Generation-based interest point recommendation method and device, storage medium and terminal

Abstract

The invention discloses a generation-based interest point recommendation method and device, a storage medium and a terminal. The interest point recommendation method comprises the following steps: acquiring a historical access sequence of a user; creating a generative interest point recommendation model, wherein the model comprises a word segmentation device and a Transform architecture; a time expert network, a space expert network and a point-of-interest expert network are arranged in a decoder of the Transform architecture and are respectively used for time prediction, position prediction and point-of-interest recommendation; training the generative interest point recommendation model based on the historical access sequence; and predicting the current user by using the trained generative point-of-interest recommendation model to obtain a point-of-interest recommendation result of the current user. According to the method, complex interaction relations among time, space and semantic features can be fully mined, the limitation that geographic information is understood insufficiently in a traditional method is broken through, the model characterization capacity is enhanced, the problems of data offset and missing are effectively solved, and interest point recommendation is more accurate.

Description

Technical Field

[0001] This invention relates to the field of point-of-interest (POI) recommendation technology, and more specifically to generative POI recommendation methods, apparatus, storage media, and terminals. Background Technology

[0002] The core challenge of recommending the next point of interest lies in accurately understanding and utilizing the contextual information of user access behavior. In real-world scenarios, each access behavior contains multi-dimensional heterogeneous features such as time, space, and semantics. These heterogeneous features have complex interaction relationships, which significantly influence the user's next access choice.

[0003] In recent years, researchers have proposed various methods to effectively integrate this contextual information. From early feature engineering methods to recent deep learning models, feature utilization strategies have continuously evolved. Especially within the deep learning framework, by designing specialized feature interaction modules and attention mechanisms, the correlation between different features can be adaptively captured, thereby improving the model's ability to understand user access intentions.

[0004] However, the long-tailed distribution characteristics of heterogeneous information and the deep interactive dependencies between features remain challenging research areas that urgently need to be addressed. GeoSAN was the first to propose a dedicated geographic feature encoding mechanism, effectively capturing the local correlations of spatial information. STAN proposed a unified modeling framework for spatiotemporal features, achieving synergistic optimization of time intervals and geographic distances. However, existing methods still fall short in their understanding of geographic information. Most studies are limited to simple processing of distance information and superficial modeling of spatiotemporal correlations, failing to effectively capture the inherent cultural connotations, geographic semantics, and other essential attributes, and lacking a deep semantic understanding of location features. Secondly, existing methods often employ simple feature concatenation or shallow fusion strategies, making it difficult to fully capture the complex interaction patterns between temporal, spatial, and semantic features, and failing to effectively utilize the complementary information between features to enhance the model's representational capabilities. Moreover, existing sequence recommendation methods based on cross-entropy loss have limitations in their optimization objectives. These methods only perform simple ranking of candidate interest points during prediction, failing to fully consider the data bias and missing data problems commonly found in user access data, resulting in suboptimal model performance.

[0005] Therefore, how to address the shortcomings and deficiencies of existing technologies through effective point-of-interest recommendation methods has become an important issue that researchers in this field urgently need to solve. Summary of the Invention

[0006] The purpose of this invention is to address the above-mentioned problems by providing a generative point-of-interest recommendation method, apparatus, storage medium, and terminal.

[0007] The technical solution of this invention is: a generative point-of-interest recommendation method, comprising the following steps: obtaining user... Historical access sequence ,in For users In timestamp Access records Let be the sequence length, where and All are positive integers. Create a generative point of interest recommendation model, which includes a word segmenter and a Transformer architecture; the word segmenter is used to segment the historical access sequence. The token sequence is converted into a token sequence. The Transformer architecture includes an encoder and a decoder, which encode and decode the token sequence to generate a candidate interest point set. The decoder includes a time expert network, a spatial expert network, and an interest point expert network, used for time prediction, location prediction, and interest point recommendation, respectively. During prediction, based on the last hidden state of the decoder, the current prediction task type, and the feedforward network parameters corresponding to the current prediction task type, a nonlinear transformation is performed using an activation function. Then, probability normalization is used to obtain the probability distribution of the prediction result. Based on the historical access sequence... The generative point of interest recommendation model is trained; the trained generative point of interest recommendation model is used to recommend the current user. Prediction is performed to obtain the current user Interest-based recommendation results.

[0008] As an improvement to an embodiment of the present invention, the access record The features include: user ID, time ID, location ID, and point of interest ID; the word segmenter includes a word list aggregation unit and a vocabulary representation unit, the word list aggregation unit being used to segment the historical access sequence. Each access record The features are summarized to obtain the user vocabulary. Time vocabulary Location vocabulary and interest point vocabulary The vocabulary representation unit is used to represent the access records. The token sequence is obtained by lexical representation of the features. .

[0009] As an improvement to this embodiment of the invention, the prediction process of the decoder is as follows: ; ; , , , ;in, This is the hidden state of the last layer of the decoder. For the task The weights of the first layer feedforward network, For the task The first layer bias, For the task The weights of the second layer feedforward network, For the task The second layer bias, It is a task The corresponding vocabulary size; It is the hidden dimension of the generative interest point recommendation model. It is the intermediate layer dimension of the feedforward network.

[0010] As an improvement to this embodiment of the invention, when the time expert network makes predictions, the current prediction task type is a time task. It is the time lexicon The size; when the spatial expert network makes predictions, the current prediction task type is a location task. It is the positional vocabulary The size; when the interest point expert network makes predictions, the current prediction task type is an interest point task. It is the aforementioned interest point vocabulary. Size.

[0011] As an improvement to this embodiment of the invention, the word segmenter uses a vector quantization variational autoencoder (VQ-VAE) to obtain a positional vocabulary. .

[0012] As an improvement to this embodiment of the invention, the objective function of the generative point of interest recommendation model is the cross-entropy loss function, which is: ,in, Indicates task Authentic labels in vocabulary One-hot encoding on This indicates that the model predicts samples in the vocabulary. The probability of it.

[0013] As an improvement to this embodiment of the invention, the beam search method for location prediction in the generative point of interest recommendation model is an adaptive width beam search.

[0014] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides a generative point-of-interest recommendation device, comprising the following modules: a data acquisition module, used to acquire user... Historical access sequence ,in For users In timestamp Access records Let be the sequence length, where and All are positive integers. The model creation module is used to create a generative interest point recommendation model, which includes a word segmenter and a Transformer architecture; the word segmenter is used to segment the historical access sequence. The token sequence is converted into a token sequence. The Transformer architecture includes an encoder and a decoder, which encode and decode the token sequence to generate a candidate interest point set. The decoder includes a time expert network, a spatial expert network, and an interest point expert network, used for time prediction, location prediction, and interest point recommendation, respectively. During prediction, based on the last hidden state of the decoder, the current prediction task type, and the feedforward network parameters corresponding to the current prediction task type, a nonlinear transformation is performed using an activation function. Then, probability normalization is used to obtain the probability distribution of the prediction result. A model training module is used to train the model based on the historical access sequence. The generative point of interest recommendation model is trained to obtain a generative point of interest recommendation model; the point of interest recommendation module is used to use the generative point of interest recommendation model to recommend the current user. Prediction is performed to obtain the current user Interest-based recommendation results.

[0015] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides a storage medium storing program instructions, which, when executed, implement the point of interest recommendation method as described in any of the preceding claims.

[0016] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides a terminal, including a processor and a memory, wherein the memory stores program instructions, and the processor executes the program instructions to implement the point of interest recommendation method as described in any of the preceding claims.

[0017] The generative point-of-interest (POI) recommendation method, apparatus, storage medium, and terminal provided in this invention have the following advantages: This invention uses a word segmenter to represent heterogeneous features in user access sequences lexically, and sets up temporal, spatial, and POI expert networks in the Transformer architecture decoder for prediction, fully exploring the complex interaction relationships between temporal, spatial, and semantic features, overcoming the limitations of traditional methods in understanding geographic information, and enhancing the model's representation ability; and uses adaptive width beam search to effectively solve the problems of data offset and missing data, making POI recommendation more accurate. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the generative point-of-interest recommendation method described in this invention. Figure 2 This is a schematic diagram of the architecture of the generative interest point recommendation model described in this invention; Figure 3 This is a comparison chart of the performance of the generative interest point recommendation model described in this invention and the traditional model; Figure 4 This is a schematic diagram of the generative point-of-interest recommendation device described in this invention; Figure 5 This is a schematic diagram of the structure of the electronic terminal described in this invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0020] If the present invention involves orientation (e.g., up, down, left, right, front, back, outside, inside, etc.) when described, then the orientations involved need to be defined.

[0021] The scope of the embodiments described herein includes the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another without requiring or implying any actual relationship or order between the elements. Indeed, a first element can also be referred to as a second element, and vice versa. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a structure, apparatus, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a structure, apparatus, or device. Without further limitations, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the structure, apparatus, or device that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

[0022] The terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" used in this document to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings and are used only for the convenience of describing this document and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. In the description herein, unless otherwise specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements, or direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0023] Embodiment 1 of this invention provides a generative point-of-interest recommendation method, such as... Figure 1 As shown, it includes the following steps: Step 101: Obtain User Historical access sequence ,in For users In timestamp Access records Let be the sequence length, where and All are positive integers. ; In practice, the user Historical access sequence Data can be obtained from publicly available datasets such as FourSquare-NYC, FourSquare-TKY, or Gowalla-CA. FourSquare-NYC represents the New York area, containing 1075 users, 5099 points of interest, and 104074 interaction records, with a sparsity of 98.10%, suitable for geospatial analysis and user behavior modeling. FourSquare-TKY represents the Tokyo area, containing 2281 users, 7844 points of interest, and 361430 interactions, with a sparsity of 97.98%, helpful for understanding Tokyo user preferences. Gowalla-CA, built by Stanford University based on user social media check-in data, contains 4318 users, 9923 points of interest, and 250780 interactions, with a sparsity of 99.41%.

[0024] After acquiring the data, it needs to be cleaned. Access records lacking key identifiers, such as user IDs or point-of-interest (POI) IDs, should be removed because the association between the user and the POI cannot be clearly established. If the POI location information is abnormal, it should first be compared and corrected with surrounding coordinates and map data; abnormal records that cannot be verified and corrected should be removed. Simultaneously, timestamps should be checked to ensure correct format and within a reasonable time range to avoid time errors. After cleaning, the access records for each user are sorted by timestamp. Considering the research objectives, data characteristics, and the impact of different sequence lengths on accuracy, recall, and other metrics during model training and validation, a suitable sequence length N is determined. Corresponding segments are then extracted from the sorted records to obtain the final historical access sequence. .

[0025] Step 102: Create a generative point-of-interest recommendation model, such as Figure 2 As shown, the generative interest point model includes a word segmenter and a Transformer architecture; the word segmenter is used to segment the historical access sequence. The token sequence is converted into a token sequence. The Transformer architecture includes an encoder and a decoder. The Transformer architecture is used to encode and decode the token sequence and generate a set of candidate interest points. The decoder is equipped with a time expert network, a spatial expert network, and an interest point expert network, which are used for time prediction, location prediction, and interest point recommendation, respectively. During prediction, based on the last hidden state of the decoder, the current prediction task type, and the feedforward network parameters corresponding to the current prediction task type, a nonlinear transformation is performed using an activation function. Then, the probability distribution of the prediction result is obtained through probability normalization. Here, the access record The features include: user ID, time ID, location ID, and point of interest ID; the word segmenter includes a word list aggregation unit and a vocabulary representation unit, the word list aggregation unit being used to segment the historical access sequence. Each access record The features are summarized to obtain the user vocabulary. Time vocabulary Location vocabulary and interest point vocabulary The vocabulary representation unit is used to represent the access records. The token sequence is obtained by lexical representation of the features. .

[0026] Preferably, the word segmenter uses a vector quantization variational autoencoder (VQ-VAE) to obtain a positional vocabulary. Specifically, VQ-VAE first uses an encoder The feature vector of the location ID Mapped to 3D continuous latent space is obtained Subsequently, in the context of... Codebook of code vectors Searching for and nearest neighbor discrete codes This quantization process can be represented as a discrete index mapping. Finally, the decoder The quantized representation is reconstructed into the original feature space. The training objective of VQ-VAE , , ,in, To rebuild the losses, To quantify the loss, This indicates a stop-gradient operation. These are coefficients used to balance the two losses. The positional vocabulary is obtained by mapping the positional IDs to the feature vector space through the vector quantization variational autoencoder (VQ-VAE). For time IDs, the word segmenter uses hierarchical quantization to divide a day into four time periods: morning, afternoon, evening, and late night, and constructs a time vocabulary of size 28 based on the seven days of the week. The user IDs and point-of-interest IDs are converted into corresponding tags through a one-to-one mapping to obtain the user vocabulary. and interest point vocabulary It is understandable that a positional vocabulary is constructed using VQ-VAE. It can map high-dimensional sparse location IDs to a low-dimensional continuous latent space, significantly compressing data dimensionality while preserving key semantic information and improving model processing efficiency. The quantization process forces the discriminativeness and semantic consistency of location features, making similar geographical locations more closely clustered in the feature space, which is conducive to the model capturing spatial distribution patterns. The unified vocabulary mapping provides standardized input for the Transformer architecture, effectively alleviating the problem of long-tail distribution of heterogeneous information, optimizing the model's understanding of users' historical access sequences, and improving the accuracy and rationality of interest point recommendations.

[0027] In practice, a word segmenter is first used to map the historical access sequence to a vocabulary based on user ID, time ID, location ID, and point of interest ID. <u-u> 、 <t-t> 、 <l-l> 、 <p-p>A sequence of tokens in a specific format; then construct a Transformer architecture containing an encoder and decoder, such as... Figure 2 As shown, the encoder consists of a multi-layered substructure consisting of a normalization layer, a feedforward network layer, a normalization layer, and a self-attention layer, which encodes the token sequence and outputs global semantic features. The decoder, through a multi-layered substructure consisting of a normalization layer, a feedforward network layer, a cross-attention layer, and a self-attention layer, and combined with a temporal expert network, a spatial expert network, and an interest point expert network, adapts to multiple tasks. During prediction, it calls the hidden state of the last layer of the decoder, loads the corresponding feedforward network parameters according to the task type, and outputs the probability distribution of the prediction result after nonlinear transformation and probability normalization by the activation function.

[0028] Preferably, the prediction process of the decoder is as follows: ; ; , , , ;in, This is the hidden state of the last layer of the decoder. For the task The weights of the first layer feedforward network, For the task The first layer bias, For the task The weights of the second layer feedforward network, For the task The second layer bias, It is a task The corresponding vocabulary size; It is the hidden dimension of the generative interest point recommendation model. This is the intermediate layer dimension of the feedforward network. When the time expert network makes predictions, the current prediction task type is a time-based task. It is the time lexicon The size; when the spatial expert network makes predictions, the current prediction task type is a location task. It is the positional vocabulary The size; when the interest point expert network makes predictions, the current prediction task type is an interest point task. It is the aforementioned interest point vocabulary. Size.

[0029] Understandably, the decoder assigns corresponding feedforward network parameters and vocabularies to different tasks, enabling the model to accurately adapt to the data distribution and semantic requirements of each task. This unified architecture ensures efficient collaboration, guarantees flexibility in multi-task processing, strengthens feature interaction and semantic association, improves the accuracy and relevance of predictions for each task, and makes interest point recommendations more aligned with users' actual behavioral logic.

[0030] In this embodiment, the objective function of the generative point of interest recommendation model is the cross-entropy loss function, which is as follows: ,in, Indicates task Authentic labels in vocabulary One-hot encoding on This indicates that the model predicts samples in the vocabulary. The probability of the predicted label is calculated using the cross-entropy loss function. Here, the cross-entropy loss function accurately measures the fit between the model's predicted probability distribution and the true label distribution, adapting to multi-task classification scenarios, guiding the model to optimize parameters, and improving the accuracy of word prediction for different tasks.

[0031] Step 103: Based on the historical access sequence The generative point of interest recommendation model is trained. During training, historical access sequences are divided into training, validation, and test sets in an 8:1:1 ratio. Initial parameters such as VQ-VAE codebook size, VQ-VAEBeta, Dropout mask, sequence length N, and temperature coefficient are then set. The VQ-VAE codebook size determines the granularity of positional semantic characterization, balancing accuracy and cost. VQ-VAEBeta adjusts the balance between reconstruction and quantization loss, adapting to semantic preservation requirements. The Dropout mask randomly masks neurons to enhance generalization and prevent overfitting. The sequence length N adapts to user access patterns, balancing behavioral dependence and computational noise. The temperature coefficient adjusts the smoothness of the probability distribution. During training, samples from the training set are input into the model in batches. Error is calculated using cross-entropy loss, and parameters are updated via backpropagation. The validation set is used to assess performance in each round, and hyperparameters are adjusted accordingly. After training, the test set is used for evaluation, and accuracy and other metrics are used to judge the recommendation effectiveness.

[0032] A comparison of the recommendation performance of the GenNext generative interest point recommendation model described in this invention with that of traditional models is as follows: Figure 3 As shown in the comparison, the improvements in Acc and MRR for GenNext on each dataset when a is 5, 10, and 20 are between 6.4.0%-7.8%, 2.3%-24.9%, 3.2%-40.3%, and 1.6%-5.8%, respectively. The overall performance advantage is significant, indicating that it can more accurately predict user interests in the interest point recommendation task and provide users with more suitable recommendation results. Preferably, the model parameters selected for the above three datasets are shown in Table 1.

[0033] Table 1. Model parameter selection for each dataset

[0034] Step 104: Use the trained generative interest point recommendation model to recommend the current user. Prediction is performed to obtain the current user Interest-based recommendation results.

[0035] In practice, a test set can be used to test the generative point of interest recommendation model and evaluate its performance based on the test results. The test set includes the current user... The historical access sequence is input into the generative point of interest recommendation model. Preferably, the beam search method for location prediction in the generative point of interest recommendation model is an adaptive width beam search, and the beam width function is: ,when When predicting the location of a point of interest, the beam width is The rest are This is to adapt to the needs of different candidate exploration ranges. Then, according to the formula... From the candidate sequence set Based on predicted probability Filter out the top with the highest probability The candidate sequences are used as the current user. Interest-based recommendation results.

[0036] Understandably, the adaptive width beam search increases the beam width to ensure sufficient search coverage for key interest point prediction locations, while maintaining a normal beam width at other locations. This avoids the waste of computational resources caused by large beam widths during full sequence decoding. Compared to traditional beam search, it significantly improves the accuracy of interest point prediction while maintaining a similar inference speed, achieving a balance between search efficiency and recommendation quality.

[0037] Embodiment 2 of the present invention provides a generative point-of-interest recommendation device, such as... Figure 3 As shown, it includes the following modules: Data acquisition module 401 is used to acquire user data. Historical access sequence ,in For users timestamp Access records Let be the sequence length, where and All are positive integers. ; Model creation module 402 is used to create a generative interest point recommendation model, which includes a word segmenter and a Transformer architecture; the word segmenter is used to segment the historical access sequence The token sequence is converted into a token sequence. The Transformer architecture includes an encoder and a decoder. The Transformer architecture is used to encode and decode the token sequence and generate a set of candidate interest points. The decoder is equipped with a time expert network, a spatial expert network, and an interest point expert network, which are used for time prediction, location prediction, and interest point recommendation, respectively. During prediction, based on the last hidden state of the decoder, the current prediction task type, and the feedforward network parameters corresponding to the current prediction task type, a nonlinear transformation is performed using an activation function. Then, the probability distribution of the prediction result is obtained through probability normalization. Model training module 403 is used to train models based on the historical access sequence. The generative interest point recommendation model is trained to obtain the generative interest point recommendation model; The point of interest recommendation module 404 is used to recommend points of interest to the current user using the generative point of interest recommendation model. Prediction is performed to obtain the current user Interest-based recommendation results.

[0038] Embodiment 3 of the present invention provides a storage medium storing program instructions, which, when executed, implement the point of interest recommendation method as described in any of the preceding embodiments.

[0039] Embodiment 4 of the present invention provides a terminal, such as Figure 4 As shown, it includes a processor and a memory, the memory storing program instructions, and the processor executing the program instructions to implement the point of interest recommendation method as described in any of the preceding embodiments.

[0040] This invention can be an apparatus, method, and / or computer program product. A computer program product may include a readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0041] Storage media can be tangible devices that hold and store instructions for use by instruction execution devices. Storage media can include, but are not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0042] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0043] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention. < / l-l> < / t-t> < / u-u>

Claims

1. A generative-based point of interest recommendation method, characterized by, The method comprises the following steps: Acquiring users Historical access sequence ,in For users timestamp Access records Let be the sequence length, where and All are positive integers. ; creating a generative point-of-interest recommendation model, the generative point-of-interest model comprising a tokenizer and a Transformer architecture; The tokenizer is configured to convert the historical access sequence into a token sequence; the Transformer architecture comprising an encoder and a decoder, the Transformer architecture being used for encoding and decoding the token sequence and generating a candidate point-of-interest set; a time expert network, a space expert network and a point-of-interest expert network are arranged in the decoder, and are respectively used for time prediction, location prediction and point-of-interest recommendation; when predicting, based on the last layer hidden state of the decoder, the current prediction task type and the feedforward network parameters corresponding to the current prediction task type, a nonlinear transformation is completed by using an activation function, and then a probability distribution of a prediction result is obtained by probability normalization; based on the historical access sequence training the generative point of interest recommendation model; The trained generative point of interest recommendation model is used to predict a point of interest recommendation result of the current user. The trained generative point of interest recommendation model is used to predict a point of interest recommendation result of the current user. The trained generative point of interest recommendation model is used to predict a point of interest recommendation result of the current 2.The point of interest recommendation method of claim 1, wherein, The access record The features in the access record include: user ID, time ID, location ID, and point of interest ID. The word segmenter comprises a word list summarizing unit and a word representation unit, the word list summarizing unit is used for summarizing the features of each access record in the historical access sequence to obtain a user word list , a time word list , a location word list and a point of interest word list ; and the word representation unit is used for word representation of the features of the access record to obtain a token sequence . The word segmenter comprises a word list summarizing unit and a word representation unit, the word list summarizing unit is used for summarizing the features of each access record in the historical access sequence to obtain a user word list , a time word list , a location word list and a point of interest word list ; and the word representation unit is used for word representation of the features of the access record to obtain a token sequence . The word segmenter comprises a word list summarizing unit and a word 3.The point of interest recommendation method of claim 2, wherein, the prediction process of the decoder is as follows: ； ； ， ， ， ； wherein, is the hidden state of the last layer of the decoder, is the first layer feedforward network weight of the task , is the first layer bias of the task , is the second layer feedforward network weight of the task , is the second layer bias of the task , is the vocabulary size of the task ; is the hidden dimension of the generative point of interest recommendation model, is the intermediate layer dimension of the feedforward network. 4.The point of interest recommendation method of claim 3, wherein, The time expert network predicts when, the current prediction task type is a time task, is the size of the time vocabulary The space expert network predicts where, the current prediction task type is a location task, is the size of the location vocabulary The point of interest expert network predicts what, the current prediction task type is a point of interest task, is the size of the point of interest vocabulary . 5.The point of interest recommendation method of claim 2, wherein, The tokenizer uses a vector-quantized variational autoencoder, VQ-VAE, to obtain a position vocabulary . 6.The point of interest recommendation method of claim 3, wherein, The objective function of the generated interest point recommendation model is a cross-entropy loss function, and the cross-entropy loss function is: Wherein, Indicates the true label of the task One-hot encoding on the vocabulary , Indicates the probability of the model predicting the sample on the vocabulary . 7.The point of interest recommendation method of claim 1, wherein, an adaptive width beam search is used for the beam search method of the location prediction in the generative point-of-interest recommendation model. 8.A point of interest recommendation apparatus based on generative, the apparatus comprising: The program instructions are executed to implement the point-of-interest recommendation method according to any one of claims 1 to 7. A data acquisition module is configured to acquire a historical access sequence of a user wherein is an access record of the user at a time stamp , and is a sequence length, wherein and are both positive integers, ;​ The model creation module is used to create a generative point of interest recommendation model, which includes a word segmenter and a Transformer architecture; the word segmenter is used to segment the historical access sequence. The token sequence is converted into a token sequence. The Transformer architecture includes an encoder and a decoder. The Transformer architecture is used to encode and decode the token sequence and generate a set of candidate interest points. The decoder is equipped with a time expert network, a spatial expert network, and an interest point expert network, which are used for time prediction, location prediction, and interest point recommendation, respectively. During prediction, based on the last hidden state of the decoder, the current prediction task type, and the feedforward network parameters corresponding to the current prediction task type, a nonlinear transformation is performed using an activation function. Then, the probability distribution of the prediction result is obtained through probability normalization. a model training module, configured to train the generative point of interest recommendation model based on the historical access sequence training the generative point of interest recommendation model to obtain a generative point of interest recommendation model; The point-of-interest recommendation module is used to recommend points of interest to the current user using the generative point-of-interest recommendation model. Prediction is performed to obtain the current user Interest-based recommendation results.

9. A storage medium storing program instructions, characterized in that, The processor executes the program instructions to implement the point-of-interest recommendation method according to any one of claims 1 to 7.

10. A terminal, characterized by comprising: ​

Images