Next interest point recommendation method based on continuous learning

Abstract

The invention discloses a next interest point recommendation method based on continuous learning, and relates to the technical field of interest point recommendation, and the method comprises the following steps: S1, according to a memory entry of a user, constructing an interest memory for the user; s2, generating key representation for interest memory; s3, generating a long-term interest vector according to the key representation; and S4, generating a final interest vector according to the long-term interest vector and the short-term user behavior of the user, and completing interest point recommendation. Interest memory, context-aware key coding, generative interest retrieval and adaptive interest updating and fusion are integrated, and an existing next interest point recommendation model can be seamlessly integrated.

Description

Technical Field

[0001] The present invention relates to the technical field of point-of-interest recommendation, and particularly to a method for recommending the next point of interest based on continuous learning. Background Art

[0002] With the popularization of location-based services, next point-of-interest recommendation has become a crucial component in enhancing the user experience of various applications, including navigation systems, travel platforms, and food delivery services. Effective recommendation requires modeling user interests based on their historical interaction behaviors. Technologies such as recurrent neural networks and graph neural networks have been widely used to capture sequential patterns and spatio-temporal relationships. However, these methods typically focus on static models, assuming that user interests remain fixed and ignoring the regular updates brought about by new interaction behaviors.

[0003] However, this assumption contradicts the dynamic nature of user interests, as user interests may change due to factors such as time, location, and point-of-interest categories. For example, seasonal changes may cause users to prefer indoor venues in winter and outdoor venues in summer. This context-driven interest evolution highlights the need for a model that can adapt to changes in user interests through continuous updates.

[0004] Traditional next point-of-interest recommendation uses a static training model during the deployment phase, while next point-of-interest recommendation based on continuous learning involves periodically updating the model. As user preferences evolve, the performance of the static model degrades over time, highlighting the necessity of continuous model updates. An intuitive solution is to retrain the model periodically using all the observed data, but this method is computationally expensive. Another straightforward approach is to fine-tune the model using the most recently observed data. However, this may lead to catastrophic forgetting, i.e., the loss of previously learned knowledge, resulting in performance lower than that of a fully retrained model. To address this issue, next point-of-interest recommendation based on continuous learning aims to dynamically adjust the model through continuous updates to adapt to the changing user interests without full retraining. This method can provide timely and context-relevant recommendations while maintaining computational efficiency.

[0005] However, designing an efficient and effective continual learning-based next interest point recommendation framework faces numerous challenges. The first challenge is how to retain historical preferences, i.e., preserve past knowledge while adapting to users' dynamic interactions to avoid catastrophic forgetting. This problem is particularly complex because human mobility preferences are context-sensitive and influenced by spatio-temporal factors in next interest point recommendations. The second challenge is how to retrieve relevant historical preferences. Effective retrieval is crucial as not all historical preferences remain relevant or valuable as users' interests change. Correct retrieval ensures accurate inference of continual interests. The difficulty of this challenge lies in that the system must prioritize relevant information based on multiple factors (such as location, time, and activity) while avoiding redundancy or omission of key details. The third challenge is how to adaptively balance historical information and current information. Successful recommendations require combining continual interests and recent interests, where continual interests provide a stable foundation for understanding long-term behavior, and recent interests reflect immediate needs. Balancing these interests is difficult because the importance of historical information and current information usually varies depending on the user and context. The fourth challenge is to ensure model-agnostic adaptability. The continual framework needs to be able to integrate with various next interest point recommendation models (such as recurrent neural networks, Transformers, and diffusion models) without significant modification. Since different models have different characteristics, designing a model-agnostic framework that can maintain adaptability without sacrificing performance is highly challenging. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a method for next interest point recommendation based on continual learning.

[0007] The technical solution of the present invention is: A method for next interest point recommendation based on continual learning includes the following steps:

[0008] S1. Construct an interest memory for the user according to the user's memory entries;

[0009] S2. Generate a key representation for the interest memory;

[0010] S3. Generate a long-term interest vector according to the key representation;

[0011] S4. Generate a final interest vector based on the user's long-term interest vector and short-term user behavior to complete the interest point recommendation.

[0012] Further, in S1, the user's interest memory has the following expression:

[0013] ;

[0014] In the formula, represents the user In the key vector in the th memory entry, represents the recommended probability distribution, represents the timestamp of the interaction,

[0015] Further, S2 includes the following sub-steps:

[0016] S21. Normalize the user's GPS coordinates and generate coordinate embeddings based on the normalized GPS coordinates;

[0017] S22. Generate region embeddings based on the GPS coordinates;

[0018] S23. Generate geographical embeddings based on the coordinate embeddings and region embeddings;

[0019] S24. Generate discrete time embeddings and periodic position encodings based on the user's access time information;

[0020] S25. Generate time embeddings based on the discrete time embeddings and periodic position encodings;

[0021] S26. Generate category embeddings for the user;

[0022] S27. Generate key representations for the interest memory based on the user's geographical embeddings, time embeddings, and category embeddings.

[0023] Further, in S21, the coordinate embedding has the following expression:

[0024] ;

[0025] In the formula, represents the weight matrix, represents the bias vector, represents the normalized trajectory coordinates;

[0026] In S22, the region embedding has the following expression:

[0027] ;

[0028] In the formula, represents the embedding layer, represents the region ID corresponding to the trajectory point;

[0029] In S23, the geographical embedding has the following expression:

[0030] ;

[0031] In S24, discrete-time embedding has the following expression:

[0032] ;

[0033] In the formula, represents the first embedding matrix, represents hours, represents the second embedding matrix, represents the week;

[0034] In S24, periodic positional encoding has the following expression:

[0035] ;

[0036] In the formula, represents a single time frequency, represents the normalized intra-day time, represents a set of frequencies;

[0037] In S25, time embedding has the following expression:

[0038] ;

[0039] In S26, class embedding has the following expression:

[0040] ;

[0041] In the formula, represents the first embedding layer, represents the second embedding layer, represents the original class, represents the derived class;

[0042] In S27, key representation has the following expression:

[0043] ;

[0044] In the formula, represents the long short-term memory network, represents the linear layer.

[0045] Furthermore, S3 includes the following sub-steps:

[0046] S31. Use the encoder to map the original key to the latent space, obtain the Gaussian distribution, and generate the latent representation according to the Gaussian distribution;

[0047] S32. Use the decoder to generate several keys from the latent representation;

[0048] S33. Calculate the mean squared error loss between the key representations and the original keys based on the generated keys;

[0049] S34. Calculate the KL divergence loss between the key representations and the original keys;

[0050] S35. Calculate the diversity loss between the key representations and the original keys;

[0051] S36. Generate the final loss function based on the mean squared error loss, KL divergence loss, and diversity loss between the key representations and the original keys;

[0052] S37. Train the key generator using the final loss function, and generate a number of query keys using the trained key generator;

[0053] S38. Calculate the RRF score based on the number of query keys;

[0054] S39. Generate the long-term interest vector based on the RRF score.

[0055] Furthermore, in S31, the mean corresponding to the Gaussian distribution has the following expression:

[0056] ;

[0057] In the formula, represents the original key, represents a multi-layer perceptron;

[0058] In S31, the latent representation has the following expression:

[0059] ;

[0060] In the formula, represents a normal distribution, represents the standard deviation of the Gaussian distribution;

[0061] In S32, the -th key has the following expression:

[0062] ;

[0063] In the formula, represents the Sigmoid activation function, represents the latent representation of the -th key;

[0064] In S33, the mean squared error loss between the key representations and the original keys has the following expression:

[0065] ;

[0066] In the formula, represents the number of generated keys;

[0067] In S34, the key represents the KL divergence loss between the generated key and the original key The expression is:

[0068] ;

[0069] In the formula, represents the mathematical expectation;

[0070] In S35, the key represents the diversity loss between the generated key and the original key The expression is:

[0071] ;

[0072] In the formula, represents the th key;

[0073] In S36, the expression of the final loss function is:

[0074] ;

[0075] In the formula, represents the first hyperparameter, represents the second hyperparameter;

[0076] In S38, the expression of the RRF score is:

[0077] ;

[0078] In the formula, represents the set of keys, represents the smoothing factor, represents the memory entry 's rank under the current key;

[0079] In S39, the expression of the long-term interest vector is:

[0080] ;

[0081] In the formula, represents the activation function, represents the memory entry 's corresponding value.

[0082] Furthermore, S4 includes the following sub-steps:

[0083] S41. Calculate the update weight according to the user's consistency score;

[0084] S42. Update the memory entry for the user according to the update weight;

[0085] S43. After completing the update of the memory entry, generate the final interest representation according to the user's consistency score;

[0086] S44. Generate the final interest vector based on the long-term interest vector and the final interest representation to complete the point-of-interest recommendation.

[0087] Furthermore, in S41, the update weight has the following expression:

[0088] ;

[0089] In the formula, represents a predefined base value, represents the user's consistency score, represents the average consistency score of all users, represents a parameter for controlling the adjustment sensitivity;

[0090] In S42, the expression for updating the memory entry is:

[0091] ;

[0092] ;

[0093] In the formula, represents the updated key, represents the updated value, represents the matched key, represents the matched value, represents the newly input key, represents the newly input value;

[0094] In S43, the final interest representation has the following expression:

[0095] ;

[0096] In the formula, represents a predefined base value;

[0097] In S44, the final interest vector has the following expression:

[0098]

[0099] In the formula, represents the recent user behavior, represents the long-term interest vector.

[0100] The beneficial effects of the present invention are as follows: The present invention integrates interest memory, context-aware key encoding, generative interest retrieval, and adaptive interest update and fusion, and can seamlessly integrate existing next interest point recommendation models, thereby improving their performance in next interest point recommendation based on continuous learning, enabling them to adapt to the evolution of user preferences without the need for a complete retraining, and achieving a balance between long-term and recent interests, highlighting its ability to effectively cope with the dynamic evolution of user interests in the next interest point recommendation task based on continuous learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 is a flowchart of a method for next interest point recommendation based on continuous learning;

[0102] Figure 2 is a comparison chart of update times;

[0103] Figure 3 is an experimental result chart of updated weights;

[0104] Figure 4 is an experimental result chart of fusion weights;

[0105] Figure 5 is an experimental result chart of the number of generated keys. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0106] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0107] As Figure 1 shown, the present invention provides a method for next interest point recommendation based on continuous learning, including the following steps:

[0108] S1. Construct an interest memory for the user according to the user's memory entries;

[0109] S2. Generate a key representation for the interest memory;

[0110] S3. Generate a long-term interest vector according to the key representation;

[0111] S4. Generate a final interest vector according to the user's long-term interest vector and short-term user behavior to complete the interest point recommendation.

[0112] The next interest point recommendation based on continuous learning in the present invention can integrate context-related historical information and current information. The present invention proposes a generative interest retrieval and adaptive modeling framework, which includes four main components, specifically: an interest memory module, which preserves the user's historical preferences by storing context information as keys and the output of the fine-tuned next interest point recommendation model as values, rather than the original data samples; a generative interest retrieval module, which generates multiple potential keys to capture the interest diversity of users in different contexts. Multi-key retrieval ensures comprehensive access to continuous interests because a single key cannot fully represent diverse context requirements; an adaptive interest update and fusion module, which dynamically balances historical information and current information using a consistency score. This mechanism enhances the update of interest memory and the recommendation performance by adapting to changes in user behavior; a context-aware key encoding module, which encodes context information into a unified key vector. This design ensures seamless integration with different next interest point recommendation models.

[0113] The continuous learning process of the generative interest retrieval and adaptive modeling framework includes two stages: an update stage and a deployment stage. In the update stage, the context-aware key encoding module generates interest keys by encoding spatio-temporal features and categorical features into a unified key vector. These keys are paired with the output of the fine-tuned next interest point recommendation model to form structured user preferences and are stored in the interest memory module. During the update process, a consistency score is introduced to adaptively balance historical preferences and new preferences. In the deployment stage, the generative interest retrieval module uses a model based on conditional variational autoencoder to generate multiple similar but independent keys to retrieve continuous interests from the interest memory module, thus ensuring that the recommendation is context-related. Subsequently, the continuous interests and recent interests are fused, and the contributions of both are adaptively balanced through the consistency score.

[0114] In the embodiment of the present invention, in S1, the user's interest memory has the following expression:

[0115] ;

[0116] In the formula, represents the key vector of the user in the th memory entry, represents the recommendation probability distribution, represents the timestamp of the interaction, represents the maximum interest storage capacity of the user.

[0117] To optimize storage space, the interest memory only stores the probability values of the top 50 interest point candidates in each value vector, and sets the remaining POI probabilities to 0. These representations are stored using a sparse matrix, which greatly reduces memory consumption while retaining the most relevant information. In addition, each user is assigned a maximum interest storage capacity, denoted as .

[0118] In the embodiment of the present invention, S2 includes the following sub-steps:

[0119] S21. Normalize the user's GPS coordinates and generate a coordinate embedding based on the normalized GPS coordinates;

[0120] S22. Generate a region embedding based on the GPS coordinates;

[0121] S23. Generate a geographical embedding based on the coordinate embedding and the region embedding;

[0122] S24. Generate a discrete time embedding and a periodic position encoding based on the user's access time information;

[0123] S25. Generate a time embedding based on the discrete time embedding and the periodic position encoding;

[0124] S26. Generate a category embedding for the user;

[0125] S27. Generate a key representation for the interest memory based on the user's geographical embedding, time embedding, and category embedding.

[0126] In S2, a model-agnostic context-aware key encoding module is proposed. This module is used to generate keys for interest memories and decouple them from the recommendation output. The module fuses geographical, temporal, and category embeddings and combines them through a key encoder to form a unified context representation, thereby enhancing the retrieval ability of interest memories. Geographical embedding effectively fuses fine-grained geographical coordinate information with macroscopic regional features to enhance the spatial feature representation ability. Temporal embedding is used to encode temporal information, including hours, days of the week, and normalized intra-day time, combining discrete time representation with periodic positional encoding, thereby enhancing the representation ability of temporal features and enabling the model to learn the user's temporal preferences. Category embedding is crucial for the next point-of-interest recommendation because the category of a point of interest can reflect the user's activity type. Simply using the original category ID cannot fully capture the semantic relationship. Therefore, in the present invention, the original category and derived category information are combined. The original category represents the initial category of the point of interest, while the derived category is automatically generated by ChatGPT using context information; the derived category can provide a higher level of abstraction and capture the semantic meaning and underlying concepts of the point of interest. The final representation effectively captures the semantic information of the category by combining the original category embedding and the derived category embedding, thereby improving the accuracy of point-of-interest recommendation. The dimensionality of the fused representation is adjusted through a linear transformation, and then an LSTM is used to learn sequence dependencies and extract temporal dynamic features. The final output is used as the key representation for the update of interest memory and generative interest retrieval, thereby achieving efficient user interest modeling.

[0127] In the embodiment of the present invention, in S21, the coordinate embedding has the following expression:

[0128] ;

[0129] In the formula, represents the weight matrix, represents the bias vector, represents the normalized trajectory coordinates;

[0130] In S22, the regional embedding has the following expression:

[0131] ;

[0132] In the formula, represents the embedding layer, represents the regional ID corresponding to the trajectory point; the embedding layer first converts the discrete numerical input into a one-hot vector and then applies a linear transformation to generate a low-dimensional embedding representation.

[0133] In S23, the geographical embedding has the following expression:

[0134] ;

[0135] In S24, the discrete-time embedding has the expression:

[0136] ;

[0137] In the formula, represents the first embedding matrix, represents hours, represents the second embedding matrix, represents the week; the embedding matrix maps the input to a latent vector.

[0138] In S24, the periodic positional encoding has the expression:

[0139] ;

[0140] In the formula, represents a single time frequency, represents the normalized intraday time, represents a set of frequencies; is used to capture time patterns at different scales.

[0141] In S25, the time embedding has the expression:

[0142] ;

[0143] In S26, the categorical embedding has the expression:

[0144] ;

[0145] In the formula, represents the first embedding layer, represents the second embedding layer, represents the original category, represents the derived category;

[0146] In S27, the key encoder fuses the above embeddings (geographical, temporal, and categorical) into a unified representation and processes it through a sequence model to capture temporal dependencies. The key representation has the expression:

[0147] ;

[0148] In the formula, represents a long short-term memory network, represents a linear layer.

[0149] In an embodiment of the present invention, S3 includes the following sub-steps:

[0150] S31. Use an encoder to map the original key to the latent space to obtain a Gaussian distribution, and generate a latent representation according to the Gaussian distribution;

[0151] S32. Use a decoder to generate several keys from the latent representation;

[0152] S33. Calculate the mean squared error loss between the key representation and the original key according to the generated several keys;

[0153] S34. Calculate the KL divergence loss between the key representation and the original key;

[0154] S35. Calculate the diversity loss between the key representation and the original key;

[0155] S36. Generate a final loss function according to the mean squared error loss, KL divergence loss, and diversity loss between the key representation and the original key;

[0156] S37. Use the final loss function to train the key generator, and use the trained key generator to generate several query keys;

[0157] S38. Calculate the RRF score according to the several query keys;

[0158] S39. Generate a long-term interest vector according to the RRF score.

[0159] The simplest approach of the retrieval method based on interest memory is to find the most similar key and directly use its corresponding value. However, a single key cannot comprehensively capture the diverse interests of users in different contexts. For this reason, the present invention proposes a generative interest retrieval module, which performs interest retrieval by generating multiple relevant but independent keys. Specifically, the present invention adopts a method based on a variational autoencoder to generate keys, and uses these keys to retrieve diverse long-term interests in the interest memory. This strategy ensures a broader interest representation, enabling the recommendation results to balance between relevance and diversity.

[0160] The key generator is a conditional variational autoencoder model, which consists of an encoder, a reparameterization module, and a decoder. This architecture can generate diverse and context-compliant query keys through sampling in the latent space. Different from traditional autoencoders that cannot model uncertainty, and generative adversarial networks that are unstable in training and prone to mode collapse, the conditional variational autoencoder achieves a balance between diversity and semantic relevance.

[0161] The encoder uses a multi-layer perceptron to map the real key to the latent space, and outputs the mean and log variance, which are used to parameterize the latent Gaussian distribution.

[0162] Using the trained key generator, the decoder can sample a latent vector from a standard Gaussian distribution based on the original key, and then generate diverse and context-compliant keys. These generated keys are then used to query the user's interest memory to retrieve long-term interests.

[0163] The retrieval process adopts the Reciprocal Rank Fusion (RRF) mechanism, which aggregates the retrieval results of multiple keys by assigning a ranking score related to the query key to each memory entry. The resulting long-term interest vector can capture the individual's historical preferences and serve as one of the key factors for the final recommendation.

[0164] In the embodiment of the present invention, in S31, the mean value corresponding to the Gaussian distribution has the following expression:

[0165] ;

[0166] In the formula, represents the original key, represents a multi-layer perceptron;

[0167] In S31, the latent representation has the following expression:

[0168] ;

[0169] In the formula, represents a normal distribution, represents the standard deviation of the Gaussian distribution;

[0170] In S32, the th key has the following expression:

[0171] ;

[0172] In the formula, represents the Sigmoid activation function, represents the th latent representation of the key;

[0173] In S33, the mean squared error loss between the key representation and the original key has the following expression:

[0174] ;

[0175] In the formula, represents the number of generated keys;

[0176] In S34, the KL divergence loss between the key representation and the original key has the following expression:

[0177] ;

[0178] In the formula, represents the mathematical expectation;

[0179] In S35, the key represents the diversity loss with respect to the original key The expression is:

[0180] ;

[0181] In the formula, represents the th key;

[0182] In S36, the expression of the final loss function is:

[0183] ;

[0184] In the formula, represents the first hyperparameter, represents the second hyperparameter; the hyperparameters are used to balance the contributions of the KL divergence and the diversity loss.

[0185] In S38, the expression of the RRF score is:

[0186] ;

[0187] In the formula, represents the set of keys, represents the smoothing factor, represents the memory entry 's rank under the current key; the smoothing factor is used to prevent over - weighting of the entries with higher ranks.

[0188] In S39, the expression of the long - term interest vector is:

[0189] ;

[0190] In the formula, represents the activation function, represents the value corresponding to the memory entry .

[0191] In the embodiments of the present invention, S4 includes the following sub - steps:

[0192] S41. Calculate the updated weight according to the user's consistency score;

[0193] S42. Update the memory entry for the user according to the updated weight;

[0194] S43. After completing the update of the memory entry, generate a final interest representation based on the user's consistency score;

[0195] S44. Generate a final interest vector based on the long-term interest vector and the final interest representation to complete the point-of-interest recommendation.

[0196] In the update stage, refine the interest representation according to the change of user preferences; in the deployment stage, fuse the long-term interest and the recent interest to generate personalized recommendations. Although their purposes are different, the update process and the fusion process are closely related, and both require a refined combination of the long-term interest and the recent interest. Therefore, the present invention introduces a consistency score to adaptively guide the update and fusion processes to ensure the balance between historical information and new information. Intuitively, a higher consistency score means that the user's interest is consistent with the overall trend, indicating that the recent trend has a greater influence. For example, when users visit popular attractions (such as parks and beaches) in summer, these interests are consistent with the behavior of most users, so the recommendation should focus on the recent trend (such as swimming or ice cream). On the contrary, a lower consistency score indicates that the user's interest is more unique, and more attention should be paid to their personalized preferences. For example, a running enthusiast still frequently visits the park even in winter, and the recommendation system needs to pay more attention to their personalized behavior rather than the overall trend.

[0197] This mechanism dynamically balances historical preferences and current trends in interest memory update and recommendation generation. The user 's consistency score is calculated by the cosine similarity between the output of the personalized model and the output of the collective model:

[0198] ;

[0199] where and are the outputs of the collective model and the personalized model on the user training data, respectively.

[0200] In the embodiment of the present invention, in S41, the expression of the update weight is:

[0201] ;

[0202] In the formula, represents a predefined base value, represents the user 's consistency score, represents the average value of the consistency scores of all users, represents a parameter controlling the adjustment sensitivity;

[0203] In S42, the cosine similarity is used to evaluate the matching degree between the newly input key and the existing memory entries. If the maximum similarity exceeds the threshold , the matching memory entry is updated. The expression for updating the memory entry is:

[0204] ;

[0205] ;

[0206] In the formula, represents the updated key, represents the updated value, represents the matched key, represents the matched value, represents the newly input key, represents the newly input value; if the maximum similarity does not exceed the threshold , a new entry is added if the storage space permits, otherwise the oldest entry is replaced. This strategy ensures that the interest memory can be dynamically updated and efficiently managed, reflecting both changes in user preferences and maintaining reasonable utilization of the storage space.

[0207] In S43, the expression for the final interest representation is:

[0208] ;

[0209] In the formula, represents a predefined base value;

[0210] In S44, the expression for the final interest vector is:

[0211]

[0212] In the formula, represents the recent user behavior, represents the long-term interest vector. is generated by a fine-tuned personalized model and reflects the recent user behavior.

[0213] The following is illustrated with specific embodiments.

[0214] The present invention conducts experiments on three publicly available real-world datasets, including Foursquare-NYC, Foursquare-TKY, and Gowalla-CA. The processing method of the datasets is as follows: First, the present invention filters out POIs and users with less than 10 check-ins; then, the check-in records are split in chronological order, where the first 50% of the records form the basic data block , the remaining check-in records are evenly divided into five data blocks; then, the user check-in records in each data block are divided into trajectories at one-week intervals, and trajectories containing only one check-in are discarded, finally forming trajectory data blocks . The model is first trained on the basic data block and validated on the first incremental data block . Subsequently, the model is updated on the complete and the next incremental data block is randomly divided into two parts for validation and testing respectively. This process is repeated for subsequent data blocks. The statistical information of the processed dataset is shown in Table 1.

[0215] Table 1

[0216]

[0217] The present invention applies the proposed framework to three representative next POI recommendation models: Flashback, GETNext, and DiffPOI. Flashback is an RNN-based next POI recommendation model. This method utilizes spatio-temporal context information to efficiently retrieve historical hidden states with high predictive power, thereby improving recommendation accuracy. GETNext captures the general movement patterns of users through trajectory flow mapping and introduces a graph-enhanced Transformer model to fully utilize rich collaborative signals. DiffPOI is a recommendation method based on the diffusion model. This model encodes the user access sequence and spatial features through two specially designed graph encoding modules and adopts a diffusion sampling strategy to effectively capture and explore the user's spatial access patterns and behavioral trends. The present invention selects these models as baselines based on their different architectural characteristics: Flashback adopts an RNN structure, GETNext combines GCN and Transformer, while DiffPOI integrates GCN, CNN, and the diffusion model. This diverse selection enables the present invention to comprehensively evaluate GIRAM on different structures.

[0218] The present invention uses two widely used evaluation metrics: Top-k accuracy (Acc@k) and mean reciprocal rank (MRR). Among them, Acc@k measures whether the true visited POI appears in the top k recommended results generated by the model. In this study, the present invention reports Acc@5, Acc@10, and Acc@20 to evaluate the recommendation effect; MRR evaluates the ranking position of the correct POI in the recommendation list. The higher its value, the higher the ranking of the correct POI in the recommendation results, thus reflecting better recommendation quality.

[0219] The present invention compares the proposed method with the following methods: Retrain: Retrain a completely new model using all historical data, and its result is used as a reference benchmark and not included in the ranking comparison; Fine - tune: Only use the latest data block for training, without considering prior information, to perform incremental updates on the model; ADER: Select samples based on item frequencies for retraining to ensure the representativeness of the training data; IncCTR: A method based on knowledge distillation, which uses the output of the previous model as a supervision signal and combines new training data to update the current model; ReLoop2: The current best - performing replay - mechanism continuous - learning recommendation method, which uses an error - memory mechanism to estimate the deviation between the model prediction value and the true label; CMuST: The latest spatio - temporal continuous - learning method. Due to the discrete characteristics of POI data, the present invention removes its task - prompting module to adapt to the scenario of this study.

[0220] Figures 3 - 5 Shows the average update time of three basic models on each data block. In the efficiency evaluation experiment, the present invention uniformly sets the number of training rounds to 10. From the perspective of the running efficiency of each method, Fine - tune has the fastest update speed because it completely ignores historical information, so the computational overhead is the lowest. While GIRAM slightly increases the computational overhead due to the need to update the interest memory, its running time is still much lower than that of Retrain, and at the same time, it maintains a performance comparable to Retrain in terms of recommendation performance. These results indicate that GIRAM can maintain good recommendation effects while ensuring efficient processing of continuous updates.

[0221] To evaluate the effectiveness of each component in the proposed GIRAM model, the present invention designs and tests the following five variants: (1) w / o CKE: Remove the context - aware key - encoding module and replace the generated key with the latent vector in the basic model; (2) w / o GIR: Remove the generative interest - retrieval module and directly use the value of the most similar key in the interest memory as the long - term interest; (3) w / o CS: Remove the consistency - scoring strategy and use a fixed reference value during interest update and fusion instead of an adaptive adjustment strategy; (4) w / o SI: Remove the long - term interest and only use the recent interest for recommendation; (5) w / o RI: Remove the recent interest and only rely on the long - term interest for recommendation.

[0222] The present invention calculates the average results of Acc@5 and MRR on three datasets, and the specific results are shown in Table 2. The experiments show that GIRAM performs optimally on all variants, verifying the effectiveness of its core components. Among them, the performance degradation of w / oCKE indicates the importance of the context-aware key encoding module in continuous next POI recommendation. This module can effectively utilize context information such as location, time, and POI category to improve the recommendation effect. The experimental results of w / oGIR show that the generative interest retrieval module significantly improves the retrieval accuracy by generating multiple potential query keys, thus more accurately identifying relevant long-term interests. The performance degradation of w / oCS further verifies the effectiveness of the consistency scoring strategy, which can adaptively update the interest memory and achieve a dynamic balance between long-term interests and recent interests. In addition, the results of w / oSI and w / oRI show that both long-term interests and recent interests are crucial for accurate recommendation. The lack of any part will cause the system to be difficult to effectively integrate diverse interest patterns, thereby reducing the recommendation quality.

[0223] Table 2

[0224]

[0225] The present invention conducts hyperparameter sensitivity experiments on the NYC dataset to analyze the impact of key hyperparameters on the performance of GIRAM. The experimental results are reported in terms of Acc@5 and analyzed for three key hyperparameters: the update weight , the fusion weight , and the number of generated keys . Among them, and vary between 0.1 and 0.9, while varies between 5 and 50. The experimental results are shown in Figure 3 respectively.

[0226] The analysis results show that when is set to 0.5 (Flashback and DiffPOI) or 0.3 (GETNext), Acc@5 reaches the maximum value. Similarly, when is set to 0.3 (Flashback) or 0.7 (GETNext and DiffPOI), the fusion effect is the best. It should be noted that when and are both set to 0.5, the model performance is close to optimal. This result indicates that extreme or may disrupt the balance between long-term interests and recent interests, thereby affecting the update of the interest memory and the recommendation effect.

[0227] In addition, when the number of generated keys When set to 20 (Flashback) or 30 (GETNext and DiffPOI), the GIRAM performance is optimal. If the number of keys is too small, it may not be able to fully capture the diversity of user interests; while if the number of keys is too large, it may introduce noise and lead to a decline in the model effect. This result verifies that the diversity control of GIRAM in long-term interest retrieval is crucial.

[0228] Those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on these technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A method for recommending the next point of interest based on continuous learning, characterized in that: The following steps are involved: S1. Build interest memory for the user based on the user's memory items; S2, generate key representations for interesting memories; S3, generating a long-term interest vector according to the key representation; S4. Generate the final interest vector based on the user's long-term interest vector and short-term user behavior to complete the recommendation of interest points.

2. The method for recommending the next point of interest based on continuous learning according to claim 1, characterized in that: In S1, the user's interest memory The expression is: ； In the formula, Indicates user In the The key vector in the memory entry, represents the recommendation probability distribution, The timestamp of the interaction. Indicates the user's maximum interest storage capacity.

3. The method for recommending the next point of interest based on continuous learning according to claim 1, characterized in that: The S2 comprises the following sub-steps: S21, normalizing the user's GPS coordinates, and generating coordinate embedding according to the normalized GPS coordinates; S22, generating a region embedding according to the GPS coordinates; S23, generating geographic embedding according to the coordinate embedding and the region embedding; S24, generating discrete time embedding and periodic position coding according to the user's access time information; S25, generating a time embedding according to the discrete time embedding and the periodic position encoding; S26, generating category embedding for the user; S27. Generate key representation for interest memory based on the user's geographic embedding, time embedding and category embedding.

4. The method for recommending the next point of interest based on continuous learning according to claim 3, characterized in that: In S21, the coordinates are embedded The expression is: ； In the formula, represents the weight matrix, represents the bias vector, represents the normalized trajectory coordinates; In S22, the region is embedded The expression is: ； In the formula, represents the embedding layer, Indicates the area ID corresponding to the trajectory point; In S23, geography is embedded The expression is: ； In S24, discrete time embedding The expression is: ； In the formula, represents the first embedding matrix, Indicates hours, represents the second embedding matrix, Indicates the day of the week; In S24, periodic position coding The expression is: ； In the formula, represents a single time frequency, represents the normalized intraday time, represents a set of frequencies; In S25, time embedding The expression is: ； In S26, the category embedding The expression is: ； In the formula, represents the first embedding layer, represents the second embedding layer, represents the original category, Indicates a derived category; In S27, the key indicates The expression is: ； In the formula, represents the long short-term memory network, Represents a linear layer.

5. The method for recommending the next point of interest based on continuous learning according to claim 1, characterized in that: The S3 comprises the following sub-steps: S31, using the encoder to map the original key to the latent space, obtain a Gaussian distribution, and generate a latent representation according to the Gaussian distribution; S32, generating a plurality of keys from the latent representation using a decoder; S33, calculating the mean square error loss between the key representation and the original key according to the generated plurality of keys; S34, calculating the KL divergence loss between the key representation and the original key; S35, calculating the diversity loss between the key representation and the original key; S36, generating a final loss function according to the mean square error loss, KL divergence loss and diversity loss between the key representation and the original key; S37, using the final loss function to train the key generator, and using the trained key generator to generate a number of query keys; S38. Calculate RRF scores based on a number of query keys; S39. Generate a long-term interest vector based on the RRF score.

6. The method for recommending the next point of interest based on continuous learning according to claim 5, characterized in that: In S31, the mean corresponding to the Gaussian distribution The expression is: ； In the formula, represents the original key, represents a multi-layer perceptron; In S31, the potential representation The expression is: ； In the formula, represents a normal distribution, represents the standard deviation of the Gaussian distribution; In the S32, Keys The expression is: ； In the formula, represents the Sigmoid activation function, Indicates potential representation of a key; In S33, the mean square error loss between the key representation and the original key The expression is: ； In the formula, Indicates the number of generated keys; In S34, the KL divergence loss between the key representation and the original key The expression is: ； In the formula, represents mathematical expectation; In S35, the diversity loss between the key representation and the original key The expression is: ； In the formula, Indicates keys; In S36, the final loss function The expression is: ； In the formula, represents the first hyperparameter, represents the second hyperparameter; In S38, RRF score The expression is: ； In the formula, represents a collection of keys, represents the smoothing factor, Represents a memory entry Rank under the current key; In S39, the long-term interest vector The expression is: ； In the formula, represents the activation function, Represents a memory entry The corresponding value.

7. The method for recommending the next point of interest based on continuous learning according to claim 1, characterized in that: The S4 comprises the following sub-steps: S41. Calculate update weight according to the consistency score of the user; S42, updating the memory entry for the user according to the update weight; S43, after completing the memory entry update, generating a final interest representation according to the user's consistency score; S44: Generate a final interest vector based on the long-term interest vector and the final interest representation to complete the recommendation of interest points.

8. The method for recommending the next point of interest based on continuous learning according to claim 7, characterized in that: In S41, the weight is updated The expression is: ； In the formula, Represents a predefined base value, Indicates user The consistency score, represents the average consistency score of all users, Represents the parameter that controls the adjustment sensitivity; In S42, the expression for updating the memory entry is: ； ； In the formula, represents the updated key, Represents the updated value. Indicates the matched key. Indicates the matched value. Indicates the newly entered key, Represents the value of the new input; In S43, the final interest expression The expression is: ； In the formula, Indicates a predefined base value; In S44, the final interest vector The expression is: ； In the formula, Recent user behavior. Represents the long-term interest vector.

Images