Cross-domain interest point recommendation method and system based on double prototype alignment and reinforcement learning

Abstract

The application relates to the field of point of interest recommendation, and provides a cross-domain point of interest recommendation method and system based on double prototype alignment and reinforcement learning, which comprises the following steps: inputting point of interest text data into a large language model to obtain preprocessed text data; encoding to obtain an encoding vector and calculating a clustering classification loss to obtain a clustering center; obtaining a state space according to the clustering center; inputting the state space into an action network to obtain a residual offset, estimating through a value network to obtain a cumulative advantage parameter; calculating an action probability distribution, using the action probability distribution and the cumulative advantage parameter to calculate action network optimization parameters, inputting the state space into the value network to obtain a truncated cumulative value, and obtaining value network optimization parameters according to the truncated cumulative value; updating the action network and the value network through the action network optimization parameters and the value network optimization parameters to obtain a target action network, and outputting a recommended point of interest; and the application can be used for point of interest recommendation.

Description

Technical Field

[0001] This invention relates to the field of interest point recommendation technology, and in particular to a cross-domain interest point recommendation method and system based on dual prototype alignment and reinforcement learning. Background Technology

[0002] With the booming development of mobile internet, point-of-interest (POI) recommendation has become a core technology for improving user experience and promoting offline consumption. Traditional POI recommendation mainly focuses on a single city or specific region; however, in real-world scenarios, users' activities often span geographical areas. In this situation, systems often face severe data sparsity and cold-start problems, meaning that user behavior records are lacking or nonexistent in the target domain. Therefore, effectively transferring user behavioral preferences accumulated in the source domain to the target domain to achieve cross-domain POI recommendation is crucial for building seamless personalized services.

[0003] Current single-domain interest point recommendation methods primarily rely on collaborative filtering, deep neural networks, and graph neural networks to capture users' spatiotemporal preferences. While these methods have achieved significant results in scenarios with abundant data, they are highly dependent on the distribution of data within the domain. When users enter a completely new environment, the generalization ability of single-domain models rapidly declines due to the drastic changes in the underlying feature space. Furthermore, single-domain methods often struggle to capture higher-order preference logic common across different regions, making them inadequate for handling user migration behavior.

[0004] To address the aforementioned issues, existing research on cross-domain interest point recommendation attempts to achieve feature transfer through embedding alignment or mapping networks. However, current research still faces the following challenges: Fragmentation of Points of Interest (POIs): In the real world, POIs contain multi-dimensional semantic information, including both function-oriented category labels and fine-grained descriptive attributes. This discreteness and fragmentation of information makes it difficult for traditional models to capture deep semantic logic in complex feature spaces, thereby increasing the difficulty of cross-domain knowledge alignment.

[0005] Neglecting negative preferences: In the real environment, the points of interest that users have visited may not reflect their true preferences. Most cross-domain models only focus on what users "like" when extracting transferable user profiles, but ignore the key role of "what they dislike" in defining the boundaries of user preferences.

[0006] The diversity of user preferences: In the real world, user interests exhibit extremely high heterogeneity, and their preferences are often influenced by multiple factors such as personal habits, cultural background, and immediate needs. This highly complex and individualized preference characteristic makes it difficult for models to construct robust user profiles from limited interaction data, thus limiting the accuracy of recommendation systems in cross-domain scenarios. Summary of the Invention

[0007] This invention aims to at least address one of the technical problems existing in related technologies. To this end, this invention provides a method and system for cross-domain interest point recommendation based on dual-prototype alignment and reinforcement learning, enabling cross-domain interest point recommendation.

[0008] This invention provides a cross-domain interest point recommendation method based on dual-prototype alignment and reinforcement learning, including: S1: Obtain the text data of points of interest, select a large language model, input the text data of points of interest into the large language model, and obtain preprocessed text data; Among them, the large language model scores the sentiment preference of interest points, and obtains preprocessed text data including preprocessed user preference prototypes and preprocessed user dislike prototypes; the preprocessed user preference prototypes and preprocessed user dislike prototypes constitute a dual prototype. S2: Encode the preprocessed text data to obtain the encoded vector, calculate the clustering classification loss of the encoded vector, and obtain the cluster center based on the clustering classification loss; S3: Obtain the state space based on the cluster center, construct the action network, input the state space into the action network to obtain the residual offset, construct the value network, and estimate the cumulative advantage parameter through the value network; S4: Calculate the action probability distribution using the residual offset, use the action probability distribution and cumulative advantage parameters to calculate the action network optimization parameters, input the state space into the value network and perform truncated cumulative value calculation to obtain the truncated cumulative value, and obtain the value network optimization parameters based on the truncated cumulative value; S5: Update the action network and value network by optimizing the parameters of the action network and value network to obtain the target action network. Then, output the recommended interest points through the target action network to complete the interest point recommendation.

[0009] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S2, a text encoder is constructed, and the preprocessed text data is encoded by the text encoder to obtain the encoded vector.

[0010] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S2, the uniform loss and separation loss of the encoding vector are calculated, the interest point weights are calculated according to the encoding vectors, the alignment loss is obtained through the interest point weights, and the uniform loss, alignment loss and separation loss are weighted and summed to obtain the clustering classification loss.

[0011] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S3, the preference cluster center is obtained from the cluster center, the state space is input into the action network to obtain the action network output, and the residual offset is obtained according to the preference cluster center and the action network output.

[0012] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S3, the state space is input into the value network to obtain the temporal difference error, and the cumulative advantage parameter is obtained through the temporal difference error.

[0013] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S4, the action probability distribution including the updated action probability distribution and the original action probability distribution is calculated by the residual offset, the ratio of the new and old probability distributions is obtained by using the action probability distribution, and the action network optimization parameters are obtained by using the ratio of the new and old probability distributions and the cumulative advantage parameters.

[0014] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S4, the state space is input into the value network to obtain an advantage estimate including an advantage estimate before the update and an advantage estimate after the update. The truncated cumulative value is obtained through the advantage estimate. The target gain is obtained using the advantage estimate before the update and the cumulative advantage parameter. The value network optimization parameters are obtained through the target gain and the truncated cumulative value.

[0015] According to the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by the present invention, in step S5, target optimization parameters are constructed through the action network optimization parameters and the value network optimization parameters, and the action network and value network are updated using the target optimization parameters to obtain the target action network.

[0016] This invention also provides a cross-domain interest point recommendation system based on dual prototype alignment and reinforcement learning, including: Text preprocessing module: used to acquire text data of points of interest, select a large language model, input the text data of points of interest into the large language model, and obtain preprocessed text data; Among them, the large language model scores the sentiment preference of interest points, and obtains preprocessed text data including preprocessed user preference prototypes and preprocessed user dislike prototypes; the preprocessed user preference prototypes and preprocessed user dislike prototypes constitute a dual prototype. Clustering classification module: used to encode preprocessed text data to obtain encoded vectors, calculate the clustering classification loss of the encoded vectors, and obtain cluster centers based on the clustering classification loss; Cumulative Advantage Parameter Module: This module is used to obtain the state space based on the cluster center, construct the action network, input the state space into the action network to obtain the residual offset, construct the value network, and estimate the cumulative advantage parameter through the value network. The network optimization parameter module is used to calculate the action probability distribution through residual offset, calculate the action network optimization parameters using the action probability distribution and cumulative advantage parameters, input the state space into the value network and perform truncated cumulative value calculation to obtain the truncated cumulative value, and obtain the value network optimization parameters based on the truncated cumulative value. Interest point recommendation module: It is used to update the action network and value network by optimizing the parameters of the action network and the value network to obtain the target action network, and then output recommended interest points through the target action network to complete the interest point recommendation.

[0017] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: This invention provides a cross-domain interest point recommendation method and system based on dual-prototype alignment and reinforcement learning. It effectively alleviates data sparsity and cold-start problems by using a large language model to perform semantic understanding and sentiment preference scoring on interest point texts. By utilizing dual-prototype alignment of user preferences and dislikes, and designing a clustering classification loss that integrates uniformity, alignment, and separation, the encoded vectors form a compact and distinguishable cluster structure in the semantic space, thereby more accurately characterizing user interest boundaries and improving the accuracy and personalization of recommendations. Simultaneously, the cluster centers are constructed as a state space, and the recommendation strategy is adaptively adjusted using the residual offset output by the action network. Furthermore, the cumulative advantage parameter is estimated and the cumulative value is truncated through a value network, and a new-to-old probability ratio truncation mechanism limits the policy update magnitude, enhancing the stability and convergence of the training process. The synergistic optimization of the action network and value network achieves a good balance between exploration and utilization, ultimately significantly improving the overall performance, robustness, and long-term benefits of interest point recommendation.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1This is a flowchart illustrating the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning provided by this invention.

[0021] Figure 2 This is a schematic diagram of the cross-domain interest point recommendation system based on dual prototype alignment and reinforcement learning provided by the present invention.

[0022] Figure label: 100. Text preprocessing module; 200. Cluster classification module; 300. Cumulative advantage parameter module; 400. Network optimization parameter module; 500. Interest point recommendation module. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0024] In the description of the embodiments of the present invention, it should be noted that the terms "first", "second" and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention based on the specific circumstances.

[0026] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0027] The following is combined with Figures 1 to 2Specific embodiments of the present invention are described below. Figure 1 This is a flowchart illustrating the cross-domain interest point recommendation method based on dual-prototype alignment and reinforcement learning provided by the present invention, including: S1: Obtain the text data of points of interest, select a large language model, input the text data of points of interest into the large language model, and obtain preprocessed text data; Among them, the large language model scores the sentiment preference of interest points, and obtains preprocessed text data including preprocessed user preference prototypes and preprocessed user dislike prototypes; the preprocessed user preference prototypes and preprocessed user dislike prototypes constitute a dual prototype. Furthermore, the objective of this stage is to obtain preprocessed text data. The specific implementation steps in this embodiment are as follows: First, we need to acquire text data of users' historical comments on various points of interest (POIs) and text data describing the POIs themselves. This data is referred to as POI text data. The text data describing the POIs includes their attributes and descriptions. Then, we select an existing large language model and input the POI text data into it. The large language model, based on pre-written prompts, can score the sentiment preference of each POI based on the text data of users' comments. When scoring the sentiment preference of each POI, if a user has not commented on a particular POI, their sentiment preference is obtained from the text data of their comments on other similar POIs.

[0028] Here, a higher sentiment preference score indicates that the user likes the point of interest more, and vice versa. This allows us to use a preset sentiment preference score threshold to preprocess user preference prototypes for points with sentiment preference scores above the threshold, and user dislike prototypes for points with sentiment preference scores below the threshold. Concatenating the interest point text data, the preprocessed user preference prototypes, and the preprocessed user dislike prototypes yields the preprocessed text data. These preprocessed user preference and dislike prototypes constitute a dual prototype.

[0029] S2: Encode the preprocessed text data to obtain the encoded vector, calculate the clustering classification loss of the encoded vector, and obtain the cluster center based on the clustering classification loss; Furthermore, the objective of this stage is to calculate the clustering classification loss of the encoded vector and obtain the cluster centers based on the clustering classification loss. Specifically, in step S2, a text encoder is constructed, and the preprocessed text data is encoded using the text encoder to obtain the encoded vector.

[0030] In step S2, the uniform loss and separation loss of the encoding vector are calculated, the interest point weights are calculated based on the encoding vector, the alignment loss is obtained through the interest point weights, and the uniform loss, alignment loss, and separation loss are weighted and summed to obtain the clustering classification loss.

[0031] The specific implementation method for the above steps in this embodiment is as follows: First, a transformer-based text encoder needs to be built. This encoder encodes the preprocessed text data to obtain encoded vectors. Specifically, encoding the preprocessed user preference prototype yields the encoded user preference prototype, and encoding the preprocessed user dislike prototype yields the encoded user dislike prototype. The encoded vectors also include encoded interest point text data, i.e., interest point vectors. Next, the uniform loss of the encoded vectors needs to be calculated. : in, This indicates the expected value of the expression in parentheses. This represents the softplus activation function. This represents the encoded user preference prototype of the i-th user. Let T represent the encoded user non-preference prototype of the i-th user, and let T denote the transpose. This indicates the preset temperature hyperparameter.

[0032] Subsequently, the points of interest that can be recommended to users are taken as potential points of interest, and the weight of the j-th potential point of interest recommended to the i-th user is calculated based on the encoding vector. : in, The prototype of user preferences encoded by the i-th user who has commented. This is the prototype for encoding user preferences for the i-th user who has not left a comment. The non-preference prototype for the coded user who has commented on the i-th user. Let i be the non-preference prototype of the i-th user who has not commented. Let be the interest point vector of the j-th potential interest point. This allows us to calculate the alignment loss. : in, For user preference probability, The probability of user aversion is given by exp(), which represents the exponential operation.

[0033] Separation loss also needs to be calculated. : in, This represents the encoded user preference prototype for the u-th user other than the i-th user. This represents the encoded user non-preference prototype of the u-th user other than the i-th user.

[0034] Here, in order to perform a weighted summation of the uniformity loss, alignment loss, and separation loss, a hyperparameter Decay needs to be constructed: in, For the current training iteration of the encoder, This represents the total number of training epochs for the encoder. Subsequently, the uniformity loss, alignment loss, and separation loss are weighted and summed to obtain the clustering classification loss. : in, To align the loss weights, The uniform loss weights are used. Hyperparameters monotonically decrease with increasing training epochs, and each hyperparameter value corresponds to both alignment loss weights and uniform loss weights. As the hyperparameters decrease, the alignment loss weights increase, and the uniform loss weights decrease, but the alignment loss weights will not exceed the uniform loss weights. The text encoder is trained to achieve clustered classification loss until the predetermined number of training epochs is reached. The encoded user preference prototypes obtained through the trained text encoder are used as preference cluster centers. The encoded user dislike prototypes obtained through the trained text encoder are used as dislike cluster centers. This yields the cluster center. During encoder training, the parameters of the text encoder, obtained after training with data from n users, can be fed back to the text encoder to help it update its local parameters.

[0035] S3: Obtain the state space based on the cluster center, construct the action network, input the state space into the action network to obtain the residual offset, construct the value network, and estimate the cumulative advantage parameter through the value network; Furthermore, the objective of this stage is to obtain the residual offset and the cumulative advantage parameter. Specifically, in step S3, the preference cluster center is obtained from the cluster centers, the state space is input into the action network, the action network output is obtained, and the residual offset is obtained based on the preference cluster center and the action network output.

[0036] In step S3, the state space is input into the value network to obtain the temporal difference error, and the cumulative advantage parameter is obtained through the temporal difference error.

[0037] The specific implementation method for the above steps in this embodiment is as follows: Here, we first need to extract t priority interest points from the recommendable interest points according to their weights, thus obtaining the state space for the t-th step selection: in, This represents vector concatenation. Let represent the mean of the interest point vectors of the t preferred interest points. Then, an action network A() is constructed, with the state space input into the action network to obtain the action network output. The action network can select interest points to recommend to the user from the prioritized interest points. The residual offset at step t is obtained based on the preference cluster center and the action network output. : in, The residual offset training parameters are set as follows: the higher the training epoch, the lower the value of the residual offset training parameters.

[0038] Subsequently, a value network V() is constructed, which can evaluate the quality of the action network's output. The state space is input into the value network to obtain the temporal difference error at step t. : in, The single-step reward for step t is the sum of the interest point weights of the t priority recommended interest points extracted in step t. As the first hyperparameter, The value network output selected at step t+1 is the advantage estimate at step t+1. The value network output selected at step t is the advantage estimate at step t.

[0039] Finally, based on the timing difference error selected in step t+1... This yields the cumulative advantage parameters for the t-th step selection. : in, For cumulative parameters, This is the second hyperparameter.

[0040] S4: Calculate the action probability distribution using the residual offset, use the action probability distribution and cumulative advantage parameters to calculate the action network optimization parameters, input the state space into the value network and perform truncated cumulative value calculation to obtain the truncated cumulative value, and obtain the value network optimization parameters based on the truncated cumulative value; Furthermore, the objective of this stage is to obtain the action network optimization parameters and the value network optimization parameters. Specifically, in step S4, the action probability distribution, including the updated action probability distribution and the original action probability distribution, is calculated using the residual offset. The ratio of the new and old probability distributions is obtained using the action probability distribution, and the action network optimization parameters are obtained using the ratio of the new and old probability distributions and the cumulative advantage parameter.

[0041] In step S4, the state space is input into the value network to obtain an advantage estimate including the advantage estimate before the update and the advantage estimate after the update. The truncated cumulative value is obtained through the advantage estimate. The target return is obtained using the advantage estimate before the update and the cumulative advantage parameter. The value network optimization parameters are obtained through the target return and the truncated cumulative value.

[0042] The specific implementation method for the above steps in this embodiment is as follows: Here, the updated action probability distribution can first be calculated based on the residual offset of the current action network. : in, Let t be the action selected at step t. This action is to recommend the v-th point of interest to the user. Here, the v-th point of interest is the one that maximizes the probability distribution of the action after selection. Let v be the interest point vector of the v-th interest point. Substituting the residual offset of the action network before the update into the formula above, we can obtain the action probability distribution before the update. This allows us to calculate the ratio of the old to the new probability at step t. : This allows us to obtain the action network optimization parameters. : Here, min() means to retrieve the minimum value within the parentheses, and clip() means to use the clip function. This indicates that hyperparameters are truncated.

[0043] The state space is then input into the value network before the update to obtain the advantage estimate before the update. The state space is input into the current value network to obtain the updated advantage estimate. This yields the advantage estimate. From this, the truncated cumulative value can be calculated. : Based on the pre-update advantage estimate and cumulative advantage parameters, the target return at step t can also be calculated. : This allows us to obtain the value network optimization parameters. : In this context, max() represents finding the maximum value within the parentheses.

[0044] S5: Update the action network and value network by optimizing the parameters of the action network and value network to obtain the target action network. Then, output the recommended interest points through the target action network to complete the interest point recommendation.

[0045] Furthermore, the objective of this stage is to output recommended interest points through the target action network, thereby completing the interest point recommendation. Specifically, in step S5, target optimization parameters are constructed using the action network optimization parameters and the value network optimization parameters. The action network and value network are then updated using the target optimization parameters to obtain the target action network.

[0046] The specific implementation method for the above steps in this embodiment is as follows: Here, the target optimization parameters can be constructed using the action network optimization parameters and the value network optimization parameters. : Here, H represents the entropy of the interest points, describing their diversity. The more types and quantities of interest points in the text data, the higher this value. α is the entropy weight that decreases with each training epoch. The goal is to minimize the target optimization parameters by updating the parameters of the action network and value network. After multiple updates, the target action network is obtained. Here, a user's interest point text data is used for only four updates; after that, the next user's interest point text data is used to update the parameters of the action network and value network. Finally, the target action network can be used to recommend interest points to the user's first location. By using the target action network to obtain the action probability distribution, the recommended interest points can be obtained, completing the interest point recommendation process.

[0047] The effectiveness of the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning was also verified here. The method and other existing methods were validated on different publicly available cross-domain interest point text datasets, namely Dataset 1 and Dataset 2. The results are shown in Tables 1 and 2. Table 1. A schematic diagram of the interest point recommendation test results of this method and other methods on Dataset 1.

[0048] Table 2. A schematic diagram of the interest point recommendation test results of this method and other methods on Dataset 2.

[0049] Among them, KDDC, CLSPRec, TMHKG, MGSTDN, UPTDNet, TCKT, CCPR, and PC-GDN are all existing interest point recommendation methods. `recall20` represents the recall of the first 20 recommendations, `recall15` represents the recall of the first 15 recommendations, `recall10` represents the recall of the first 10 recommendations, `recall5` represents the recall of the first 5 recommendations, `ndcg20` represents the normalized depreciation cumulative gain of the first 20 recommendations, and `ndcg`... 15 represents the normalized cumulative gain of the first 15 recommendations, ndcg10 represents the normalized cumulative gain of the first 10 recommendations, ndcg5 represents the normalized cumulative gain of the first 5 recommendations, mrr20 represents the average reciprocal rank of the first 20 recommendations, mrr15 represents the average reciprocal rank of the first 15 recommendations, mrr10 represents the average reciprocal rank of the first 10 recommendations, and mrr5 represents the average reciprocal rank of the first 5 recommendations. The bolded numbers indicate that they achieved the best performance in that metric. It can be seen that this method achieves good interest point recommendation results on both Dataset 1 and Dataset 2.

[0050] The cross-domain interest point recommendation system based on dual prototype alignment and reinforcement learning provided by this invention will be described below. The cross-domain interest point recommendation system based on dual prototype alignment and reinforcement learning described below can be referred to in correspondence with the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning described above.

[0051] Figure 2 A schematic diagram of the cross-domain interest point recommendation system based on dual prototype alignment and reinforcement learning is shown, as follows: Figure 2 As shown, the method for performing cross-domain interest point recommendation based on dual-prototype alignment and reinforcement learning as described above includes: Text preprocessing module 100: used to acquire interest point text data, select a large language model, input the interest point text data into the large language model, and obtain preprocessed text data; Among them, the large language model scores the sentiment preference of interest points, and obtains preprocessed text data including preprocessed user preference prototypes and preprocessed user dislike prototypes; the preprocessed user preference prototypes and preprocessed user dislike prototypes constitute a dual prototype. Clustering classification module 200: used to encode preprocessed text data to obtain encoded vectors, calculate the clustering classification loss of the encoded vectors, and obtain cluster centers based on the clustering classification loss; Cumulative Advantage Parameter Module 300: It is used to obtain the state space based on the cluster center, construct the action network, input the state space into the action network to obtain the residual offset, construct the value network, and estimate the cumulative advantage parameter through the value network. Network optimization parameter module 400: It is used to calculate the action probability distribution through residual offset, calculate the action network optimization parameters using the action probability distribution and cumulative advantage parameters, input the state space into the value network and perform truncated cumulative value calculation to obtain the truncated cumulative value, and obtain the value network optimization parameters based on the truncated cumulative value. Interest point recommendation module 500: It is used to update the action network and value network by optimizing the parameters of the action network and the value network to obtain the target action network, and output recommended interest points through the target action network to complete the interest point recommendation.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-domain interest point recommendation method based on dual-prototype alignment and reinforcement learning, characterized in that, include: S1: Obtain the text data of points of interest, select a large language model, input the text data of points of interest into the large language model, and obtain preprocessed text data; Among them, the large language model scores the sentiment preference of interest points, and obtains preprocessed text data including preprocessed user preference prototypes and preprocessed user dislike prototypes; the preprocessed user preference prototypes and preprocessed user dislike prototypes constitute a dual prototype. S2: Encode the preprocessed text data to obtain the encoded vector, calculate the clustering classification loss of the encoded vector, and obtain the cluster center based on the clustering classification loss; S3: Obtain the state space based on the cluster center, construct the action network, input the state space into the action network to obtain the residual offset, construct the value network, and estimate the cumulative advantage parameter through the value network; S4: Calculate the action probability distribution using the residual offset, use the action probability distribution and cumulative advantage parameters to calculate the action network optimization parameters, input the state space into the value network and perform truncated cumulative value calculation to obtain the truncated cumulative value, and obtain the value network optimization parameters based on the truncated cumulative value; S5: Update the action network and value network by optimizing the parameters of the action network and value network to obtain the target action network. Then, output the recommended interest points through the target action network to complete the interest point recommendation.

2. The cross-domain interest point recommendation method based on dual-prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S2, a text encoder is constructed to encode the preprocessed text data to obtain the encoded vector.

3. The cross-domain interest point recommendation method based on dual-prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S2, the uniform loss and separation loss of the encoding vector are calculated, the interest point weights are calculated based on the encoding vector, the alignment loss is obtained through the interest point weights, and the uniform loss, alignment loss, and separation loss are weighted and summed to obtain the clustering classification loss.

4. The cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S3, the preference cluster center is obtained from the cluster center, the state space is input into the action network, the action network output is obtained, and the residual offset is obtained based on the preference cluster center and the action network output.

5. The cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S3, the state space is input into the value network to obtain the temporal difference error, and the cumulative advantage parameter is obtained through the temporal difference error.

6. The cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S4, the action probability distribution, including the updated action probability distribution and the original action probability distribution, is calculated using the residual offset. The ratio of the old and new probabilities is obtained using the action probability distribution, and the action network optimization parameters are obtained using the ratio of the old and new probabilities and the cumulative advantage parameter.

7. The cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S4, the state space is input into the value network to obtain an advantage estimate including the advantage estimate before the update and the advantage estimate after the update. The truncated cumulative value is obtained through the advantage estimate. The target return is obtained using the advantage estimate before the update and the cumulative advantage parameter. The value network optimization parameters are obtained through the target return and the truncated cumulative value.

8. The cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning according to claim 1, characterized in that, In step S5, target optimization parameters are constructed using the action network optimization parameters and the value network optimization parameters. The action network and the value network are then updated using the target optimization parameters to obtain the target action network.

9. A cross-domain interest point recommendation system based on dual prototype alignment and reinforcement learning, used to execute the cross-domain interest point recommendation method based on dual prototype alignment and reinforcement learning as described in any one of claims 1 to 8, characterized in that, include: Text preprocessing module: used to acquire text data of points of interest, select a large language model, input the text data of points of interest into the large language model, and obtain preprocessed text data; Among them, the large language model scores the sentiment preference of interest points, and obtains preprocessed text data including preprocessed user preference prototypes and preprocessed user dislike prototypes; the preprocessed user preference prototypes and preprocessed user dislike prototypes constitute a dual prototype. Clustering classification module: used to encode preprocessed text data to obtain encoded vectors, calculate the clustering classification loss of the encoded vectors, and obtain cluster centers based on the clustering classification loss; Cumulative Advantage Parameter Module: This module is used to obtain the state space based on the cluster center, construct the action network, input the state space into the action network to obtain the residual offset, construct the value network, and estimate the cumulative advantage parameter through the value network. The network optimization parameter module is used to calculate the action probability distribution through residual offset, calculate the action network optimization parameters using the action probability distribution and cumulative advantage parameters, input the state space into the value network and perform truncated cumulative value calculation to obtain the truncated cumulative value, and obtain the value network optimization parameters based on the truncated cumulative value. Interest point recommendation module: It is used to update the action network and value network by optimizing the parameters of the action network and the value network to obtain the target action network, and then output recommended interest points through the target action network to complete the interest point recommendation.

Images