A point of interest recommendation debiasing method based on a causal diagram

Abstract

The present application relates to a kind of point of interest recommendation debiasing method based on causal diagram, it is a unified debiasing method, by structure causal diagram to attraction, by realizing the balance between positive and positive influence, the method of the present application helps to reduce bias and realize balance.The method uses counterfactual reasoning and self-supervised learning to reduce the influence of negative attraction, which is achieved by maximizing the consistency between counterfactuals and observed data in vector space.In addition, the attraction features learned from multi-dimensional attraction are combined with user features to take advantage of the positive effects of attraction.In the experiment, relative to other models, the method improves the recommendation frequency of low-attraction sites while maintaining good overall performance.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a bias removal method for interest point recommendation based on causal graphs. Background Technology

[0002] Next-Point-of-Interest (NPI) recommendation aims to suggest locations that a user might be interested in at the next moment. It is a sequence recommendation method based on historical data. It plays an important role in many fields, such as user exploration, business decision-making, and urban planning, and has received increasing attention from researchers in recent years. However, due to the sparser nature of location data and the dynamic changes in user behavior, it is more susceptible to biases in the recommendation process.

[0003] Despite some progress in existing research, several challenges remain. Firstly, different methods have been used to address biases arising from various causes, resulting in a fragmented approach to bias correction and a lack of a unified paradigm. For instance, to mitigate popularity bias, the do-calculus method in statistics is employed, reducing popular location recommendations by cutting off the backdoor path of popularity's influence on users. Secondly, a counterfactual contrastive learning method (CCL) is proposed to address exposure bias. This method utilizes three unique positive sample sampling strategies, combined with estimated exposure probabilities or random counterfactual samples, to minimize exposure bias through a comparison of self-supervised learning and Inverse Phenomenon Indicator (IPS). Thirdly, a structural causal graph is constructed to depict the impact of consistency bias in a dynamic recommendation environment. A variational autoencoder is used to learn the distribution of consistency bias, thereby generating bias data and ultimately reconstructing user interaction characteristics unaffected by it.

[0004] These methods use different models and techniques to address biases caused by different reasons, making it difficult to unify the bias removal paradigm. Ultimately, different bias removal methods may produce different recommendation results, thereby reducing the consistency and reliability of the system. On the other hand, most existing research has ignored the positive impact of bias. When a location is recommended to a large number of users, in addition to being affected by the Matthew effect caused by bias, the location itself must have unique advantages (well-known brands, current fashion). If all biases are treated as negative impacts and mitigated, it may damage the advantages of these high-quality locations, thereby affecting the quality of recommendations. Summary of the Invention

[0005] In view of the above-mentioned problems in the existing technology, the technical problem to be solved by the present invention is: how to construct a unified debiasing paradigm that can both mitigate the negative impact of bias and utilize its positive impact.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a bias removal method for interest point recommendation based on causal graphs, comprising the following steps:

[0007] S1: Obtain the user's historical trajectory, let U = (u1, u2, ..., u3) N Let L be the set of users, where N is the number of users, and L = (l1, l2, ..., ln). M Let M be the set of locations, and l be the number of locations. i =(id) i ,cat i ,lon i ,lat i ), representing a detailed description of the i-th check-in location, id i ,cat i ,lon i ,lat i These represent the ID, category, longitude, and latitude of the i-th check-in location, respectively; user u i The travel trajectory is defined as Check-in location c for each trajectory j =(u i ,l j ,t j ) represents user u i At time t j Go to location l j Sign in.

[0008] S2: Construct the WNA module to mitigate the negative impact of attraction on users, and calculate... Corresponding counterfactual enhancement data T, unaffected by negative attraction ca And the observational augmentation data T influenced by attraction oa In each batch of training, T generated by the same user will be used. ca With T oa As positive samples, the other users in this batch are used as negative samples to train WNA, when the self-supervised learning loss... Convergence indicates the end of training, and the trained WNA is obtained. The input of trained WNA yields user features whose attractiveness is weakened.

[0009] S3: Construct the LPA (Limited Approach Module) which utilizes the positive influence of attractiveness, and use the user features obtained in S2 that weaken the influence of attractiveness as input to train the LPA. During training, gradient descent is used to update the model parameters of both the LPA and the trained WNA. When the total loss... Convergence indicates the end of training, yielding the optimal LPA and optimal WNA.

[0010] S4: Obtain the trajectory of a new user, and extract user features h' from the trajectory using the feature extractors in the optimal LPA and optimal WNA. P , will h'P Embedded e in different locations p Matrix multiplication is performed, and the final value is used as the recommendation score for each location. The higher the recommendation score, the more likely the location is to be selected by new users. The locations are then sorted in descending order according to their recommendation scores to obtain the recommendation list.

[0011] Furthermore, in S2, the calculation The corresponding T ca At that time, first Check-in location c j Classified as either highly attractive or weakly attractive locations;

[0012] A comprehensive description of the multidimensional attractiveness associated with each location is given by MA = (LA, CA, SA, TA), where LA is the inherent attractiveness of the current location, CA is the attractiveness associated with the location category, SA is the attractiveness derived from spatial features, and TA is the attractiveness caused by the location check-in time.

[0013]

[0014]

[0015]

[0016]

[0017] Where CTR(·) represents the function used to calculate the location click-through rate, d ij Location calculated based on latitude and longitude i and location l j The spatial distance between them, where δ is the threshold. For the indicator function, if condition d ij If δ is true, return 1; otherwise, return 0. (CTR) weekend (·) represents a function used to calculate weekend location click-through rates.

[0018] Specify different thresholds θ for the multidimensional attraction MA A =(θ l ,θ c ,θ s ,θ t If c j The corresponding MA>θ A Then c j Classified as a highly attractive location Otherwise, c j Location with low attraction The following is the equation for determining the location attraction:

[0019]

[0020] Where, θ l ,θ c ,θ s ,θ t These represent the thresholds for LA, CA, SA, and TA, respectively.

[0021] Furthermore, in S2, the calculation The corresponding T ca The method is as follows:

[0022] Based on the question, "How would users act if they weren't influenced by the allure?", we design a scenario where excessive attraction leads to a shift in travel intentions, assuming a specific trajectory within this scenario. for in Indicates a more attractive location. express The next location is a weak attraction location, c n This is the nth location.

[0023] Using cosine similarity calculation and The formula for determining the similarity between them is as follows:

[0024]

[0025] Where E(·) represents the embedding method, used to map locations to a vector space. This represents the similarity function. The following is the equation for counterfactual data augmentation in the scenario of travel intention replacement:

[0026]

[0027] Replace(·) represents generating counterfactual trajectories that are unaffected by attraction by replacing them. for The multi-dimensional attraction of Tr is used as Tca.

[0028] Furthermore, in S2, the calculation The corresponding T ca The method is as follows:

[0029] Based on the question, "How would users act if they were unaffected by attractions?", we designed scenarios for eliminating and increasing travel intention, assuming specific trajectories within these scenarios. for

[0030] Predicting user intent based on previous user activity patterns k :

[0031]

[0032] Wherein, pre(T) i-1 Predict the location where the user is most likely to go. k The choice is less attractive than and And the position with the highest predicted probability, c k As the initial intention of a user before being negatively influenced by attraction, when the consecutive check-in interval between two consecutive locations exceeds the user's average interval, a less attractive location c is introduced between those two consecutive locations. k The specific equation for achieving counterfactual data augmentation is shown below:

[0033]

[0034] Where Insert(·) represents generating a counterfactual trajectory unaffected by attraction through insertion, and ACI(T) represents the average check-in interval of the trajectory. Location c i-1 The multi-dimensional appeal of location c i The multi-dimensional appeal of location c k The multi-dimensional appeal of location c i Check-in time and location c i-1 The sign-in time; will T i As T ca .

[0035] Furthermore, in S2, the calculation The corresponding T ca The method is as follows:

[0036] Based on the question, "How would users act if they weren't influenced by attractions?", we designed a scenario that would increase travel intention, in which we hypothesized a trajectory. for

[0037] If two consecutive positions and If the time interval between the two locations is less than the user's average interval, and both locations belong to the strong attraction category, then the location... These are new travel intentions influenced by attraction. To restore the original user intent unaffected by attraction, location-based services should be removed. To achieve counterfactual data augmentation in this scenario, the specific equation is as follows:

[0038]

[0039] Where Delete(·) represents generating counterfactual augmentation data that is unaffected by negative attraction by deleting it; Td As T ca .

[0040] Furthermore, in S2, the calculation The corresponding T oa The method is as follows: Three methods from self-supervised learning are employed: masking, cropping, or reordering; the masking method involves random masking. Some check-in locations received T Mask The cutting method is to randomly select from... T is obtained from the pruning subsequence. Crop The reordering method, on the other hand, is to shuffle. The order of some check-in locations is used to obtain T. Reorder ;T oa =T Mask or T Crop or T Reorder Through these methods, we can comprehensively represent user characteristics from various perspectives, ultimately obtaining enhanced observational data T influenced by attraction. oa .

[0041] Furthermore, the S2 trains the WNA using a self-attention mechanism as the encoder for feature extraction. The embedding is represented as T ca The embedding is represented as T oa The embedding is represented as Where n is the length of the trajectory sequence and d is the dimension of the feature; construct three vectors. The attention score is calculated using the following equation:

[0042] Q = L(e ca W Q (12)

[0043] K = L(e ca W K (13)

[0044] V = L(e ca W V (14)

[0045] Where L(·) represents layer normalization, This represents the learnable parameter matrix. Let n represent the set of real numbers in the form n×d.

[0046] Calculate counterfactual user characteristics h ca As shown below:

[0047]

[0048]

[0049]

[0050] in It is a learnable parameter matrix, hd num It's about the number of heads, hd dim It is the dimension of each attention head, head i `Attention(.)` and `Attention(.)` are the attention heads, respectively. The attention mechanisms `MH(.)` and `MultiHead(.)` are both multi-head attention mechanisms. `Concat(.)`... Wo represents the feature concatenation operation, the last attention head, and the weight matrix.

[0051] Furthermore, the self-supervised learning loss during training of the WNA in S2... Calculated using the InfoNCE method:

[0052]

[0053] Where τ represents the temperature coefficient, a hyperparameter of self-supervised learning. Let represent the cosine similarity function shown in formula (7).

[0054] Furthermore, in S3, during the LPA training process, a neural network is used to learn the multidimensional attractiveness MA = (LA, CA, SA, TA) as a unified attractiveness index UA:

[0055] UA = f θ (MA) (19)

[0056] Among them, f θ (·) represents the embedding layer controlled by parameter θ, where θ includes the weight matrix and the bias vector.

[0057] After learning the unified attraction features, a two-stream attention mechanism is used to learn user features. One attention stream receives only user features and location embeddings, while the other receives user features and user agents (UAs). A bidirectional mask is used to fuse global attraction information. Finally, after utilizing positive attraction, the features learned from the two attention streams are fused to obtain the user features influenced by positive attraction. The process is as follows:

[0058]

[0059]

[0060]

[0061] Where Position represents location encoding information, γ is a hyperparameter that affects the attraction weights, and h A h represents the user characteristics obtained through a bidirectional mask. U This represents traditional user characteristics learned through a one-way mask. P To leverage user characteristics influenced by positive attraction, MH BiMask (.) and MH UniMask (.) represent the two-stream multi-head attention mechanism and the single-stream multi-head attention mechanism, respectively.

[0062] Furthermore, in S3, during the LPA training process, the user's actual check-in location at the next moment is recorded as c. pos Let c be the location of the randomly sampled negative sample. neg The recommended loss function is the binary cross-entropy function.

[0063]

[0064] Total loss as follows:

[0065]

[0066] Where λ1 and λ2 are the weights for controlling the recommendation loss and the self-supervised learning loss, respectively.

[0067] Compared with the prior art, the present invention has at least the following advantages:

[0068] 1. Unlike traditional recommendation systems, this invention considers the impact of bias on recommendations. By intervening in changes in attractiveness caused by bias, it employs counterfactual self-supervised learning to obtain unbiased user features. Furthermore, this invention uses a structural causal graph to capture the causal relationships behind user behavior, thereby mitigating the impact of relevance training on the model and generating richer and more accurate user features.

[0069] 2. The method of this invention extracts multidimensional attraction from location data and intervenes in attraction by means of structural causal graph analysis, thus constructing a unified debiasing paradigm. This solves the problem of model inconsistency caused by using different methods to address different biases.

[0070] 3. The method of this invention reduces excessively high attractiveness position weights through counterfactual self-supervised learning. Furthermore, it utilizes the positive influence of attractiveness to mitigate its negative impact, and then employs feature fusion to maintain the model's high-quality positional advantage features using this positive influence. Ultimately, a balance is achieved between the positive and negative influences of attractiveness, helping to alleviate bias while maintaining a balance in the impact of bias. Attached Figure Description

[0071] Figure 1 A structural causal graph describing the influence of attractiveness on recommendations is presented, where, for example... Figure 1 (a) represents the influence of attractiveness in traditional recommendation systems. Figure 1 (b) shows a structural cause-effect diagram to mitigate the negative impact of bias. Figure 1 (c) represents a structural causal graph that utilizes both positive and negative effects.

[0072] Figure 2 This diagram presents the overall framework of an interest point recommendation bias removal method based on causal graphs. The central section shows the overall model architecture. On the left, WNA utilizes counterfactual self-supervised learning to learn user features unaffected by negative attraction effects. On the right, LPA employs feature fusion to enhance the feature quality of high-quality locations.

[0073] Figure 3 Three scenarios where attractiveness influences user behavior.

[0074] Figure 4 This is a two-stream attention mechanism.

[0075] Figure 5 The experiment uses loss weight hyperparameters, where (a) and (b) show the results of the NDGC metric and the HR metric, respectively, for four different datasets under different loss weights.

[0076] Figure 6 The experiment is a hyperparameter experiment for the attraction index, where (a) and (b) are the results of the NDGC index and the HR index in four different datasets under different attraction indices, respectively. Detailed Implementation

[0077] The present invention will now be described in further detail.

[0078] The inventors believe that although the causes of bias differ, their ultimate result can be attributed to directly or indirectly affecting the attractiveness of a location to users. This attractiveness, in turn, influences users' location choices, leading to homogenized and undiversified recommendation results. By extracting this abstract attractiveness as concrete features for intervention, the problem can be addressed directly from the consequences of bias, thereby constructing a unified debiasing paradigm.

[0079] To analyze the causal relationship between attractiveness and other variables in the recommendation process, we constructed a structural causal graph as follows: Figure 1As shown in the diagram, we use nodes U, V, Z, and Y to represent the user, location, deviation, and probability of a user going to a certain location, respectively. Considering the characteristics of location data, we represent the attractiveness (A) of a location from four aspects: the location itself (P), location type (C), the location's spatial location (S), and the user's check-in time (T). The impact of attractiveness on traditional recommendation systems is as follows... Figure 1 As shown in (a), the edge Z→A represents the impact of biases such as the Matthew effect on attractiveness A, which makes popular locations even more popular and unpopular locations go unnoticed. There are two paths through which attractiveness A affects recommendation result Y. The edge A→U→Y reflects the influence of attractiveness on users, making them more inclined to choose popular, high-exposure, and mass-selected locations, ignoring their personal preferences, ultimately leading to homogenized recommendation results. The other edge A→Y represents that attractiveness directly affects the probability of a user going to a certain location. An attractive location naturally has superior features compared to other locations, and these features naturally increase the probability of a user choosing that location. Therefore, we should retain this path that benefits the recommendation service. To weaken the negative impact of attractiveness, A→U needs to be cut off... Figure 1 As shown in (b). Based on this, in order to utilize the positive attraction effect, the A→Y edge needs to be preserved as shown. Figure 1 As shown in (c). Finally, by cutting off A→U and keeping A→Y, we can construct a unified debiasing model that balances the positive and negative effects of attraction.

[0080] This invention proposes a bias removal method, CBRec, for point-of-interest (POI) recommendation based on causal graphs. This unified bias removal paradigm balances the positive and negative impacts of bias by intervening in location attractiveness. Based on mobile data characteristics, it extracts multi-dimensional attractiveness information of locations from four dimensions: location, type, space, and time. The model consists of two modules. The first module is the Weaken Negative Attraction (WNA) module. In this module, three different scenarios are established based on the counterfactual assumption of "how users would act if not influenced by attractiveness": replacing user travel intentions, deleting user travel intentions, and forcing the addition of new user travel intentions. Counterfactual data is generated based on these three scenarios, and self-supervised learning is used to compare the counterfactual data with observed data. This module enables the model to learn user characteristics even without the influence of attractiveness, helping to weaken the attractiveness impact of overheated locations and eliminate spurious correlations between attractiveness and user behavior. The second module is the Leverage Positive Attraction (LPA) module. After the first module weakens the negative impact of attractiveness, the influence of attractiveness in recommendation is as follows: Figure 1(c). By utilizing the edge between A and Y, the attraction feature is incorporated into the model training process using feature fusion operations, thus fusing the attraction feature with the user feature.

[0081] Let U = (u1, u2, ..., u) N Let L be the set of users, where N is the number of users. M Let M be the set of locations, and L be the number of locations. i =(id) i ,cat i ,lon i ,lat i This represents a detailed description of the i-th check-in location, including its ID, category, longitude, and latitude. User u i The travel trajectory is defined as Check-in location c for each trajectory j =(u i ,l j ,t j ) represents user u i At time t j Go to location l j Sign in.

[0082] The goal of the next point of interest recommendation task is to be based on user u i Historical trajectory The recommended position is c, which has the highest probability. n+1 The goal is to predict a user's most likely next location based on their past interaction history. To maximize the probability of predicting the next location, we minimize the actual next location c given the user's historical trajectory. n+1 The negative log-likelihood of . Let . Given the probability distribution of possible next positions, the optimization objective can be expressed as follows:

[0083]

[0084] Where θ represents the model parameters used for predicting the next location.

[0085] A bias removal method for interest point recommendation based on causal graphs, see [link to relevant documentation]. Figure 2 , Figure 2 The central section illustrates the overall model architecture. On the left, WNA utilizes counterfactual self-supervised learning to learn user features unaffected by negative attraction effects. On the right, LPA employs feature fusion to enhance the feature quality of high-quality locations. This includes the following steps:

[0086] S1: Obtain the user's historical trajectory, let U = (u1, u2, ..., u3) NLet L be the set of users, where N is the number of users, and L = (l1, l2, ..., ln). M Let M be the set of locations, and l be the number of locations. i =(id) i ,cat i ,lon i ,lat i ), representing a detailed description of the i-th check-in location, id i ,cat i ,lon i ,lat i These represent the ID, category, longitude, and latitude of the i-th check-in location, respectively; user u i The travel trajectory is defined as Check-in location c for each trajectory j =(u i ,l j ,t j ) represents user u i At time t j Go to location l j Sign in;

[0087] S2: Construct the WNA module to mitigate the negative impact of attraction on users, and calculate... Corresponding counterfactual enhancement data T, unaffected by negative attraction ca And the observational augmentation data T influenced by attraction oa In each batch of training, T generated by the same user will be used. ca With T oa As positive samples, the other users in this batch are used as negative samples to train WNA, when the self-supervised learning loss... Convergence indicates the end of training, and the trained WNA is obtained. The input WNA after training yields user features with reduced attractiveness; T generated from a user trajectory ca and T oa As positive samples, other users in the batch are designated as negative samples. After embedding the positive and negative samples into a vector space, the consistency among the positive samples is maximized to obtain user features unaffected by attractiveness. At the same time, we maximize the difference between positive and negative samples, thereby enhancing the model's ability to distinguish them and improving the overall generalization ability of the model.

[0088] During gradient descent optimization, WNA updates its parameters in a direction that minimizes the negative logarithm of the scores, effectively minimizing the similarity between positive and negative samples while maximizing the similarity between positive samples. This process iterates until convergence, ultimately yielding user features that mitigate the influence of attractiveness.

[0089] S3: Construct the LPA (Limited Approach Module) which utilizes the positive influence of attractiveness, and use the user features obtained in S2 that weaken the influence of attractiveness as input to train the LPA. During training, gradient descent is used to update the model parameters of both the LPA and the trained WNA. When the total loss... Convergence indicates the end of training, yielding the optimal LPA and optimal WNA.

[0090] S4: Obtain the trajectory of a new user, and extract user features h' from the trajectory using the feature extractors in the optimal LPA and optimal WNA. P , will h' P Embedded e in different locations p Matrix multiplication is performed, and the final value is used as the recommendation score for each location. The higher the recommendation score, the more likely the location is to be selected by new users. The locations are then sorted in descending order according to their recommendation scores to obtain the recommendation list.

[0091] Specifically, in S2, the calculation The corresponding T ca At that time, first Check-in location c j Classified as either highly attractive or weakly attractive locations;

[0092] A comprehensive description of the multidimensional attractiveness associated with each location is given by MA = (LA, CA, SA, TA), where LA is the inherent attractiveness of the current location, CA is the attractiveness associated with the location category, SA is the attractiveness derived from spatial features, and TA is the attractiveness caused by the location check-in time.

[0093]

[0094]

[0095]

[0096]

[0097] Where CTR(·) represents the function used to calculate the location click-through rate, d ij Location calculated based on latitude and longitude i and location l j The spatial distance between them, where δ is the threshold. For the indicator function, if condition d ij If δ is true, return 1; otherwise, return 0. (CTR) weekend (·) represents a function used to calculate weekend location click-through rates.

[0098] Specify different thresholds θ for the multidimensional attraction MAA =(θ l ,θ c ,θ s ,θ t If c j The corresponding MA>θ A Then c j Classified as a highly attractive location Otherwise, c j Location with low attraction The following is the equation for determining the location attraction:

[0099]

[0100] Where, θ l ,θ c ,θ s ,θ t These represent the thresholds for LA, CA, SA, and TA, respectively.

[0101] Specifically, in S2, the calculation The corresponding T ca The method is as follows:

[0102] Based on the question, "How would users act if they weren't influenced by the allure?", we design a scenario where excessive attraction leads to a shift in travel intentions, assuming a specific trajectory within this scenario. for in Indicates a more attractive location. express The next location is a weak attraction location, c n For the nth location

[0103] In the scenario of replacing travel intentions ( Figure 3 (Left side) The user originally intended to visit a low-attractive location. Under the influence of MA, users switched to a more attractive location. Calculate the multidimensional attractiveness of each location visited by the user. In a counterfactual construction unaffected by attractiveness, the user's intention to visit that location should remain constant. Therefore, and The similarity between them should be very high. Its appeal should be lower than The attraction of [the target]. Cosine similarity is used to calculate... and The formula for determining the similarity between them is as follows:

[0104]

[0105] Where E(·) represents the embedding method, used to map locations to a vector space. This represents the similarity function. In this scenario, after obtaining the location similarity metric, we can infer user trajectories under the counterfactual assumption of being unaffected by attractiveness: choosing a weakly attractive point similar to a strongly attractive point to replace the original location. The following is the equation for counterfactual data augmentation in the scenario of travel intention replacement:

[0106]

[0107] Replace(·) represents generating counterfactual trajectories that are unaffected by attraction by replacing them. for The multi-dimensional appeal of T r As T ca .

[0108] Specifically, in S2, the calculation The corresponding T ca The method is as follows:

[0109] Based on the question, "How would users act if they were unaffected by attractions?", we designed scenarios for eliminating and increasing travel intention, assuming specific trajectories within these scenarios. for

[0110] In scenarios where travel intention is eliminated (see...) Figure 3 (Central part), the location the user initially intended to access However, due to the location's low appeal, users lost interest in visiting it and ultimately did not go. In the counterfactual construction where attractiveness is not a factor, a user's travel intention is influenced only by personal preference, leading the user to reconsider visiting this weakly attractive location. When located in and The consecutive check-in interval between two locations exceeds the user's average interval, and both locations belong to category c. s When categorizing, we assume there's a point between these two points that is eliminated due to attraction. To find this point, we predict the user's intent c based on their previous trajectory. k :

[0111]

[0112] Wherein, pre(T) i-1 Predict the location where the user is most likely to go. k The choice is less attractive than and And the position with the highest predicted probability, c kAs the initial intention of a user before being negatively influenced by attraction, when the consecutive check-in interval between two consecutive locations exceeds the user's average interval, a less attractive location c is introduced between those two consecutive locations. k The specific equation for achieving counterfactual data augmentation is shown below:

[0113]

[0114] Where Insert(·) represents generating a counterfactual trajectory unaffected by attraction through insertion, and ACI(T) represents the average check-in interval of the trajectory. Location c i-1 The multi-dimensional appeal of location c i The multi-dimensional appeal of location c k The multi-dimensional appeal of location c i Check-in time and location c i-1 The sign-in time; will T i As T ca .

[0115] Specifically, in S2, the calculation The corresponding T ca The method is as follows:

[0116] Based on the question, "How would users act if they weren't influenced by attractions?", we designed a scenario that would increase travel intention, in which we hypothesized a trajectory. for

[0117] In scenarios where travel intentions increase ( Figure 3 (On the right), the user initially lacks travel intention, but is forced to consider visiting a highly attractive location due to strong appeal. Such a trip does not reflect the user's preferences and may adversely affect the model's effectiveness. In the counterfactual construction unaffected by appeal, we should delete the corresponding interaction record when the user lacks travel intention. If two consecutive locations... and If the time interval between the two locations is less than the user's average interval, and both locations belong to the strong attraction category, then the location... These are new travel intentions influenced by attraction. To restore the original user intent unaffected by attraction, location-based services should be removed. To achieve counterfactual data augmentation in this scenario, the specific equation is as follows:

[0118]

[0119] Where Delete(·) represents generating counterfactual augmentation data that is unaffected by negative attraction by deleting it; T dAs T caz .

[0120] The counterfactual hypothesis proposed in this article is: "How would users behave if they were not influenced by attraction?" This hypothesis... Figure 1 Node U in (b) has an impact that can cut off the A→U transition at the counterfactual level, mitigating the negative impact of attractiveness. Based on this assumption, this invention designs three scenarios and generates counterfactual data based on these scenarios. These three scenarios include travel intention replacement caused by excessive attractiveness, travel intention elimination, and travel intention increase. Figure 3 This demonstrates the impact of attractiveness on user behavior. For ease of subsequent explanation, we assume the trajectory is as follows in the following scenario. T is generated based on three scenarios. r T i and T d Three types of counterfactual enhancement data unaffected by negative attraction, T ca Take T r T i or T d .

[0121] We obtained comparative samples for self-supervised learning by employing data augmentation methods in three counterfactual scenarios. By maximizing the consistency between counterfactual data and observed data, we make user characteristics more closely resemble counterfactual environments unaffected by attraction, thereby mitigating the adverse effects of attraction.

[0122] Specifically, in S2, the calculation The corresponding T oa The method is as follows: Three methods from self-supervised learning are employed: masking, cropping, or reordering; the masking method involves random masking. Some check-in locations received T Mask The cutting method is to randomly select from... T is obtained from the pruning subsequence. Crop The reordering method, on the other hand, is to shuffle. The order of some check-in locations is used to obtain T. Re order ;T oa =T Mask or T Crop or T Re order Through these methods, we can comprehensively represent user characteristics from various perspectives, ultimately obtaining enhanced observational data T influenced by attraction. oa .

[0123] Specifically, the S2 trains the WNA using a self-attention mechanism as the encoder for feature extraction. The embedding is represented as T ca The embedding is represented as Toa The embedding is represented as Where n is the length of the trajectory sequence and d is the dimension of the feature; construct three vectors. The attention score is calculated using the following equation:

[0124] Q = L(e ca W Q (12)

[0125] K = L(e ca W K (13)

[0126] V = L(e ca W V (14)

[0127] Where L(·) represents layer normalization, This represents the learnable parameter matrix. Let n represent the set of real numbers in the form n×d.

[0128] Calculate counterfactual user characteristics h ca As shown below

[0129]

[0130]

[0131]

[0132] in It is a learnable parameter matrix, hd num It's about the number of heads, hd dim It is the dimension of each attention head, head i `Attention(.)` and `Attention(.)` are the attention heads, respectively. The attention mechanisms `MH(.)` and `MultiHead(.)` are both multi-head attention mechanisms. `Concat(.)`... Wo represents the feature concatenation operation, the last attention head, and the weight matrix.

[0133] Specifically, we will represent the features learned from counterfactual augmented trajectory data and the features learned from observed trajectory data as h, respectively. ca and h oa Paired features (h) ca ,h oa We consider other features h in the batch as positive samples to maximize consistency. j Negative samples are considered to maximize their inconsistency with positive samples. The self-supervised learning loss used when training the WNA in S2 is described. Calculated using the InfoNCE method:

[0134]

[0135] Where τ represents the temperature coefficient, a hyperparameter of self-supervised learning. Let represent the cosine similarity function shown in formula (7).

[0136] Specifically, the LPA module combines attractiveness features extracted from location, category, spatial, and temporal dimensions with negative attractiveness features weakened by the WNA module. Through an attention mechanism, it leverages the positive effects of attractiveness. Ultimately, during training, the WNA and LPA modules mutually constrain each other, achieving a balance between the positive and negative effects of attractiveness. The architecture of LPA is as follows: Figure 2 As shown on the right.

[0137] To leverage the positive impact of location attractiveness, we integrate location attractiveness embeddings into user feature embeddings during training. This integration helps enrich user features with location attractiveness information, thereby preserving high-quality location features.

[0138] In S3, during the LPA training process, a neural network is used to learn the multidimensional attractiveness MA = (LA, CA, SA, TA) as a unified attractiveness index UA.

[0139] UA = f θ (MA) (19)

[0140] Among them, f θ (·) represents the embedding layer controlled by parameter θ, where θ includes the weight matrix and the bias vector.

[0141] After learning the characteristics of unified attraction, a two-stream attention mechanism is adopted, such as... Figure 4 As shown, we learn user features. In this mechanism, we use different attention streams to learn traditional user features and user features that include attraction. One attention stream only receives user features and location embeddings, and we use a one-way mask to hide information after the current timestamp, representing a traditional user analysis and recommendation task. Simultaneously, the other attention stream receives user features and user agents (UAs), and uses a two-way mask to fuse global attraction information. Finally, after utilizing positive attraction, the features learned by the two attention streams are fused to obtain the user features influenced by positive attraction, as shown below:

[0142]

[0143]

[0144]

[0145] Where Position represents location encoding information, γ is a hyperparameter that affects the attraction weights, and h A This represents user characteristics obtained through bidirectional masking, used to capture global attractiveness information, h U This represents traditional user characteristics learned through a one-way mask. P To leverage user characteristics influenced by positive attraction, MH BiMask (.) and MH UniMask (.) represent the two-stream multi-head attention mechanism and the single-stream multi-head attention mechanism, respectively.

[0146] Specifically, in S3, during the LPA training process, after integrating the WNA and LPA modules, the model ultimately obtains user features that balance the positive and negative attraction effects. These features are then fed into the prediction layer to recommend locations to users. To enhance the model's generalization ability, we also randomly sample negative samples to improve its robustness. The actual location where the user checks in at the next moment is denoted as c. pos Let c be the location of the randomly sampled negative sample. neg The recommended loss function is the binary cross-entropy function.

[0147]

[0148] During the training of CBRec, we divided it into two tasks: one focusing on passing through One way to achieve a unified de-biasing of the negative impact on attractiveness is through... We utilize positive attraction for location recommendation. To achieve a balance in the impact of attraction, we employ a multi-task strategy to collaboratively optimize these tasks, with a total loss... as follows:

[0149]

[0150] Where λ1 and λ2 are the weights for controlling the recommendation loss and the self-supervised learning loss, respectively.

[0151] Experiments and Analysis

[0152] 1 Dataset

[0153] Experiments were conducted using four real-world location datasets from the FourSquare platform: NYC, TKY, SG, and GB. FourSquare is a mobile service website centered on user geolocation information. The NYC and TKY datasets recorded user check-in behavior in New York City (NYC) and Tokyo, Japan (TKY) from April 2012 to February 2013. Additionally, we partitioned the global user check-in dataset recorded from April 2012 to September 2013, separating user check-in data from Singapore (SG) and the United Kingdom (GB) by country.

[0154] To maintain data quality during the experiment, we employed a filtering criterion to exclude users with fewer than 10 check-in records and locations with fewer than 10 visitors. After filtering, the statistics for each dataset are shown in Table 1, including the number of users, number of locations, location category, number of check-ins, and average trajectory length.

[0155] Table 1 Summary of Data Statistics

[0156]

[0157] 2 baselines

[0158] To evaluate the effectiveness of the method of this invention, we compare it with excellent sequence recommendation models in recent years. Furthermore, to further verify the impact of the model's balancing attractiveness, we compare it with debiased sequence recommendation models and data-augmented sequence recommendation models. The selected baselines are as follows.

[0159] (1) Sequence Recommendation Model

[0160] SASRec (ICDM18): A sequence recommendation model that effectively captures long-term semantics and balances model complexity by using self-attention.

[0161] SSEPT20 (RecSys20): A personalized Transformer model designed to enhance sequential recommendations by incorporating personalized user embeddings.

[0162] TiSASRec(WSDM20): A self-attention-based time-interval-aware sequence recommendation system that explicitly models interaction timestamps and explores the impact of time intervals on the prediction of the next item.

[0163] AC-TSR (CIKM23): A framework that addresses the inaccuracy of attention weights in Transformer-based sequence recommendation models. It introduces a spatial calibrator to capture the spatial relationships between items and an adversarial calibrator to redistribute attention weights.

[0164] (2) Partially de-biased sequence recommendation model

[0165] PDA (CIKM21): It addresses the popularity bias problem in recommender systems by leveraging beneficial popularity patterns while mitigating bias. It eliminates bias during training and adjusts recommendation scores during the inference phase.

[0166] CauseRec (SIGIR21): Enhances recommender systems by learning robust user representations through modeling counterfactual data distributions.

[0167] (3) Data-augmented sequence recommendation model

[0168] Cl4SRec (ICDE22): This is an advanced framework for sequence recommendation that improves user representation by combining traditional prediction tasks with contrastive learning.

[0169] DuoRec (WSDM22): Improves sequence recommendation models by addressing representation degradation through novel contrastive regularization and model-level enhancements.

[0170] FEARec (SIGIR23): This model enhances sequence recommendation by introducing a frequency-enhanced hybrid attention network, addressing the limitations of capturing high-frequency information and periodic features by improving both the time and frequency domains.

[0171] 3 Evaluation Indicators

[0172] We use the commonly used evaluation metrics NDCG@K and HR@K in current recommendation systems to evaluate the quality of the model, where K takes values ​​of 5, 10, and 20.

[0173] Normalized Discounted Cumulative Gain (NDCG@K) is a commonly used evaluation metric in recommender systems to measure the quality of item ranking in a recommendation list. It normalizes the cumulative gain by considering the relevance of recommended items at different positions and the impact of their arrangement on user satisfaction. Here, rel... i The correlation of the i-th item is represented by DCG, which represents the cumulative loss gain. The calculation formula is as follows:

[0174]

[0175] The final NDCG expression is as follows:

[0176]

[0177] Wherein, IDCG@K represents the ideal depreciation cumulative gain based on descending correlation ranking.

[0178] The hit rate (HR@K) is defined as the ratio of the number of correctly predicted samples to the total number of samples in the predicted results list. Essentially, it measures the consistency between the recommended items and user preferences, emphasizing the accuracy of the prediction process. Its calculation formula is as follows:

[0179]

[0180] Here, hit(i) indicates whether the recommendation result hits user i, where 1 indicates a hit and 0 indicates a miss.

[0181] 4 Implementation Details

[0182] In the experiments, the model of this invention was built using the PyTorch framework. We set the learning rate to 0.001, the embedding dimension to 50, and the batch size to 128 samples. Given the varying trajectory lengths across different datasets, we set a trajectory length threshold of 120. Trajectories exceeding this threshold were ignored, while those below were padded with zeros to reach 120. We used the Adam optimizer to iteratively optimize the model parameters.

[0183] In WNA, we use a temperature coefficient τ = 0.01. When generating counterfactual augmentation data, we need to manipulate the original trajectories: replace travel intentions, remove travel intentions, and add travel intentions, with 30% of the trajectories remaining unchanged. The weights of the recommendation loss and self-supervised learning loss are used in the hyperparameter experiments, respectively: {λ1 = 0.2, λ2 = 0.8}, {λ1 = 0.4, λ2 = 0.6}, {λ1 = 0.6, λ2 = 0.4}, and {λ1 = 0.8, λ2 = 0.2}.

[0184] In LPA, we integrate four dimensions: location, category, space, and time to aggregate attraction data. Furthermore, we conducted hyperparameter experiments on the attraction index with γ values ​​set to {1, 0.1, 0.01, 0.001}.

[0185] 5 Results Analysis

[0186] (1) Recommended performance

[0187] Table 2 presents the comparative experimental results of the CBRec method with several high-quality methods in recent years on four real-world datasets. Bold text indicates the best performance data for each metric, while underlined text indicates the second-best performance data for each metric. Table 2 clearly demonstrates the superior performance of CBRec across all evaluation metrics on the four datasets. Notably, CBRec shows a more significant improvement on the NYC and TKY datasets, with an average improvement of approximately 3%. Although CBRec also achieves best performance on the SG and GB datasets, the improvement is smaller. This may be due to the shorter average user trajectories in the SG and GB datasets, and the relatively less active users compared to the NYC and TKY datasets, resulting in a smaller bias effect.

[0188] Table 2 shows the comparative experiments on the NYC, TKY, SG, and GB datasets.

[0189]

[0190] (2) Effects of different components

[0191] To assess the effectiveness of each component in the invention, we divided CBRec into two components: the WNA module, which only mitigates the negative effects of attraction, and the LPA module, which only utilizes the positive effects of attraction. The experimental results are shown in Table 3. Bold text indicates the best performance for each metric, underlined text indicates the second-best performance for each metric, and up arrows mark results where a component outperforms the overall model.

[0192] The module that only mitigates the negative impact of attractiveness does not utilize the positive impact of attractiveness. Guided by the hypothesis "how would users behave if they were unaffected by attractiveness," this module proposes three scenarios involving replacement, elimination, and addition of user intent to generate counterfactual data. In the ablation experiments, we designed four cases: using only one scenario or using all three scenarios simultaneously. As shown in Table 3, when the positive impact of attractiveness is not utilized, the performance of all CBReces degrades across all datasets and all metrics, highlighting the positive effect of utilizing positive attractiveness on recommendation results. In the comparison of different scenarios, only the scenario utilizing all three scenarios achieved optimal performance on the NYC dataset. In most models on the other three datasets, the scenario utilizing only user intent replacement yielded better results. We argue that, compared to the other two scenarios, only replacing trajectories ensures that user characteristics do not change significantly, leading to better recommendation performance. However, sacrificing some accuracy is necessary to construct comprehensive counterfactual data.

[0193] Building only the attractiveness module to positively influence the model did not mitigate the negative effects of attractiveness. This component extracts multidimensional attractiveness from location, category, space, and time, and integrates this attractiveness feature into user features to leverage the model's positive effects. In ablation experiments, we compared the results of using only a single-dimensional attractiveness with using unified attractiveness. As shown in Table 3, the performance of all metrics in the CBRec dataset degrades when the effects of negative attractiveness are not eliminated, demonstrating the necessity of mitigating negative effects to achieve enhanced recommendation results. Comparative analysis shows that, except for the TKY dataset, unified attractiveness features perform best on most metrics in other datasets. For example, on the NYC dataset, using only location or spatial attractiveness can also produce excellent results. However, it is unrealistic to determine which form of attractiveness yields the best results across different datasets. Therefore, utilizing unified attractiveness features is a more practical approach.

[0194] Table 3 Ablation Experiment Results of Different Modules

[0195]

[0196] (3) The influence of different hyperparameters

[0197] Hyperparameter experiments were conducted, covering two experiments: loss weight and attraction index.

[0198] As shown in Equation (24), our final loss consists of two parts: recommendation loss and self-supervised learning loss. The weights of these two losses are denoted as λ1 and λ2, and were explored experimentally, testing four sets of values: λ1 = 0.8, λ2 = 0.2; λ1 = 0.6, λ2 = 0.4; λ1 = 0.4, λ2 = 0.6; λ1 = 0.2, λ2 = 0.8. The final result is as follows: Figure 5 As shown, the results of the second and third groups are similar in the NDCG@20 metric, both achieving the best performance. However, the second group clearly performs better in the HR@20 metric, so we ultimately chose the combination of λ1 = 0.6 and λ2 = 0.4 for model training.

[0199] When leveraging the positive impact of attraction, we integrate the unified attraction feature into user characteristics after applying an index calculation. To determine the attraction index γ, we conducted experiments with γ taking values ​​of 1, 0.1, 0.01, and 0.001. Figure 6 The results show that when γ = 0.01, the results are unstable, while the effects are similar when γ is 0.1 or 0.001. Considering the overall experimental results, we chose a value of 0.001 for γ.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A bias removal method for interest point recommendation based on causal graphs, characterized in that: Includes the following steps: S1: Obtain the user's historical trajectory, set... For a set of users, where For the number of users, To meet at the location, For the number of locations, This represents a detailed description of the i-th check-in location. These represent the ID, category, longitude, and latitude of the i-th check-in location, respectively; user The travel trajectory is defined as , Check-in locations for each track On behalf of users In time Go to the location Sign in; S2: Construct the WNA (Warning Appeal Negativity) module to mitigate the negative impact of attraction on users and calculate... Corresponding counterfactual enhancement data unaffected by negative attraction And enhanced observational data influenced by attraction In each training batch, the data will be generated by the same user. and As positive samples, the other users in this batch are used as negative samples to train WNA, when the self-supervised learning loss... Convergence indicates the end of training, and the trained WNA is obtained. The input of trained WNA yields user characteristics with reduced attractiveness influence; S3: Construct the LPA (Limited Approach Module) which utilizes the positive influence of attractiveness, and use the user features obtained in S2 that weaken the influence of attractiveness as input to train the LPA. During training, gradient descent is used to update the model parameters of both the LPA and the trained WNA. When the total loss... Convergence indicates the end of training, yielding the optimal LPA and optimal WNA. S4: Obtain the trajectory of a new user, and extract user features from the trajectory using the feature extractors in the optimal LPA and optimal WNA. ,Will Embedded in different locations Matrix multiplication is performed, and the final value is used as the recommendation score for each location. The higher the recommendation score, the more likely the location is to be selected by new users. The locations are then sorted in descending order according to their recommendation scores to obtain the recommendation list.

2. The interest point recommendation bias removal method based on causal graph as described in claim 1, characterized in that: In S2, the calculation corresponding At that time, first Sign-in location Classified as either highly attractive or weakly attractive locations; A comprehensive description of the multidimensional attractions associated with each location. , Due to the inherent appeal of the current location, For attractiveness associated with location categories, Attraction stemming from spatial features, The attraction stems from the location's check-in time; in, This represents a function used to calculate location click-through rate. Locations calculated based on latitude and longitude and location Spatial distance between them For the threshold, For indicator functions, if the condition < If true, return 1; otherwise, return 0. This represents a function used to calculate weekend location click-through rates; For multidimensional attraction Specify different thresholds ;if Correspondingly Then Classified as a highly attractive location ;otherwise, Location with low attraction The following is the equation for determining the location attraction: in, They represent The threshold.

3. The bias removal method for interest point recommendation based on causal graphs as described in claim 2, characterized in that: In S2, the calculation corresponding The method is as follows: Based on the question, "How would users act if they weren't influenced by the allure?", a scenario is designed to explore how excessive attraction can lead to a shift in travel intentions. In this scenario, a hypothetical trajectory is assumed. for ,in Indicates a more attractive location. express The next location is a weak attraction location. This is the nth location; Using cosine similarity calculation and The formula for determining the similarity between them is as follows: in This represents an embedding method used to map locations to a vector space. The similarity function is represented below; the equation for counterfactual data augmentation in the scenario of travel intention replacement is as follows: in The representative generates counterfactual trajectories unaffected by attraction by replacing [the data]. for Its multi-dimensional appeal will As .

4. The bias removal method for interest point recommendation based on causal graphs as described in claim 2, characterized in that: In S2, the calculation corresponding The method is as follows: Based on the question, "How would users act if they were unaffected by attractions?", we designed scenarios for eliminating and increasing travel intention, assuming a specific trajectory within each scenario. for ; Predicting user intent based on previous user behavior : in, Predict the location where users are most likely to go. Choosing an attractive option is less than and And the position with the highest prediction probability As the initial intention before a user is negatively influenced by an attractive location, if the consecutive check-in interval between two consecutive locations exceeds the user's average interval, then a less attractive location is introduced between those two consecutive locations. The specific equation for achieving counterfactual data augmentation is shown below: in This indicates that counterfactual trajectories unaffected by attraction are generated through insertion. This represents the average check-in time interval of the trajectory. , , , , Locations Multi-dimensional appeal, location Multi-dimensional appeal, location Multi-dimensional appeal, location Check-in time and location The sign-in time; As .

5. The interest point recommendation bias removal method based on causal graph as described in claim 2, characterized in that: In S2, the calculation corresponding The method is as follows: Based on the question, "How would users act if they weren't influenced by attractions?", we designed a scenario that would increase travel intention, in which we hypothesized a trajectory. for ; If two consecutive positions and If the time interval between the two locations is less than the user's average interval, and both locations belong to the strong attraction category, then the location... These are new travel intentions influenced by attractiveness; to restore the original user intent unaffected by attractiveness, location is removed. To achieve counterfactual data augmentation in this scenario, the specific equation is as follows: in This indicates that counterfactual augmentation data, which is unaffected by negative attraction, is generated by removing it; As .

6. A bias removal method for interest point recommendation based on causal graphs as described in any one of claims 2-5, characterized in that: In S2, the calculation corresponding The method is as follows: Three methods from self-supervised learning are employed: masking, cropping, or reordering; the masking method involves random masking. Some of the check-in locations received The cutting method is to randomly select from... The middle cut subsequence is obtained The reordering method, on the other hand, is to shuffle. The order of some check-in locations was obtained ; These methods represent user characteristics from various perspectives, ultimately yielding augmented observation data influenced by attractiveness. .

7. The interest point recommendation bias removal method based on causal graph as described in claim 6, characterized in that: The S2 training WNA uses a self-attention mechanism as the encoder for feature extraction. The embedding is represented as ,Will The embedding is represented as ,Will The embedding is represented as ,in It is the length of the trajectory sequence. It is the dimension of the feature; Construct three vectors The attention score is calculated using the following equation: in Representation layer normalization, This represents the learnable parameter matrix. express The set of real numbers; Calculate counterfactual user characteristics As shown below: in , It is a learnable parameter matrix. It's about the number of heads to focus on. It is the dimension of each attention head. and These are the attention head and the attention mechanism, respectively. and All are bullish attention. , , What do they represent respectively? Feature concatenation operation, the last attention head, and the weight matrix? 8. The interest point recommendation bias removal method based on causal graph as described in claim 7, characterized in that: The self-supervised learning loss during training of WNA in S2 is mentioned above. Calculated using the InfoNCE method: in The temperature coefficient represents the hyperparameter of self-supervised learning. Let represent the cosine similarity function shown in formula (7).

9. The interest point recommendation and bias removal method based on causal graphs as described in claim 8, characterized in that: In S3, during the LPA training process, a neural network is used to learn multidimensional attraction. For a unified attractiveness index : in, Indicates by parameters The embedded layer of control, and Includes the weight matrix and the bias vector; After learning the unified attraction features, a two-stream attention mechanism is used to learn user features. One attention stream receives only user features and location embeddings, while the other attention stream receives user features and A bidirectional mask is used to fuse global attraction information; finally, after utilizing positive attraction, the features learned from the two attention streams are fused to obtain user features influenced by positive attraction, as shown below: in, Indicates location encoding information, It is a hyperparameter that affects the attractiveness weight. This represents the user characteristics obtained through a two-way mask. This represents traditional user characteristics learned through a one-way mask. To leverage user characteristics influenced by positive attraction, and These are the two-stream multi-head attention mechanism and the single-stream multi-head attention mechanism, respectively.

10. The interest point recommendation and bias removal method based on causal graphs as described in claim 9, characterized in that: In S3, during the LPA training process, the user's actual check-in location at the next moment is recorded as... The locations of the randomly sampled negative samples are denoted as We recommend using the binary cross-entropy function to calculate the loss. Total loss as follows: in, and These are the weights for controlling the recommendation loss and the self-supervised learning loss, respectively.

Images