveness. It assigns low weights to hotspot clusters and unusual attraction interactions dominated by strong bias, while retaining high weights for demand-driven and long-tail interactions related to weak bias. This achieves the effect of balancing bias at the user interaction behavior level. In the attraction weight map, if there is continuous interaction between two locations, it is considered that there is an edge between the two locations, and the weight of the edge is calculated based on the average of their attractiveness, thereby balancing the attractiveness of different locations. By constructing a multi-level dynamic attraction perception map, we achieve temporal correlation modeling of users' explicit behavior patterns and quantification of interaction depth, alleviating the problem of recommendation results deviating from users' personalized preferences due to confounding variable debiasing in statistical calculations.

[0065] 2. To mine implicit semantic associations related to attraction in the data, the model constructs dynamic semantic templates from three perspectives: user spatiotemporal patterns, social connections, and regional ecology. The spatiotemporal pattern template highlights actual user behavior patterns, strengthening the model's capture of personalized preferences and mitigating the negative impact of bias. The social connection template compares the preferences of core user groups with those of the general public, capturing the behavioral specificity of niche but highly sticky groups, strengthening targeted exposure and community cohesion in long-tail scenarios, and effectively leveraging the positive impact of bias. The regional ecology template retains the characteristics of high-quality locations by mining implicit attractions within spatiotemporal regions, thereby leveraging the positive impact of bias. A pre-trained language model is used to mine implicit semantic associations in the templates, enhancing the debiasing effect.

[0066] 3. The present invention conducted experiments on two real-world location datasets. Compared with the recent high-quality recommendation system models, the method proposed in the present invention has relatively strong performance. At the same time, this method can give more traffic to long-tail locations and give popular locations a lower recommendation frequency, which is also effective in alleviating popularity bias. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a schematic diagram of the model framework of the present invention;

[0068] Figure 2 Schematic diagram of the dual-stream self-attention mechanism used in the present invention;

[0069] Figure 3 This is the interactive category weight hyperparameter experiment in the experiment of the present invention;

[0070] Figure 4 This is the semantic length threshold hyperparameter experiment in the experiment of the present invention;

[0071] Figure 5 This is the experiment of semantic and graph feature weight hyperparameters in the experiment of this invention;

[0072] Figure 6 This is the recommended result coverage experiment in the experiment of this invention;

[0073] Figure 7 This is an experiment to alleviate the long tail effect in the experiment of the present invention. DETAILED DESCRIPTION

[0074] The present invention is described in further detail below.

[0075] This paper proposes a popularity bias removal model based on an attractiveness perception graph and an enhancement module. The attractiveness perception graph module constructs two graph structures: an interaction graph dynamically adjusts weights based on the attractiveness strength of pairs of users' consecutive interaction locations, and a weighted graph uses the mean attractiveness of interaction nodes as edge weights. Together, these two graphs mine explicit user interaction patterns. The semantic enhancement module constructs dynamic semantic templates based on spatiotemporal patterns, social connections, and regional ecology to mine implicit semantic associations of attractiveness. These two modules collaborate on data-level biases, focusing on explicit behavioral patterns and implicit semantic associations, respectively. The effectiveness of this approach has been validated on real-world datasets.

[0076] See also Figure 1-Figure 7 A popularity bias mitigation method for location recommendation includes the following steps:

[0077] S100: Select a public user trajectory dataset, which includes a user set U and the action trajectory of each user. The expression of U is as follows:

[0078] U=(u1,u2,…,u N )

[0079] Where N represents the total number of users, and the kth user u k The action trajectory is expressed as:

[0080]

[0081] c i =(u k ,l i ,t i ,ts i )

[0082] Among them, c i is the trajectory information at the check-in location i, specifically represented by u k At time t i At the check-in location i Check-in record, ts i Indicates the specific time period of check-in. Each check-in location has a serial number label, k∈[1,N], i∈[1,M], where M represents the total number of check-in locations.

[0083] S200: Construct a popularity bias mitigation model AGSRec. AGSRec includes a feature extraction module and a POI recommendation module. The feature extraction module includes an attractiveness perception map bias balance module AGB and a semantic enhancement bias balance module SEB. The AGB module mines users' explicit interaction patterns by constructing two different types of attractiveness perception maps to balance the positive and negative effects of bias. The SEB module is designed to mine implicit semantic associations related to attractiveness in the data. The model constructs dynamic semantic templates from three perspectives: user spatiotemporal patterns, social associations, and regional ecology. The spatiotemporal patterns extract users' personal interest preferences to mitigate the negative effects of bias, while social associations and regional ecology assist in enhancing the positive effects of attractiveness information on bias. The POI recommendation module includes a feature fusion module, a dual-stream self-attention mechanism module, and a prediction layer module. The dual-stream self-attention mechanism module includes a multi-head attention mechanism and a feedforward neural network. The multi-head attention mechanism and the feedforward neural network are existing technologies.

[0084] S300: k The action trajectory is input into the feature extraction module to obtain u k Attractiveness perception map features and semantic enhancement features

[0085] In S300, the process of obtaining the attractiveness perception map features is as follows:

[0086] S310: Setting the attraction threshold θ and the distance threshold δ, Extract c i Location attraction Category attractiveness Spatial attractiveness and space-time attraction The calculation expression is as follows:

[0087]

[0088] Among them, CTR(·) represents the function of calculating the click rate of the check-in location, M c Indicates the number of check-in location categories, cat i represents category i, CTR weekend (·) represents the function used to calculate the click rate of a location during the weekend, l i represents the check-in location j, I represents the exponential function, d ij It is calculated based on longitude and latitude. i and l j The spatial distance between

[0089] S311: and Summing to get c i Multidimensional attractiveness value when When it is higher than θ, define c i Highly attractive location H ,when When it is lower than θ, define c i For low-attractive locations L ,in The calculation expression is as follows:

[0090]

[0091] S312: Repeat S310-S311 and calculate The multidimensional attractiveness values ​​of all check-in locations in the , and confirm that each check-in location corresponds to l H or l L ;

[0092] S313: u k All check-in locations interact with each other to obtain four types of interaction behaviors, namely hotspot aggregation type H2H, abnormal attraction type L2H, demand-driven H2L and long-tail interaction type L2L;

[0093] S314: Calculate the interaction value set between the two check-in locations i and j The calculation expression is as follows:

[0094]

[0095] Among them, i and j represent the check-in location, w ijrepresents the attraction interaction value between check-in location i and check-in location j, α L2H , α H2H , α L2L , α H2L Both represent the interaction type weight, and α L2H ,α H2H <1,α L2L ,α H2L >1;

[0096] S315: Calculate u k The attraction interaction value between any two check-in locations in the , sum up all the attraction interaction values ​​to get the attraction interaction value set w int ;

[0097] S316: W int Converted into attraction interaction graph G1 = (V, E, w int ), V represents the user, and E represents the edge set between check-in locations;

[0098] S320: Based on Form an attraction trajectory and calculate the average attraction weight between the two check-in locations i and j The calculation expression is as follows:

[0099]

[0100] Among them, a i and a j Denote the attractiveness of check-in location i and check-in location j, respectively. i With D j Represent the out-degree matrices of check-in location i and check-in location j respectively; the out-degree matrix refers to a diagonal matrix used to describe the out-degree of a directed graph node. Its core function is to record the number of edges from each node to other nodes;

[0101] Calculate the attraction weight between any two check-in locations and obtain the attraction interaction value set w w ;

[0102] S321: w w Converted into attraction weight graph G2 = (V, E, w w );

[0103] S330: G1=(V,E,w int ) The interaction graph feature vector is obtained through the graph convolutional network Graph convolutional network is an existing technology, and the calculation expression is as follows:

[0104]

[0105] in, represents the out-degree matrix corresponding to the attraction interaction graph, W (l-1) represents the learnable parameter matrix of the l-1th layer of the attraction interaction graph in the graph convolutional network, σ is the activation function ReLU, The attraction interaction graph adds a self-connected adjacency matrix. This adjacency matrix adds an edge pointing to each node to ensure that the node retains its original characteristics while aggregating neighbor information. Represents the interaction graph feature vector of the l-1th layer of the graph convolutional network, where when l=1, is the initialization value;

[0106] G2=(V,E,w w ) The weighted graph feature vector is obtained through the graph convolutional network The calculation expression is as follows:

[0107]

[0108] in, represents the out-degree matrix corresponding to the attraction weight graph, Θ (I-1) Represents the learnable parameter matrix of the attraction weight map in the l-1 layer of the graph convolutional network, Represents the attraction weight graph with self-connected adjacency matrix, Represents the interaction weight feature vector of the l-1th layer of the graph convolutional network, where when l=1, is the initialization value;

[0109] S340: Merger and Get the uth k Features of user's attractiveness perception graph The calculation expression is as follows:

[0110]

[0111] In S300, the process of obtaining the semantic enhancement feature is as follows:

[0112] S350: Constructing a spatiotemporal pattern dynamic prompt template M1, a social association dynamic prompt template M2, and a regional ecology dynamic prompt template M3; the dynamic prompt templates can be customized and modified as required;

[0113] S351: According to M1, M2, M3 Rewrite them into three types of information: spatiotemporal pattern, social connection and regional ecology, and then combine the three types of information to get u k The complete semantic information M of is calculated as follows:

[0114] M=[CLS]+M1+[EOS]+[CLS]+M2+[EOS]+[CLS]+M3+[EOS]

[0115] Among them, [CLS] and [EOS] both represent template identifiers;

[0116] S352: Input M into the PLM model to obtain semantic enhancement features The PLM model is an existing language model technology; the PLM model in S352 can specifically be GPT-2, and GPT-2 is an existing language model technology;

[0117] S400: c i The user features are obtained by converting them into vectors through multi-layer perceptron. and Input POI recommendation module, user characteristics and After the feature fusion module is spliced, the user's attractiveness perception map features are obtained; user features and User semantic enhancement features are obtained through splicing through feature fusion module; multi-layer perceptron is an existing technology;

[0118] S500: Embed the serial number label of the check-in location into the user's attractiveness perception map feature and the user's semantic enhancement feature. At the same time, the user's attractiveness perception map feature and the user's semantic enhancement feature embedded with the serial number label of the check-in location are simultaneously input into the dual-stream self-attention mechanism module to obtain the unbiased user feature h b The embedding here refers to the embedding layer in the multi-layer perceptron, which is a state-of-the-art technology. The role of unbiased user features is to suppress negative effects and utilize positive attraction.

[0119] S600: Repeat S300-S500 to obtain unbiased user features for all users in U, and simultaneously train AGSRec using a negative sampling strategy, which is an existing technology.

[0120] Constructing the loss function of AGSRec And according to Update the parameters of AGSRec and stop training when the maximum number of training times is reached or the loss function no longer changes, and obtain the trained popularity deviation mitigation model AGSRec';

[0121] Among them, the calculation formula of the AGSRec model loss function is as follows:

[0122]

[0123] Among them, l pos Indicates the next actual check-in point, l negRepresents the randomly sampled negative samples in the negative sampling strategy;

[0124] S700: For a new user u, obtain u's movement trajectory T u , T u Each trajectory point in T contains a position number. u Input AGSRec' to get the unbiased user feature h' corresponding to the new user b , b′ b Input the prediction layer module to get the recommended score set Score of all check-in locations of u u , the calculation formula is as follows:

[0125] Score s =h′ b ·e s

[0126] Score=(Score1,…,Score s ,…,Score n )

[0127] Among them, Score s represents the recommendation score of the sth historical check-in location, e s represents the location embedding vector of the sth historical check-in location, s∈[1,n]; e s The multi-layer perceptron converts the trajectory sequence information of point s into a position embedding vector; Score j The value of reflects the probability of location j being selected by the user, and a higher score indicates a higher recommendation priority;

[0128] S800: Arrange the scores in Score in descending order, select the check-in locations corresponding to the first P scores as the POI recommendation list and output it.

[0129] Experimental content and results

[0130] 1. Dataset

[0131] This paper uses two real-world location datasets for experiments: NYC and Tokyo, Japan. The NYC and TKY data come from the FourSquare platform, a mobile service website that uses user location information. The datasets record user check-in behavior in New York City (NYC) and Tokyo, Japan (TKY) between April 2012 and February 2013.

[0132] To maintain data quality in the experiment, the model of the present invention uses a filtering method to exclude users with fewer than 10 check-in records and locations with fewer than 10 visits. The statistical data of each dataset after filtering are shown in Table 1, which covers the number of users, number of locations, number of location categories, number of check-ins, and average trajectory length.

[0133] Table 1 Dataset details

[0134]

[0135] 2. Evaluation indicators

[0136] This paper uses the two indicators NDCG@k and HR@ to evaluate the model quality, and the k value in the experimental setting is 5, 10, and 20.

[0137] Normalized Discounted Cumulative Gain (NDCG) is an important metric for evaluating ranking validity. It quantifies the ranking value of the recommendation sequence through a dynamic weight allocation mechanism. This metric combines the position decay factor with the relevance weight. Its calculation process adopts a step-by-step strategy: related items at the top of the recommendation list will receive higher weights. i represents the relevance of the i-th location, and DCG represents the discounted cumulative gain, which is calculated as follows:

[0138]

[0139] Based on the DCG calculation, the final NDCG calculation formula is as follows, where IDCG represents the ideal discounted cumulative gain based on the ideal relevance in descending order.

[0140]

[0141] The hit ratio (HR) metric focuses on verifying prediction accuracy, using a binary decision mechanism to measure the intersection of the recommended list and the user's actual preferences. The core value of this metric lies in revealing the recommendation system's ability to accurately target users through Boolean logic. Its calculation formula is as follows, where hit(i) indicates whether the recommendation result matches user i, with 1 indicating a hit and 0 indicating a miss.

[0142]

[0143] 3. Experimental Setup

[0144] In this experiment, the model was built using the PyTorch framework. The hardware configuration was as follows: CPU: Intel(R) Core(TM) i9-10980XE, GPU: NVIDIA GeForce RTX 3060 12GB. The model parameters were iteratively optimized using the Adam optimizer, with a learning rate of 0.001, an embedding dimension of 64, and a batch size of 128. Training was performed for 500 epochs, with validation evaluation performed every 10 epochs.

[0145] In the AGB module, the interaction category weight α when constructing the attraction interaction graph is confirmed through hyperparameter experiments. The four weights are the weights of abnormal attraction, hotspot aggregation, long-tail interaction, and demand-driven. The experimental settings are {0.7, 0.9, 1.7, 1.9}, {0.5, 0.7, 1.5, 1.7}, {0.3, 0.5, 1.3, 1.5}, and {0.1, 0.3, 1.1, 1.3}. The attraction interaction graph weight β int and the attraction weight graph weight β w The hyperparameters are also confirmed by experiments, and the values ​​are 0.2, 0.4, 0.6, and 0.8 respectively. In the SEB module, the semantic length threshold δ s Through the hyperparameter experiment setting, the semantic feature weight α is set to 5, 10, 15, and 20 respectively. s In the hyperparameter experiments, it is set to 0.2, 0.4, 0.6, and 0.8.

[0146] The special hyperparameters of the baseline model in the comparative experiment were configured according to the original text and set to 500 rounds for training. The present invention used the code provided by the RecBole platform. The platform provides an integrated recommendation system model framework, which the present invention used to construct the CL4SRec, DuoRec and TiSASRec models. All models were trained for 500 epochs with a batch size of 128 and an embedding dimension of 50. The learning rate of SASRec, SSE-PT, AC-TSR, FeaRec and TiSASRec was set to 0.01, while the learning rate of Cl4SRec, DuoRec, PDA, DSDRec and CauseRec was set to 0.001. DSDRec uses GPT-2 as the pre-trained language model, and the learning rate of the GETNext model is set to 0.01. The model-specific hyperparameters are consistent with the original paper, and the dataset used also filters out users who have checked in less than 10 times and places that have checked in less than 10 times.

[0147] 4. Comparative Experiment

[0148] The baseline models selected for comparison in the experiments of the present invention are as follows:

[0149] SASRec (ICDM18): Utilizes a dynamic attention mechanism to adaptively select key historical behaviors for prediction, outperforming traditional sequence models in both sparse and dense data scenarios.

[0150] AC-TSR (CIKM23): It captures the order and distance relationship between items through a spatial calibrator and dynamically optimizes the attention weight distribution in combination with an adversarial calibrator.

[0151] Cl4SRec (ICDE22): applies self-supervised learning to recommendation systems, combining three data augmentation strategies: item masking, cropping, and reordering, to construct comparison samples, effectively alleviating the problem of data sparsity.

[0152] DuoRec (WSDM22): It combines a contrastive learning regularization strategy with Dropout-based model-level data augmentation, and innovatively adopts a sampling mechanism that uses sequences of the same target item as hard positive samples.

[0153] FEARec (SIGIR23): It captures full-band information through frequency-domain ramp structure self-attention, integrates time-frequency dual-domain attention, and introduces contrastive learning and frequency-domain regularization to achieve multi-view alignment.

[0154] TiSASRec (WSDM22): By simultaneously modeling the absolute position of items in an interaction sequence and the time interval between adjacent behaviors, it breaks through the limitation of traditional sequential recommendation that ignores temporal density.

[0155] DSDRec (Inf.Sci24): A Transformer framework based on personalized user semantic feature enhancement to optimize the accuracy of point of interest recommendations by capturing interest patterns.

[0156] GETNext (SIGIR22): Modeling global user trajectory flows based on graph structures to capture common mobility patterns, combining contextual learning with multi-task learning to optimize the POI recommendation model and achieve accurate prediction in personalized scenarios.

[0157] Table 2 Performance comparison on the NYC dataset

[0158]

[0159]

[0160] The results of the comparative experiments are shown in Tables 2 and 3, where boldface indicates the best-performing model, and underlined indicates the second-best-performing model. AGSRec achieved the best results across all metrics on both datasets. On the NYC dataset, the highest improvement over the second-best model was HR@5, with a 3.06% improvement; the lowest improvement was HR@10, with a 1.31% improvement. The average improvement was 2.32% for the NDCG metric and 1.99% for the HR metric. On the TKY dataset, the highest improvement over the second-best model was NDCG@5, with a 4.57% improvement; the lowest was HR@20, with a 0.66% improvement. On average, the model achieved superior performance on the TKY dataset compared to the NYC dataset. This may be due to the higher quality of the TKY dataset, with a larger number of users, locations, and interaction records. The data size enhances the ability of the attractiveness perception graph to mine user interaction patterns, while the scene diversity strengthens the semantic enhancement module's ability to express implicit semantic associations.

[0161] Table 3 Performance comparison on the TKY dataset

[0162]

[0163] Specifically, AGSRec's significant advantage on the TKY dataset stems from the deep fit between its model design and data characteristics. Experiments show that while traditional sequential recommendation models such as SASRec and AC-TSR can capture long-term dependencies, they lack explicit modeling of the spatiotemporal dynamics of POI scenarios, making it difficult to balance spatiotemporal proximity and semantic relevance in complex user trajectories. Debiasing models such as PDA and CauseRec can mitigate popularity bias, but they fail to address the problem of ignoring the positive impact of bias in POI scenarios. In high-quality data, they lose real signals due to over-suppression of bias. Data augmentation models such as Cl4SRec and FEARec enhance user representations through contrastive learning or frequency-domain mixing, but fail to deeply integrate multimodal semantics such as spatiotemporal periods and social connections, resulting in limited data-level interest capture. Dedicated POI models, while incorporating spatiotemporal or graph structure information, fail to achieve global multimodal collaboration. For example, TiSASRec models time intervals in isolation without incorporating spatial clustering, while GETNext relies on static graphs and ignores dynamic changes in attractiveness. In contrast, our model achieves multimodal bias balance within TKY's high-quality data through a dual-path collaboration: the Attraction Perception Graph and the Semantic Enhancement Module. The Attraction Perception Graph dynamically adjusts edge weights based on intensive interactions, balancing the attractiveness of extreme locations, suppressing the negative impact of bias while preserving the characteristics of high-quality locations. The Semantic Enhancement Module integrates spatiotemporal patterns, social connections, and regional ecological characteristics, leveraging TKY's rich multidimensional information to construct dynamic templates that capture user preferences and attractiveness. This design results in particularly strong performance on ranking-sensitive metrics, such as a 4.57% improvement in NDCG@5.

[0164] 5. Ablation Experiment

[0165] We tested the effectiveness of the attraction perception map and semantic enhancement modules in our model. "w / " represents the use of only one module. In ablation experiments, we tested the semantic enhancement module alone, the attraction interaction map alone, the attraction weight map alone, and the attraction perception map alone (interaction map + weight map).

[0166] Tables 4 and 5 present the results of ablation experiments on two datasets. Bold text represents the optimal performance for each metric, and underlined text represents the suboptimal performance for each metric. Analysis of the experimental results in both tables shows that using the semantic enhancement module alone or the attractiveness perception map alone results in decreased performance across all metrics, demonstrating the positive impact of each module on model improvement. On the NYC dataset, using the interaction graph or weight graph alone achieved NDCG@5 and HR@5 metrics of 0.4805 and 0.7922, respectively, slightly outperforming the semantic enhancement module's 0.7941 and 0.4785 results. However, directly superimposing the two yielded no benefit and even exhibited fluctuations, with HR@5 dropping to 0.7941. This suggests redundancy within the graph structure module, necessitating data-level calibration through semantic enhancement. This validates the importance of complementary multimodal information for long-tail locations. On the TKY dataset, the interaction graph alone performed better than the weight graph, achieving an NDCG@5 of 0.5133 for the former and 0.5091 for the latter, demonstrating the effectiveness of dynamic edge weight adjustment in the presence of dense interaction data. However, the independent performance of semantic enhancement is weaker than that of graph structure, indicating that the explicit behavior patterns of graph structure in high-quality data are easier to capture data-level regularities than implicit semantic associations.

[0167] Table 4 Ablation experiments on the NYC dataset

[0168]

[0169] Table 5 Ablation experiments on the TKY dataset

[0170]

[0171] 6. Hyperparameter Experimentation

[0172] This experiment is a hyperparameter experiment on some important parameters in the AGSRec model. When constructing the attraction interaction graph, four types of interactions are divided according to the attractiveness of the user interaction location, and the corresponding interaction category weight α is assigned. The present invention conducts a hyperparameter experiment on the value of α. The four values ​​correspond to abnormal attraction, hotspot aggregation, long-tail interaction and demand drive, respectively. The specific values ​​are 1 = {0.7, 0.9, 1.7, 1.9}, 2 = {0.5, 0.7, 1.5, 1.7}, 3 = {0.3, 0.5, 1.3, 1.5}, 4 = {0.1, 0.3, 1.1, 1.3}. The experimental results are as follows Figure 2As shown in the figure, in the NYC dataset, the fourth type performs significantly better than the other three types, so α = {0.1, 0.3, 1.1, 1.3} was selected for NYC. In the TKY dataset, the second type achieved the best results in the NDCG@10 metric, and the third type achieved the best results in the HR@10 metric. Overall, the difference between the two types was approximately 0.02. Finally, in the TKY dataset, α = {0.5, 0.7, 1.5, 1.7} was selected.

[0173] After constructing the semantic template, the model selects the user’s latest interaction δ according to the semantic length threshold. s Build semantic templates for each location. Here, s Conduct hyperparameter experiments and set their values ​​to 5, 10, 15, and 20. The final experimental results are as follows Figure 4 As shown in the figure, in NYC, when the threshold is 15, the best results are achieved in both indicators, with NDCG@10 being 0.4961 and HR@10 being 0.8495. In TKY, when the threshold is 15, good results are also achieved, with NDCG@10 being 0.5240 and HR@10 being 0.8971. Finally, let δ s The threshold is 15, which can capture enough user semantic features and avoid information loss due to a too small threshold, while not introducing noise due to being too long. Figure 3 shown.

[0174] The model uses semantic feature weight α during training s , gravitational interaction graph weight β int , the weight of the attraction weight graph β w This section experiments on these three hyperparameters, with the values ​​of each hyperparameter set to 0.2, 0.4, 0.6, and 0.8, respectively, and evaluates them using the NDCG@10 and HR@10 indicators. The experimental results are shown in the figure below. Figure 4 As shown. In the NYC dataset, when the semantic weight is 0.4, the indicator achieves the best performance, NDCG@10 is 0.4988, HR@10 is 0.9550; when the interaction graph weight is 0.8, the indicator achieves the best performance, NDCG@10 is 0.4939, HR@10 is 0.8458; when the weight graph weight is 0.8, the indicator achieves the best performance, NDCG@10 is 0.4986, HR@10 is 0.8578. In the TKY dataset, α s , β int , β w All of them get better results when 0.8. Finally, considering all factors, we set α s , β int , β w They are 0.4, 0.8, and 0.8 respectively.

[0175] 7. Debiasing effect experiment

[0176] The present invention conducts two different experiments on the debiasing effect. In the recommendation result coverage experiment, the model is compared with the coverage values ​​of DUORec, AC-TSR, CauseRec, FEARec, SASRec, GETNext, and DSDRec. The experimental results are as follows: Figure 5 As shown in the figure, the model achieved the highest coverage on both datasets. On the NYC dataset, the coverage was 82.56%, a maximum improvement of 3.63% and a minimum improvement of 0.57% compared to other models. On the TKY dataset, the coverage was 86.85%, a maximum improvement of 8.04% and a minimum improvement of 0.06% compared to other models. Experimental results demonstrate that this approach can provide users with more diverse choices and alleviate the problem of recommendation homogeneity. This provides users with more choices, allowing them to consider their personalized preferences when making decisions, and mitigates consistency bias.

[0177] The experimental results of the model of the present invention are compared with CL4SRec, FEARec, DuoRec, AC-TSR, GETNext and DSDRec. Figure 6 As shown, the present invention marks AGSRec as a red straight line, and other models are marked as dot-dashed lines. The curve is drawn by a b-spline curve. In the NYC dataset, the model gave higher exposure rates to groups 1, 4, and 5 with weaker appeal, and the second lowest exposure rate to group 8 with the highest appeal. In the TKY dataset, higher exposure rates were given to groups 1 and 3 with weaker appeal, and the lowest exposure rate was given to group 10 with the highest appeal. This proves that this method can give long-tail locations more traffic and is effective in alleviating exposure bias. At the same time, this method also gives popular locations a lower recommendation frequency, proving that it is also effective in alleviating popularity bias.

[0178] The above experiments prove that the model of the present invention has a certain degree of alleviating effect on exposure bias, popularity bias, and consistency bias, and proves the effectiveness of the graph-semantics-based data-level bias balancing method in debiasing.

[0179] 8. Conclusion

[0180] To fully explore the implicit information of confounding factors and alleviate the incompleteness of confounding factor calculation results caused by statistical methods, the present invention proposes a popularity bias balance model AGSRec based on the attractiveness perception graph and semantic enhancement. The attractiveness perception graph module is based on the user's explicit interaction pattern, in which the interaction graph dynamically adjusts the weights to balance positive and negative biases based on the attractiveness intensity of the user's continuous interaction locations; the weight graph balances regional attractiveness differences by calculating the interaction edge values ​​through the mean attractiveness. The semantic enhancement module is based on implicit semantic associations, constructs dynamic semantic templates based on time, space, social, and regional ecology, and mines implicit features. The two modules balance biases at the levels of explicit interaction patterns and implicit semantic associations, respectively, to achieve recommendation results with balanced popularity bias.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention 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 invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for alleviating popularity bias in location recommendations, characterized by: The steps include: S100: Select a public user trajectory dataset, which includes a user set U and the action trajectory of each user. The expression of U is as follows: U=(u1,u2,…,u N ) Where N represents the total number of users, and the kth user u k The action trajectory is expressed as: c i =(one k ,l i ,t i ,ts i ) Among them, c i is the trajectory information at the check-in location i, specifically represented by u k At time t i At the check-in location i Check-in record, ts i Indicates the specific time period of check-in. Each check-in location has a serial number label, k∈[1,N], i∈[1,M], where M represents the total number of check-in locations. S200: Constructing a popularity bias mitigation model AGSRec, which includes a feature extraction module and a POI recommendation module. The feature extraction module includes an attractiveness perception graph bias balancing module AGB and a semantic enhancement bias balancing module SEB. The POI recommendation module includes a feature fusion module, a two-stream self-attention mechanism module, and a prediction layer module. The two-stream self-attention mechanism module includes a multi-head attention mechanism and a feedforward neural network. S300: k The action trajectory is input into the feature extraction module to obtain u k Attractiveness perception map features and semantic enhancement features S400: c i The user features are obtained by converting them into vectors through multi-layer perceptron. and Input POI recommendation module, user characteristics and After the feature fusion module is spliced, the user's attractiveness perception map features are obtained; user features and User semantic enhancement features are obtained through splicing through feature fusion module; S500: Embed the serial number label of the check-in location into the user's attractiveness perception map feature and the user's semantic enhancement feature. At the same time, the user's attractiveness perception map feature and the user's semantic enhancement feature embedded with the serial number label of the check-in location are simultaneously input into the dual-stream self-attention mechanism module to obtain the unbiased user feature h b ; S600: Repeat S300-S500 to obtain the unbiased user features of all users in U, and use the negative sampling strategy to train AGSRec; Constructing the loss function of AGSRec And according to Update the parameters of AGSRec and stop training when the maximum number of training times is reached or the loss function no longer changes, and obtain the trained popularity deviation mitigation model AGSRec'; Among them, the calculation formula of the AGSRec model loss function is as follows: Among them, l pos Indicates the next actual check-in point, l neg Represents the randomly sampled negative samples in the negative sampling strategy; S700: For a new user u, obtain u's movement trajectory T u , T u Each trajectory point in T contains a position number. u Input AGSRec' to get the unbiased user feature h' corresponding to the new user b , h′ b Input the prediction layer module to get the recommended score set Score of all check-in locations of u u , the calculation formula is as follows: Score s =h′ b ·e s Score=(Score1,…,Score s ,…,Score n ) Among them, Score s represents the recommendation score of the sth historical check-in location, e s Represents the location embedding vector of the s-th historical check-in location, s∈[1,n]; S800: Arrange the scores in Score in descending order, select the check-in locations corresponding to the first P scores as the POI recommendation list and output it.

2. The method for alleviating popularity deviation for location recommendation according to claim 1, characterized in that: In S300, the process of obtaining the attractiveness perception map features is as follows: S310: Setting the attraction threshold θ and the distance threshold δ, Extract c i Location attraction Category attractiveness Spatial attractiveness and space-time attraction The calculation expression is as follows: Among them, CTR(·) represents the function of calculating the click rate of the check-in location, M c Indicates the number of check-in location categories, cat i represents category i, CTR weekend (·) represents the function used to calculate the click rate of a location during the weekend, l i represents the check-in location j, I represents the exponential function, d ij It is calculated based on longitude and latitude. i and l j The spatial distance between S311: and Summing to get c i Multidimensional attractiveness value when When it is higher than θ, define c i Highly attractive location H ,when When it is lower than θ, define c i For low-attractive locations L ,in The calculation expression is as follows: S312: Repeat S310-S311 and calculate The multidimensional attractiveness values ​​of all check-in locations in the , and confirm that each check-in location corresponds to l H or l L ; S313: u k All check-in locations interact with each other to obtain four types of interaction behaviors, namely hotspot aggregation type H2H, abnormal attraction type L2H, demand-driven H2L and long-tail interaction type L2L; S314: Calculate the interaction value set between the two check-in locations i and j The calculation expression is as follows: Among them, i and j represent the check-in location, w ij represents the attraction interaction value between check-in location i and check-in location j, α L2H , α H2H , α L2L , α H2L Both represent the interaction type weight, and α L2H ,α H2H <1,α L2L ,α H2L >1; S315: Calculate u k The attraction interaction value between any two check-in locations in the , sum up all the attraction interaction values ​​to get the attraction interaction value set w int ; S316: W int Converted into attraction interaction graph G1 = (V, E, w int ), V represents the user, and E represents the edge set between check-in locations; S320: Based on Form an attraction trajectory and calculate the average attraction weight between the two check-in locations i and j The calculation expression is as follows: Among them, a i and a j Denote the attractiveness of check-in location i and check-in location j, respectively. i With D j Represent the out-degree matrices of check-in location i and check-in location j respectively; Calculate the attraction weight between any two check-in locations and obtain the attraction interaction value set w w ; S321: w w Converted into attraction weight graph G2 = (V, E, w w ); S330: G1=(V,E,w int ) The interaction graph feature vector is obtained through the graph convolutional network The calculation expression is as follows: in, represents the out-degree matrix corresponding to the attraction interaction graph, W (l-1) represents the learnable parameter matrix of the l-1th layer of the attraction interaction graph in the graph convolutional network, σ is the activation function ReLU, Represents the attraction interaction graph with the self-connected adjacency matrix, Represents the interaction graph feature vector of the l-1th layer of the graph convolutional network, where when l=1, is the initialization value; G2=(V,E,w w ) The weighted graph feature vector is obtained through the graph convolutional network The calculation expression is as follows: in, represents the out-degree matrix corresponding to the attraction weight graph, Θ (l-1) Represents the learnable parameter matrix of the attraction weight map in the l-1 layer of the graph convolutional network, Represents the attraction weight graph with self-connected adjacency matrix, Represents the interaction weight feature vector of the l-1th layer of the graph convolutional network, where when l=1, is the initialization value; S340: Merger and Get the uth k Features of user's attractiveness perception graph The calculation expression is as follows:

3. The method for alleviating popularity deviation for location recommendation according to claim 2, characterized in that: In S300, the process of obtaining the semantic enhancement feature is as follows: S350: Constructing a spatiotemporal pattern dynamic prompt template M1, a social association dynamic prompt template M2, and a regional ecological dynamic prompt template M3; S351: According to M1, M2, M3 Rewrite them into three types of information: spatiotemporal pattern, social connection and regional ecology, and then combine the three types of information to get u k The complete semantic information M of is calculated as follows: M=[CLS]+M1+[EOS]+[CLS]+M2+[EOS]+[CLS]+M3+[EOS] Among them, [CLS] and [EOS] both represent template identifiers; S352: Input M into the PLM model to obtain semantic enhancement features 4. The method for alleviating popularity deviation for location recommendation according to claim 3, characterized in that: The PLM model in S352 may specifically be GPT-2.

Images