Hierarchical semantic label alignment based generative point of interest recommendation method and system

Abstract

The application discloses a generative interest point recommendation method and system based on hierarchical semantic identification alignment, relates to the technical field of artificial intelligence, and mainly aims to solve the problem of low accuracy of existing interest point recommendation. Mainly includes constructing a bidirectional alignment task according to a multi-level semantic identification and a structured semantic portrait, and constructing an interest point recommendation training sample according to the multi-level semantic identification of each check-in point in the sample user check-in point access sequence; based on the bidirectional alignment task and the interest point recommendation training sample, a large model is sequentially fine-tuned and trained to obtain an interest point recommendation model with multi-level semantic identification and semantic feature mutual identification capability; in response to an interest point recommendation request, the multi-level semantic identification sequence of the historical check-in point access sequence of a target user is predicted by the interest point recommendation model to generate a target recommended interest point. Mainly used for generating interest point recommendation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology and can be applied to scenarios such as geolocation services and intelligent recommendations. In particular, it relates to a generative point of interest recommendation method and system based on hierarchical semantic tag alignment. Background Technology

[0002] Point-of-interest (POI) recommendations predict a user's next likely location based on recent check-ins or browsing history, providing personalized location suggestions for mobile users and reducing information overload in location-based services. This method is widely used in local services, smart navigation, travel planning, and business recommendations.

[0003] Existing interest point recommendation methods typically input interest points as ordinary discrete numbers into recurrent neural networks, attention networks, graph neural networks, or neural network models based on self-attention mechanisms. While these methods can characterize certain trajectory sequence dependencies and spatiotemporal correlations, the original interest point numbers are usually generated by database order or random rules. The model struggles to derive transferable knowledge from the numbers themselves, resulting in insufficient generalization ability in scenarios involving large-scale candidate sets, sparse user behavior, new regions, or long-tail interest points, thus leading to low recommendation accuracy. Summary of the Invention

[0004] In view of this, the present invention provides a generative point of interest recommendation method and system based on hierarchical semantic tag alignment, the main purpose of which is to solve the problem of low accuracy of existing point of interest recommendations.

[0005] According to one aspect of the present invention, a generative interest point recommendation method based on hierarchical semantic identifier alignment is provided, comprising: Sample data is constructed based on the multidimensional feature data of the expected check-in points. Multi-granularity index encoding is performed on the multidimensional features of each sample check-in point in the sample data, and the indexes of different granularities obtained by encoding are used as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers of each sample check-in point. Multi-granularity semantic feature extraction is performed on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. Based on the multi-level semantic identifiers and the structured semantic profile, a bidirectional alignment task is constructed, and based on the multi-level semantic identifiers of each check-in point in the sample user check-in point access sequence, interest point recommendation training samples are constructed. The bidirectional alignment task is used to enable the large model to map the multi-level semantic identifiers and the structured semantic profile to each other. Based on the bidirectional alignment task and the interest point recommendation training samples, the large model is finely trained in sequence to obtain an interest point recommendation model with multi-level semantic identification and semantic feature mutual recognition capabilities. In response to a point of interest recommendation request, the point of interest recommendation model is used to predict the target point of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence, and generate the target recommended point of interest.

[0006] According to another aspect of the present invention, a generative point of interest recommendation system based on hierarchical semantic tag alignment is provided, comprising: The feature encoding module is used to construct sample data based on the multidimensional feature data of the expected check-in points, perform multi-granularity index encoding on the multidimensional features of each sample check-in point in the sample data, and use the encoded indices of different granularities as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers for each sample check-in point; and perform multi-granularity semantic feature extraction on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. The sample construction module is used to construct a bidirectional alignment task based on the multi-level semantic identifiers and the structured semantic profile, and to construct interest point recommendation training samples based on the multi-level semantic identifiers of each check-in point in the sample user check-in point access sequence. The bidirectional alignment task is used to enable the large model to map the multi-level semantic identifiers and the structured semantic profile to each other. The model training module is used to fine-tune the large model sequentially based on the bidirectional alignment task and the interest point recommendation training samples to obtain an interest point recommendation model with multi-level semantic labeling and semantic feature mutual recognition capabilities. The recommendation generation module is used to respond to the point of interest recommendation request by predicting the target point of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence through the point of interest recommendation model, and generating the target recommended point of interest.

[0007] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the generative point of interest recommendation method based on hierarchical semantic tag alignment described above.

[0008] According to another aspect of the present invention, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform the operations corresponding to the generative point of interest recommendation method based on hierarchical semantic identifier alignment described above.

[0009] By employing the above technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a generative point of interest recommendation method and system based on hierarchical semantic identifier alignment. In this embodiment, sample data is constructed based on the multidimensional feature data of expected check-in points. Multi-granularity indexing is performed on the multidimensional features of each sample check-in point in the sample data, and the resulting indices of different granularities are used as semantic identifiers for the corresponding levels of the sample check-in points, resulting in multi-level semantic identifiers for each sample check-in point. Multi-granularity semantic feature extraction is performed on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. Based on the multi-level semantic identifiers and structured semantic profiles, a bidirectional alignment algorithm is constructed. Based on the multi-level semantic identifiers of each check-in point in the sample user's check-in point access sequence, a training sample for interest point recommendation is constructed. The bidirectional alignment task enables the large model to map multi-level semantic identifiers to structured semantic profiles. The large model is then fine-tuned based on the bidirectional alignment task and the interest point recommendation training sample to obtain an interest point recommendation model capable of mutual recognition between multi-level semantic identifiers and semantic features. In response to an interest point recommendation request, the interest point recommendation model predicts recommended interest points based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence, generating target recommended interest points. By replacing the original interest point numbers with multi-level semantic identifiers, semantically similar interest points share semantic identifiers, alleviating the problem of semantic sparsity in numbering. Furthermore, through fine-tuning training using the bidirectional alignment task and the interest point recommendation training sample, the model can accurately understand semantic identifiers and perform accurate interest point recommendations, thereby significantly improving the accuracy of interest point recommendations.

[0010] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. Furthermore, in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0011] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 The flowchart of a generative point of interest recommendation method based on hierarchical semantic identifier alignment provided by an embodiment of the present invention is shown. Figure 2 The flowchart illustrates a method for fine-tuning training based on the bidirectional alignment task and the interest point recommendation training samples provided by an embodiment of the present invention. Figure 3The diagram shows a block diagram of a generative interest point recommendation system based on hierarchical semantic identifier alignment provided by an embodiment of the present invention. Figure 4 A schematic diagram of the structure of a terminal provided in an embodiment of the present invention is shown. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] To address the issue of low accuracy in existing point-of-interest (POI) recommendations, this invention provides a generative POI recommendation method based on hierarchical semantic tag alignment. This method generates hierarchical semantic tags through multi-source attributes of POIs, aligns these semantic tags to the representation and generation spaces of a generative sequence model through a multi-stage bidirectional task, and directly improves the POI recommendation quality of a large language model through supervised hot-start and reinforcement optimization, thereby enhancing the accuracy of POI recommendations. This method can be applied to scenarios such as local life services, intelligent navigation, advertising candidate generation, travel itinerary planning, and business recommendations. The executable instructions of this method can be deployed on servers, cloud platforms, mobile terminals, or edge computing devices, or implemented through program instructions in computer storage media.

[0014] like Figure 1 As shown, the method includes steps 101-104: steps 101-103 correspond to the training stage of the large model, and step 104 corresponds to the practical application stage of recommending interest points based on the large model that has been trained.

[0015] 101. Construct sample data based on the multidimensional feature data of the expected check-in points, perform multi-granularity index encoding on the multidimensional features of each sample check-in point in the sample data, and use the encoded indexes of different granularities as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers for each sample check-in point; extract multi-granularity semantic features from the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point.

[0016] In this embodiment of the invention, the expected check-in point is the location that the user expects to visit. It can recommend any accessible location within the target area covered by the point of interest, such as hotels, restaurants, scenic spots, shopping malls, barbershops, etc. The feature dimensions of the check-in point include category features of the location attribute dimension, spatial features of the spatial attribute dimension, and time features of the access cycle dimension. Specifically, the category feature can be obtained by encoding the business category name, category hierarchy text, merchant tag, or point of interest description text to which the check-in point belongs; the spatial feature can be obtained by latitude and longitude, geographic grid, administrative region, or spatial clustering results; and the time feature can be obtained by statistics of hours, days of the week, holidays, or other access cycles.

[0017] To construct a continuous, multi-layered representation of check-in points, multi-granularity indexing and encoding are performed on the multi-dimensional features of each sample check-in point to obtain indices of different granularities. Each level of granularity index serves as a semantic identifier for a specific level, ultimately resulting in multi-layered semantic identifiers for the sample check-in points. Simultaneously, for each sample check-in point, deep semantic mining and high-order feature reconstruction are performed based on the multi-dimensional features to organically link scattered, low-level heterogeneous information such as spatial coordinates, check-in frequency, temporal patterns, and user composition, extracting highly interpretable high-order semantic labels. Finally, these labels are integrated into each sample check-in point according to a pre-defined structured framework, creating a complete semantic profile. This profile can describe the functional attributes, dynamic operational status, and user group preferences of check-in points in a model-readable format, providing accurate and quantifiable support for personalized check-in point recommendations for different users.

[0018] In one embodiment of the present invention, for further explanation and limitation, the multi-level semantic identifier construction process for any check-in point includes: Category features are extracted based on the category description text of the check-in point, spatial features are extracted based on the geographical location information and geographical affiliation of the check-in point, and time features are extracted based on the access statistics of the check-in point. The category features, spatial features and time features are then fused to obtain multi-dimensional fused features. The multidimensional fusion features are residual quantized and encoded by a residual quantization variational autoencoder to obtain a multi-level codeword index with granularity from coarse to fine. Each level of the codeword index in the multi-level codeword index is used as a hierarchical semantic identifier of the same granularity to obtain the multi-level semantic identifier of the check-in point. After generating multi-level semantic identifiers for all check-in points, the method further includes: Extract conflicting check-in points that have the same multi-level semantic identifiers, and add conflict suffixes to the multi-level semantic identifiers of the conflicting check-in points to distinguish each conflicting check-in point based on the conflict suffixes.

[0019] In this embodiment of the invention, firstly, the functional attribute category features of the check-in point are extracted from the category description text of the check-in point through text encoding or keywords. Simultaneously, spatial location features of the check-in point are extracted based on its latitude and longitude coordinates and geographical affiliation information such as administrative divisions. Furthermore, based on access statistics such as the frequency of visits and the distribution of active periods in different time periods, the temporal periodic features of the check-in point are extracted. Finally, these three types of heterogeneous features are organically integrated through vector concatenation or attention fusion to form a multi-dimensional fusion feature representation covering semantic attributes, spatial location, and temporal variation patterns, thereby providing comprehensive and complementary information support for the subsequent identification and encoding process of the check-in point.

[0020] In the encoding stage, a residual quantization variational autoencoder is used as the encoding tool. First, the encoder maps the multidimensional fused features of the check-in points into latent vectors. Then, a multi-layer residual quantization module decomposes and quantizes these latent vectors step-by-step: the first quantization layer starts from the codebook... Select the closest code The first-level codeword index is obtained. ; the residual Input to the second quantization layer, from the codebook Select code The second-level codeword index is obtained. And so on, continuing until the [number]th [number]th [number]. The layers ultimately generate a multi-level codeword index sequence arranged from coarse to fine. Each level of codeword index in this sequence is considered a hierarchical semantic identifier at a corresponding granularity level. These identifiers, arranged in a macro-to-micro hierarchical relationship, collectively constitute a multi-level semantic identifier system for check-in points. Higher-level codewords reflect the broad category of the check-in point, while lower-level codewords characterize its subcategories or individual characteristics, thus achieving a structured representation of the check-in point's semantics at different levels of abstraction. It should be noted that the core innovation of this scheme lies not in the discrete encoding process itself, but in combining the multi-level semantic identifier system with subsequent large-scale model semantic alignment, Top-k hot-start, and reinforcement optimization strategies to form a complete technical process for interest-point recommendation.

[0021] After encoding, all check-in point identifiers are traversed. If multiple check-in points sharing the same multi-level semantic identifier are found, a conflict suffix is ​​added to these conflicting check-in points to distinguish them. This suffix can be a globally unique identifier such as a unique ID or an auto-incrementing sequence number for the check-in point in the original database. During the model training phase, the multi-level semantic identifiers of the check-in points in the training samples, as well as during the interest point recommendation phase, are constructed based on the above method.

[0022] 102. Based on the multi-level semantic identifiers and the structured semantic profile, construct a bidirectional alignment task, and based on the multi-level semantic identifiers of each check-in point in the sample user check-in point access sequence, construct interest point recommendation training samples.

[0023] In this embodiment of the invention, to enable a large language model to achieve bidirectional mutual recognition between multi-level semantic tags and structured semantic profiles, a bidirectional alignment task is constructed. During the execution of this task, the large model learns the mapping relationship between multi-level semantic tags and structured semantic profiles, thereby achieving bidirectional mutual recognition. The structured semantic profile includes multiple levels: Level 1 includes main categories or category clusters; Level 2 includes categories and coarse-grained regions; and Level 3 includes categories, coarse-grained regions, and local geographic buckets. This hierarchical granularity corresponds to the hierarchical granularity of the multi-level semantic tags. The bidirectional alignment task, as the name suggests, involves two directions: one requires the large model to generate corresponding structured semantic profiles based on multi-level semantic tags; the other requires the large model to generate corresponding multi-level semantic tags based on structured semantic profiles. The bidirectional alignment task can be a task that mutually recognizes semantic tags and semantic profiles at all granularities or levels, or it can be a task that mutually recognizes semantic tags and semantic profiles at only some granularities or levels.

[0024] The training samples for point-of-interest (POI) recommendation are constructed based on the historical check-in point access trajectories of a large number of users. For example, a user's historical access trajectory sequentially passes through five check-in points A, B, C, D, and E. Check-in points A, B, C, and D are extracted to construct the sample user's check-in point access sequence, and each check-in point in the sequence is encoded as a multi-level semantic identifier. The sequentially concatenated multi-level semantic identifiers of A, B, C, and D constitute the input part of the sample. The target output part of the sample is a pseudo-multi-level semantic identifier list containing the multi-level semantic identifiers corresponding to the target recommended check-in point (which could be check-in point E in the example above). The first position in this list is the multi-level semantic identifier of the target recommended check-in point. Then, pseudo-multi-level semantic identifiers sharing different levels of semantic identifiers with this identifier are constructed sequentially to complete the required number of identifiers in the list. For example, if the list selects the top 10, it is completed to 10. Training a large model based on the POI recommendation training samples enables the large model to learn a fixed number of valid and semantically related identifier lists for output.

[0025] In one embodiment of the present invention, for further illustration and limitation, the construction process of the bidirectional alignment task includes: The coarse-grained semantic identifier and the corresponding granular semantic profile are constructed into a set of coarse-grained aligned sample pairs that are mutually target outputs, and the bidirectional recognition task of the coarse-grained aligned sample pairs is constructed as a first-stage bidirectional alignment task. The multi-level semantic identifiers and the structured semantic profiles are constructed into a set of globally aligned sample pairs that are mutually target outputs, and the bidirectional recognition task of the globally aligned sample pairs is constructed as a two-stage bidirectional alignment task. The fine-grained semantic identifiers and corresponding granular semantic profiles are constructed into a set of fine-grained aligned sample pairs that are mutually target outputs, and the bidirectional recognition task of the fine-grained aligned sample pairs is constructed as a three-stage bidirectional alignment task.

[0026] In this embodiment of the invention, a three-stage progressive bidirectional alignment task is constructed to gradually establish the mapping relationship between multi-level semantic identifiers and structured semantic profiles. In the first stage, coarse-grained semantic identifiers, such as region-level and category cluster-level identifiers, are constructed with corresponding granular semantic profiles as mutually target output alignment sample pairs. A bidirectional recognition task is designed to enable the model to prioritize establishing macro-level semantic associations, providing stable high-order semantic anchors for subsequent fine-grained alignment. In the second stage, complete semantic identifiers covering all levels and structured semantic profiles are constructed with global alignment sample pairs, and a bidirectional recognition task is executed. This allows the model to further learn the complex correspondence between all semantic information and the profile based on the coarse-grained mapping, achieving cross-level overall semantic alignment. In the third stage, fine-grained semantic identifiers, such as sub-class-level, attribute-level, or local geographic buckets, are constructed with corresponding granular semantic profiles as fine-grained alignment sample pairs. A corresponding bidirectional recognition task is also constructed to guide the model to focus on the precise mapping between refined semantic features and profile details, based on macro- and global alignment. Thus, in the subsequent model fine-tuning training based on the above tasks, the model can gradually establish cross-level semantic recognition and mapping capabilities through progressive training from coarse to fine.

[0027] The bidirectional alignment task includes a one-stage bidirectional alignment task, a two-stage bidirectional alignment task, and a three-stage bidirectional alignment task. Among them, multi-level semantic identifiers encompass at least one level of coarse-grained semantic identifiers and at least one level of fine-grained semantic identifiers. The image level of the structured semantic profile corresponds one-to-one with the level granularity of the multi-level semantic identifiers; that is, multi-level semantic identifiers include at least one level of coarse-grained semantic identifiers and at least one level of fine-grained semantic identifiers, and the semantic profile level of the structured semantic profile corresponds to the level granularity of the multi-level semantic identifiers. Taking a hot pot-related point of interest as an example, its complete semantic identifier is [Northeast Region, Catering, Hot Pot, Sichuan Spicy]. The structured semantic profile includes attributes such as longitude, latitude, average consumption per person, rating, and cuisine tag. In the coarse-grained stage, the input is [Northeast Region, Catering], and the target output is a macro-level profile description, such as "Region Type: Urban Commercial Center; Major Category Feature: Catering Service; Average Consumption Range: 50-150 yuan"; correspondingly, the reverse sample inputs the macro-level profile description and outputs [Northeast Region, Catering]. In the global phase, the input is a complete identifier [Northeast China region, catering, hot pot, Sichuan spicy], and the target output is a complete profile vector, including latitude and longitude, consumption, rating, etc.; correspondingly, the reverse sample inputs the profile vector and outputs the complete identifier. In the fine-grained phase, the input is [hot pot, Sichuan spicy], and the target output is specific attribute details, such as "flavor: spicy; typical dishes: tripe, beef aorta; suitable scenario: group dining"; correspondingly, the reverse sample inputs the above specific attribute details and outputs [hot pot, Sichuan spicy]. By constructing structured semantic profiles and multi-level semantic identifiers in the above manner, it is possible to ensure that the alignment tasks at each stage are carried out at the same semantic granularity to construct the mutual target output.

[0028] 103. Based on the bidirectional alignment task and the interest point recommendation training samples, fine-tune the training model sequentially to obtain an interest point recommendation model with multi-level semantic labeling and semantic feature mutual recognition capabilities.

[0029] In this embodiment of the invention, a two-way alignment task is used to fine-tune the basic large model in the first stage, enabling it to establish a mutual mapping relationship between multi-level semantic identifiers and structured semantic profiles, thereby obtaining cross-level semantic understanding and association capabilities. Based on this, a second stage of fine-tuning is performed using interest point recommendation training samples, transferring the semantic mapping capabilities learned in the first stage of fine-tuning to the recommendation scenario. This allows the model to extract semantic identifiers from the user's historical check-in point access behavior sequences and accurately match them with the structured semantic profiles of candidate interest points. This scheme, through two-stage fine-tuning using a two-way alignment task and recommendation samples, transforms semantically meaningless numbers into a deep mapping between multi-level semantic identifiers and structured profiles, enabling the model to acquire transferable semantic reasoning capabilities. Therefore, in large-scale candidate sets, sparse behaviors, new regions, and long-tail scenarios, the model no longer relies on number co-occurrence but instead completes accurate recommendations based on semantic associations, significantly improving generalization ability and cold-start performance.

[0030] It's important to note that traditional point-of-interest (POI) recommendations typically use discrete IDs as unique identifiers for check-ins, meaning the model can only learn from co-occurrence statistics and lacks transferable semantic knowledge. To address this, this solution introduces multi-level semantic tagging, abstracting each POI into a hierarchical semantic code composed of region, category, subcategory, and core attributes, providing rich reasoning support for subsequent recommendation tasks. In terms of representation strategy, this solution does not directly use structured semantic profiles but instead uses hierarchical discrete labels. This is because semantic profiles contain dozens of continuous attributes, and direct input would cause a dramatic increase in model dimensionality. In contrast, multi-level semantic tagging compresses the representation space from high-dimensional continuous to low-dimensional discrete, reducing the input dimensionality by more than an order of magnitude while preserving the core semantic structure, thus balancing recommendation performance with the feasibility of engineering deployment.

[0031] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, based on the bidirectional alignment task and the interest point recommendation training samples, the large model is finely trained sequentially to obtain an interest point recommendation model with multi-level semantic labeling and semantic feature mutual recognition capabilities, including: 201. Based on the bidirectional alignment task, fine-tune the large model to obtain a large model with multi-level semantic labeling and semantic feature mutual recognition capabilities.

[0032] 202. Based on the interest point recommendation training samples, fine-tune the large model with multi-level semantic labeling and semantic feature mutual recognition capabilities to obtain the initial interest point recommendation model.

[0033] 203. Generate a list of multiple candidate multi-level semantic identifiers for an interest point recommendation training sample through the initial interest point recommendation model.

[0034] Based on a composite reward function that includes multi-dimensional reward items, the reward score of each candidate multi-level semantic identifier list is calculated, and the initial interest point recommendation model is trained by reinforcement learning based on the reward score to obtain the interest point recommendation model.

[0035] In this embodiment of the invention, a three-stage progressive fine-tuning training is performed on a large model based on a bidirectional alignment task. The first stage of training enables the model to recognize and map between multi-level semantic identifiers and structured semantic profiles. Based on this, a second stage of training is performed, which involves supervised fine-tuning of the fine-tuned large model using interest point recommendation training samples to obtain an initial interest point recommendation model. To further improve recommendation quality and generalization performance, multiple candidate multi-level semantic identifier lists are generated for each interest point recommendation training sample using the initial interest point recommendation model. A composite reward function containing multi-dimensional reward items is introduced, and the reward score for each candidate list is calculated. This score is used as the basis for optimization to train the initial model using reinforcement learning. This reinforcement learning training process strengthens the model's preference for high-quality recommendation results through a reward mechanism, enabling the final interest point recommendation model to continuously optimize parameters and improve the accuracy of interest point recommendations.

[0036] In one embodiment of the present invention, for further explanation and limitation, the process of fine-tuning the large model based on the bidirectional alignment task includes multiple training steps, in each training step: Extract bidirectional recognition task samples from each stage of the task to obtain a multi-stage task combination; The bidirectional recognition tasks of each stage in the multi-stage task combination are input into the large model to obtain the predicted output of each bidirectional recognition task, and the loss of each stage of the bidirectional recognition task is calculated based on the predicted output and the corresponding target output. Based on the loss of the bidirectional recognition task at each stage and the corresponding weights at each stage, the total loss of the current training step is calculated; the parameters of the large model are optimized based on the total loss to complete the training of the current training step. Each training step is executed iteratively until the preset training termination condition is met, resulting in a large model with multi-level semantic labeling and semantic feature mutual recognition capabilities.

[0037] In this embodiment of the invention, the training process based on the bidirectional alignment task includes multiple training steps. In each training step, a batch of bidirectional recognition task samples for training in the current step are sampled from the bidirectional recognition task set of each stage, forming a multi-stage task combination. This ensures that the model can simultaneously access bidirectional mapping knowledge at different granularities in each update step. The bidirectional recognition task samples of each stage are input into the large model for forward propagation. The model outputs a corresponding predicted output for each task and calculates the loss value of each sample based on the difference between the predicted output and the pre-labeled target output. Let any task in the m-th stage include an input instruction x and a target output sequence. Then the loss of any task in stage m is Represented as: ;in, Indicates the length of the target output sequence. Represents the traversal sequence Summation is performed at each position in the array; This indicates that given the input instruction x and the generated sequence Under the condition that the model predicts the first The position is correct. The conditional probability is then calculated. Based on this, the losses for each task are aggregated according to the stage dimension to obtain the stage loss. These losses are then weighted and combined with preset stage weights to calculate the total loss for the current training step. Total Loss The formula is expressed as: ;in, Let M represent the weight corresponding to the m-th stage, and M represent the set of stages corresponding to a multi-stage task. This represents the sample distribution at the m-th stage. The expected value of the model is calculated. This total loss comprehensively reflects the model's overall performance in the bidirectional recognition task across all granularity levels. Finally, with the goal of minimizing this total loss, an optimization algorithm is used to update the model parameters using gradients, completing the parameter adjustment for the current training step. This training step is iteratively executed until the model's performance on the validation set no longer improves or reaches the preset maximum number of training steps, at which point training terminates, ultimately yielding a large model capable of recognizing the bidirectional mapping between semantic labels and semantic profiles at different semantic granularities.

[0038] In one embodiment of the present invention, for further explanation and limitation, the reward score of each of the candidate multi-level semantic identifier lists is calculated based on a composite reward function that includes multi-dimensional reward items, including: For any candidate multi-level semantic tag list, calculate the scores for the format reward item, the bottom ranking reward item, the soft hit reward item, the tag level hit reward item, and the diversity reward item respectively, and then sum the scores of the reward items based on the weight coefficient of each reward item to obtain the reward score of the candidate multi-level semantic tag list. The initial interest point recommendation model is trained using reinforcement learning based on the reward score to obtain an interest point recommendation model, including: Calculate the difference between the reward score of each candidate multi-level semantic identifier list and the average of all reward scores; The strategy gradient of the model is updated based on the positive and negative values ​​of the difference to increase the generation probability of the candidate multi-level semantic identifier list corresponding to the positive difference and decrease the generation probability of the candidate multi-level semantic identifier list corresponding to the negative difference.

[0039] In this embodiment of the invention, after completing the supervised fine-tuning in the previous stage, the model already possesses basic semantic tag generation and recommendation capabilities. However, to further improve recommendation quality and ranking accuracy, a group relative strategy optimization method is introduced to enhance and fine-tune the model. Specifically, firstly, the user's historical visit check-in point sequence, predicted time, and target interest point semantic tags are constructed into conversational training samples. The model is required to directly generate a Top-K level semantic tag candidate list in descending order of probability. During training, for each input sample, the model generates multiple candidate multi-level semantic tag lists (hereinafter referred to as candidate lists). Then, a reward score is calculated for each candidate list. This reward score is provided by a composite reward function in the form of an external plugin to quantify the quality of the candidate list. Subsequently, the difference between the reward score of the candidate list and the average reward score of all candidate lists in the current batch is calculated, serving as the core indicator for measuring the relative quality of this generation.

[0040] Furthermore, based on the positive or negative direction of the aforementioned difference, the model's policy gradient is updated: when the difference is positive, it indicates that the quality of the candidate list is above average, and the model will increase the probability of it being generated again through gradient ascent; when the difference is negative, the probability of the candidate list being generated is reduced through gradient descent. This comparative optimization method enables the model to continuously tend to generate higher-quality Top-K recommendation lists during training, rather than simply fitting fixed labels. The entire enhancement and fine-tuning process can be executed based on the RLHF / GRPO training module of the MS-Swift framework, and the LoRA method is used for efficient parameter updates. The final LoRA adapter can be merged with the base model to form the final generative recommendation model for the next point of interest recommendation task, further enhancing the semantic alignment and ranking rationality of the recommendation results.

[0041] It should be noted that the composite reward function includes format reward, bottom-ranking reward, soft hit reward, identifier-level hit reward, and diversity reward. The composite reward function is expressed as: ; in, Indicates the format of the reward items. This indicates the reward for ranking last. This indicates a soft hit reward. This indicates that the reward item has been matched at the identifier level. Indicates diversity rewards, , , , , The corresponding weights for the above reward items are, in order, preferably 1, 1, 1, 0.3, and 0.2. Strengthening the large model based on the above composite reward function can comprehensively consider factors such as whether the target output label is hit, the hit ranking position, the matching degree of different levels of semantic labels, the number of labels generated in the candidate list, the legality of the output format, and the diversity of candidates. This guides the model not only to generate legal multi-level semantic labels but also to place the true target label at a higher position in the list.

[0042] The following example, using a candidate list containing 10 identifiers, illustrates each reward category: Format Reward Category The format validity of the identifiers used to evaluate the candidate multi-level semantic identifier list can be represented as: ;in, This indicates whether the output is a single line or a valid output. If it is a single line or a valid output, this value is 0.2; otherwise, it is 0. This represents the set of valid identifiers obtained through parsing. This indicates that the number of valid identifiers in the set is 10, meaning the number of valid identifiers matches the number of identifiers in the list; in this case, the value is 0.3. (This is for the bottom-ranking reward item.) Represented as: ;in, Indicates the target output identifier The sorting in the candidate list, if If it does not appear in the candidate list, then Soft hit bonus Represented as: ;in, This represents a multi-level semantic identifier set in the candidate list. If the target output identifier appears in the candidate list, 1 point is awarded; otherwise, 0 points are awarded. Identifier level hit reward item. Represented as: ;in, Indicates the identifier in the candidate list With target output identifier The number of semantic layers matched consecutively starting from level one. Diversity reward items. Represented as: ; This indicates the number of distinct identifiers in the candidate list, penalizing duplicate candidates and encouraging the generation of a complete candidate list.

[0043] 104. In response to the point of interest recommendation request, the point of interest recommendation model is used to predict the recommended points of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence, and generate the target recommended points of interest.

[0044] In this embodiment of the invention, during the model application phase, in response to a request to recommend points of interest to a target user, firstly, the trajectory of the target user's visits to several checkpoints within a historical period prior to the current moment is obtained, and multi-level semantic identifiers for each checkpoint in the access trajectory are constructed. These multi-level semantic identifiers are then arranged according to the order in which the checkpoints are accessed, resulting in the target user's historical checkpoint access sequence. Next, this historical checkpoint access sequence is input into the point of interest recommendation model to generate a multi-level semantic identifier list. This list includes multi-level semantic identifiers for multiple checkpoints arranged in descending order of recommendation degree. Finally, based on the order of the multi-level semantic identifiers in this list and the correspondence between the multi-level semantic identifiers and the checkpoints, one or more recommended points of interest are generated and displayed.

[0045] The point-of-interest (POI) recommendation request can take various triggering forms depending on the specific application scenario. For example, it could be triggered by a user actively opening the application's homepage, by a user triggering a recommendation request via a keystroke, by a user entering a query in the search box, or by a dynamic refresh triggered when the user moves or zooms the map interface. In any of these triggering forms, it is necessary to obtain the current context information of the requesting party, such as the user's geographical location, time period, historical access sequence, and filtering conditions. The user's current real-time geographical location is the last check-in point in the historical check-in point access sequence.

[0046] In one embodiment of the present invention, for further explanation and limitation, the interest point recommendation model is used to predict the recommended interest points based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence, generating target recommended interest points, including: Based on the multidimensional features of each check-in point in the historical check-in point access sequence, a multi-level semantic identifier is constructed for each check-in point to obtain the multi-level semantic identifier sequence of the target user. The order of each check-in point in the historical check-in point access sequence is consistent with the historical access trajectory of the target user. The interest point recommendation model is used to predict the recommended interest points of the multi-level semantic identifiers to obtain an identifier list containing multiple interest points to be recommended corresponding to the multi-level semantic identifiers. The multi-level semantic identifiers in the identifier list are arranged in descending order of the recommendation priority of the corresponding interest points. The validity of each multi-level semantic identifier in the identifier list is verified, and the multi-level semantic identifiers that pass the verification are decoded to obtain a candidate interest point list, wherein the candidate interest point list includes multiple candidate interest points arranged in descending order of recommendation degree. The top preset number of candidate interest points in the candidate interest point list are selected as the target recommended interest points.

[0047] In this embodiment of the invention, based on the multidimensional features of each check-in point in the target user's historical check-in point access sequence, corresponding multi-level semantic identifiers are constructed one by one, and organized into a multi-level semantic identifier sequence according to the user's historical access time order to characterize the user's behavioral trajectory pattern. Subsequently, the point of interest recommendation model uses this sequence as input to perform recommendation prediction, generating an identifier list containing multiple multi-level semantic identifiers of interest points to be recommended. The identifiers in this list are strictly arranged from high to low according to the recommendation priority of the corresponding interest points. Furthermore, the identifier validity is checked for each multi-level semantic identifier in the identifier list to ensure that its encoding structure is complete and can be correctly parsed, filtering out invalid or abnormal identifiers. Finally, the identifiers that pass the verification are decoded and restored to specific interest points, resulting in a candidate interest point list arranged from high to low recommendation degree. The top preset number of candidate interest points are then selected as the final target recommended interest points output to the user, completing the entire process from historical trajectory input to recommendation result output. The preset number can be determined by the default value or request description carried in the interest point recommendation request, such as 1, 2, 3, etc., and this embodiment of the invention does not impose a specific limitation.

[0048] To verify the effectiveness of generative point of interest (POI) recommendation based on the aforementioned technical implementation process, experiments were conducted on the public location social network datasets Foursquare-NYC, Foursquare-TKY, and Gowalla-CA. The experiments were conducted by dividing the training, validation, and test sets according to user check-in time, with a ratio of 80:10:10. Acc@k was used as the evaluation metric, representing the accuracy in the top k recommendation results, i.e., whether the true POIs appeared in the first k recommendations output by the model. To maintain the length of the documentation and highlight comparisons with more recent methods, this experimental example only lists recent deep spatiotemporal recommendation models, large-scale recommendation models, and semantic labeling-related models as baselines. Specific comparison results are shown in Table 1 below, where Foursquare-NYC corresponds to dataset 1, Foursquare-TKY to dataset 2, and Gowalla-CA to dataset 3.

[0049] Table 1: Comparison of Top-k accuracy of our method and baseline methods on different datasets

[0050] As shown in Table 1, this invention achieves superior results compared to the listed recent baseline methods in terms of Acc@5 and Acc@10 metrics across all three datasets. Specifically, compared to the strongest baseline in the table, this invention improves Acc@5 and Acc@10 by 16.72% and 6.21%, respectively, on the Foursquare-NYC dataset; by 8.01% and 1.56%, respectively, on the Foursquare-TKY dataset; and by 6.49% and 11.35%, respectively, on the Gowalla-CA dataset. This method effectively improves the overall hit rate and ranking quality of the interest point recommendation list, especially enhancing the generation of Top-k candidate lists.

[0051] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this embodiment of the invention provides a generative interest point recommendation system based on hierarchical semantic identifier alignment, such as... Figure 3 As shown, the system includes: The feature encoding module 31 is used to construct sample data based on the multidimensional feature data of the expected check-in points, perform multi-granularity index encoding on the multidimensional features of each sample check-in point in the sample data, and use the encoded indexes of different granularities as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers of each sample check-in point; and perform multi-granularity semantic feature extraction on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. The sample construction module 32 is used to construct a bidirectional alignment task based on the multi-level semantic identifier and the structured semantic profile, and to construct interest point recommendation training samples based on the multi-level semantic identifier of each check-in point in the sample user check-in point access sequence. The bidirectional alignment task is used to enable the large model to map the multi-level semantic identifier and the structured semantic profile to each other. Model training module 33 is used to fine-tune the training of the large model sequentially based on the bidirectional alignment task and the interest point recommendation training samples to obtain an interest point recommendation model with multi-level semantic identification and semantic feature mutual recognition capabilities. The recommendation generation module 34 is used to respond to the point of interest recommendation request, and to predict the target point of interest by using the point of interest recommendation model to predict the target point of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence.

[0052] Furthermore, the model training module 33 includes: The first training unit is used to fine-tune the large model based on the bidirectional alignment task to obtain a large model with multi-level semantic labeling and semantic feature mutual recognition capabilities. The second training unit is used to fine-tune the large model with multi-level semantic identification and semantic feature mutual recognition capabilities based on the interest point recommendation training samples, so as to obtain the initial interest point recommendation model. The generation unit is used to generate a list of multiple candidate multi-level semantic identifiers for an interest point recommendation training sample through the initial interest point recommendation model. The third training unit is used to calculate the reward score of each of the candidate multi-level semantic identifier lists based on a composite reward function containing multi-dimensional reward items, and to perform reinforcement learning training on the initial interest point recommendation model based on the reward score to obtain the interest point recommendation model.

[0053] Furthermore, the bidirectional alignment task includes a one-stage bidirectional alignment task, a two-stage bidirectional alignment task, and a three-stage bidirectional alignment task; the multi-level semantic identifier includes at least one level of coarse-grained semantic identifier and at least one level of fine-grained semantic identifier; the semantic profile hierarchy of the structured semantic profile corresponds to the hierarchical granularity of the multi-level semantic identifier; the sample construction module 32 includes: The first construction unit is used to construct a set of coarse-grained aligned sample pairs by combining the coarse-grained semantic identifier and the corresponding granular semantic profile, with each pair serving as a target output, and to construct a bidirectional recognition task of the coarse-grained aligned sample pairs as a first-stage bidirectional alignment task. The second construction unit is used to construct a set of globally aligned sample pairs that are mutually target outputs by the multi-level semantic identifiers and the structured semantic profiles, and to construct a bidirectional recognition task of the globally aligned sample pairs as a two-stage bidirectional alignment task. The third construction unit is used to construct a set of fine-grained aligned sample pairs by combining the fine-grained semantic identifiers and the corresponding granular semantic profiles, with each pair serving as a target output, and to construct a bidirectional recognition task for the fine-grained aligned sample pairs as a three-stage bidirectional alignment task.

[0054] Furthermore, in a specific application scenario, the first training unit is specifically used in each training step to: extract bidirectional recognition task samples for the current training step from each stage of the task to obtain a multi-stage task combination; input the bidirectional recognition tasks of each stage in the multi-stage task combination into the large model to obtain the predicted output of each bidirectional recognition task, and calculate the loss of each stage of the bidirectional recognition task based on the predicted output and the corresponding target output; calculate the total loss of the current training step based on the loss of each stage of the bidirectional recognition task and the weights corresponding to each stage; optimize the parameters of the large model based on the total loss to complete the training of the current training step; iteratively execute each training step until the preset training termination condition is met to obtain a large model with multi-level semantic labeling and semantic feature mutual recognition capabilities.

[0055] Furthermore, the composite reward function includes a format reward item, a bottom-ranking reward item, a soft hit reward item, a hierarchical hit reward item, and a diversity reward item; In a specific application scenario, the third training unit is specifically used to calculate the scores of the format reward item, the bottom ranking reward item, the soft hit reward item, the identifier level hit reward item, and the diversity reward item for any candidate multi-level semantic identifier list, and to perform a weighted summation of the scores of the reward items based on the weight coefficient of each reward item to obtain the reward score of the candidate multi-level semantic identifier list. The initial interest point recommendation model is trained using reinforcement learning based on the reward score to obtain an interest point recommendation model, including: Calculate the difference between the reward score of each candidate multi-level semantic identifier list and the average of all reward scores; The strategy gradient of the model is updated based on the positive and negative values ​​of the difference to increase the generation probability of the candidate multi-level semantic identifier list corresponding to the positive difference and decrease the generation probability of the candidate multi-level semantic identifier list corresponding to the negative difference.

[0056] Furthermore, the recommendation generation module 34 includes: The fourth construction unit is used to construct a multi-level semantic identifier for each check-in point based on the multi-dimensional features of each check-in point in the historical check-in point access sequence, thereby obtaining a multi-level semantic identifier sequence for the target user, wherein the order of each check-in point in the historical check-in point access sequence is consistent with the historical access trajectory of the target user. The prediction unit is used to predict the recommended interest points by the multi-level semantic identifiers through the interest point recommendation model, and obtain an identifier list containing multiple multi-level semantic identifiers corresponding to interest points to be recommended, wherein the multi-level semantic identifiers in the identifier list are arranged in descending order of the recommendation priority of the corresponding interest points. The decoding unit is used to perform identification validity verification on each multi-level semantic identifier in the identifier list, and decode the multi-level semantic identifiers that pass the verification to obtain a candidate interest point list, wherein the candidate interest point list includes multiple candidate interest points arranged in descending order of recommendation degree. The generation unit is used to select the top preset number of candidate interest points from the candidate interest point list as target recommended interest points.

[0057] Further, in a specific application scenario, the fourth construction unit is specifically used to construct multi-level semantic identifiers for each check-in point based on the multi-dimensional features of each check-in point in the historical check-in point access sequence, thereby obtaining a multi-level semantic identifier sequence for the target user. The order of each check-in point in the historical check-in point access sequence is consistent with the historical access trajectory of the target user. The interest point recommendation model is used to predict recommended interest points based on the multi-level semantic identifiers, resulting in an identifier list containing multiple multi-level semantic identifiers corresponding to interest points to be recommended. The multi-level semantic identifiers in the identifier list are arranged in descending order of recommendation priority for their corresponding interest points. The identifier validity of each multi-level semantic identifier in the identifier list is verified, and the verified multi-level semantic identifiers are decoded to obtain a candidate interest point list. The candidate interest point list includes multiple candidate interest points arranged in descending order of recommendation degree. The first preset number of candidate interest points in the candidate interest point list are selected as target recommended interest points.

[0058] According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, which is capable of executing the generative point of interest recommendation method based on hierarchical semantic tag alignment in any of the above method embodiments.

[0059] Figure 4 The diagram shows a structural schematic of a terminal according to an embodiment of the present invention. The specific implementation of the present invention is not limited to the specific implementation of the terminal.

[0060] like Figure 4 As shown, the terminal may include: a processor 402, a communication interface 404, a memory 406, and a communication bus 408.

[0061] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0062] Communication interface 404 is used for network communication with other devices such as clients or other servers.

[0063] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above embodiment of the generative interest point recommendation method based on hierarchical semantic tag alignment.

[0064] Specifically, program 410 may include program code that includes computer operation instructions.

[0065] Processor 402 may be a central processing unit (CPU), a specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The terminal may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0066] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0067] Specifically, program 410 can be used to cause processor 402 to perform the following operations: Sample data is constructed based on the multidimensional feature data of the expected check-in points. Multi-granularity index encoding is performed on the multidimensional features of each sample check-in point in the sample data, and the indexes of different granularities obtained by encoding are used as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers of each sample check-in point. Multi-granularity semantic feature extraction is performed on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. Based on the multi-level semantic identifiers and the structured semantic profile, a bidirectional alignment task is constructed, and based on the multi-level semantic identifiers of each check-in point in the sample user check-in point access sequence, interest point recommendation training samples are constructed. The bidirectional alignment task is used to enable the large model to map the multi-level semantic identifiers and the structured semantic profile to each other. Based on the bidirectional alignment task and the interest point recommendation training samples, the large model is finely trained in sequence to obtain an interest point recommendation model with multi-level semantic identification and semantic feature mutual recognition capabilities. In response to a point of interest recommendation request, the point of interest recommendation model is used to predict the target point of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence, and generate the target recommended point of interest.

[0068] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A generative point of interest recommendation method based on hierarchical semantic tag alignment, characterized in that, include: Sample data is constructed based on the multidimensional feature data of the expected check-in points. Multi-granularity index encoding is performed on the multidimensional features of each sample check-in point in the sample data, and the indexes of different granularities obtained by encoding are used as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers of each sample check-in point. Multi-granularity semantic feature extraction is performed on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. Based on the multi-level semantic identifiers and the structured semantic profile, a bidirectional alignment task is constructed, and based on the multi-level semantic identifiers of each check-in point in the sample user check-in point access sequence, interest point recommendation training samples are constructed. The bidirectional alignment task is used to enable the large model to map the multi-level semantic identifiers and the structured semantic profile to each other. Based on the bidirectional alignment task and the interest point recommendation training samples, the large model is finely trained in sequence to obtain an interest point recommendation model with multi-level semantic identification and semantic feature mutual recognition capabilities. In response to a point of interest recommendation request, the point of interest recommendation model is used to predict the target point of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence, and generate the target recommended point of interest.

2. The generative interest point recommendation method based on hierarchical semantic identifier alignment according to claim 1, characterized in that, Based on the bidirectional alignment task and the interest point recommendation training samples, the large model is finely trained sequentially to obtain an interest point recommendation model with multi-level semantic labeling and semantic feature mutual recognition capabilities, including: Based on the bidirectional alignment task, the large model is fine-tuned and trained to obtain a large model with multi-level semantic labeling and semantic feature mutual recognition capabilities. Based on the interest point recommendation training samples, the large model with multi-level semantic identification and semantic feature mutual recognition capabilities is fine-tuned to obtain the initial interest point recommendation model. The initial interest point recommendation model generates a list of multiple candidate multi-level semantic identifiers for an interest point recommendation training sample. Based on a composite reward function that includes multi-dimensional reward items, the reward score of each candidate multi-level semantic identifier list is calculated, and the initial interest point recommendation model is trained by reinforcement learning based on the reward score to obtain the interest point recommendation model.

3. The generative interest point recommendation method based on hierarchical semantic tag alignment according to claim 2, characterized in that, The bidirectional alignment task includes a one-stage bidirectional alignment task, a two-stage bidirectional alignment task, and a three-stage bidirectional alignment task. The multi-level semantic identifier includes at least one level of coarse-grained semantic identifier and at least one level of fine-grained semantic identifier. The semantic profile level of the structured semantic profile corresponds to the hierarchical granularity of the multi-level semantic identifier. The construction process of the bidirectional alignment task includes: The coarse-grained semantic identifier and the corresponding granular semantic profile are constructed into a set of coarse-grained aligned sample pairs that are mutually target outputs, and the bidirectional recognition task of the coarse-grained aligned sample pairs is constructed as a first-stage bidirectional alignment task. The multi-level semantic identifiers and the structured semantic profiles are constructed into a set of globally aligned sample pairs that are mutually target outputs, and the bidirectional recognition task of the globally aligned sample pairs is constructed as a two-stage bidirectional alignment task. The fine-grained semantic identifiers and corresponding granular semantic profiles are constructed into a set of fine-grained aligned sample pairs that are mutually target outputs, and the bidirectional recognition task of the fine-grained aligned sample pairs is constructed as a three-stage bidirectional alignment task.

4. The generative interest point recommendation method based on hierarchical semantic tag alignment according to claim 3, characterized in that, The process of fine-tuning the large model based on the bidirectional alignment task includes multiple training steps, in each training step: Extract bidirectional recognition task samples from each stage of the task to obtain a multi-stage task combination; The bidirectional recognition tasks of each stage in the multi-stage task combination are input into the large model to obtain the predicted output of each bidirectional recognition task, and the loss of each stage of the bidirectional recognition task is calculated based on the predicted output and the corresponding target output. Based on the loss of the bidirectional recognition task at each stage and the corresponding weights at each stage, the total loss of the current training step is calculated; the parameters of the large model are optimized based on the total loss to complete the training of the current training step. Each training step is executed iteratively until the preset training termination condition is met, resulting in a large model with multi-level semantic labeling and semantic feature mutual recognition capabilities.

5. The generative point of interest recommendation method based on hierarchical semantic identifier alignment according to claim 2, characterized in that, The composite reward function includes format reward items, bottom-ranking reward items, soft hit reward items, identifier-level hit reward items, and diversity reward items; Based on a composite reward function that includes multi-dimensional reward items, the reward score for each of the candidate multi-level semantic identifier lists is calculated, including: For any candidate multi-level semantic tag list, calculate the scores for the format reward item, the bottom ranking reward item, the soft hit reward item, the tag level hit reward item, and the diversity reward item respectively, and then sum the scores of the reward items based on the weight coefficient of each reward item to obtain the reward score of the candidate multi-level semantic tag list. The initial interest point recommendation model is trained using reinforcement learning based on the reward score to obtain an interest point recommendation model, including: Calculate the difference between the reward score of each candidate multi-level semantic identifier list and the average of all reward scores; The strategy gradient of the model is updated based on the positive and negative values ​​of the difference to increase the generation probability of the candidate multi-level semantic identifier list corresponding to the positive difference and decrease the generation probability of the candidate multi-level semantic identifier list corresponding to the negative difference.

6. The generative interest point recommendation method based on hierarchical semantic tag alignment according to claim 1, characterized in that, The interest point recommendation model predicts target interest points by performing multi-level semantic identifier sequences on the target user's historical check-in point access sequence, including: Based on the multidimensional features of each check-in point in the historical check-in point access sequence, a multi-level semantic identifier is constructed for each check-in point to obtain the multi-level semantic identifier sequence of the target user. The order of each check-in point in the historical check-in point access sequence is consistent with the historical access trajectory of the target user. The interest point recommendation model is used to predict the recommended interest points of the multi-level semantic identifiers to obtain an identifier list containing multiple interest points to be recommended corresponding to the multi-level semantic identifiers. The multi-level semantic identifiers in the identifier list are arranged in descending order of the recommendation priority of the corresponding interest points. The validity of each multi-level semantic identifier in the identifier list is verified, and the multi-level semantic identifiers that pass the verification are decoded to obtain a candidate interest point list, wherein the candidate interest point list includes multiple candidate interest points arranged in descending order of recommendation degree. The top preset number of candidate interest points in the candidate interest point list are selected as the target recommended interest points.

7. The generative interest point recommendation method based on hierarchical semantic identifier alignment according to claim 6, characterized in that, The process of constructing multi-level semantic identifiers for any check-in point includes: Category features are extracted based on the category description text of the check-in point, spatial features are extracted based on the geographical location information and geographical affiliation of the check-in point, and time features are extracted based on the access statistics of the check-in point. The category features, spatial features and time features are then fused to obtain multi-dimensional fused features. The multidimensional fusion features are residual quantized and encoded by a residual quantization variational autoencoder to obtain a multi-level codeword index with granularity from coarse to fine. Each level of the codeword index in the multi-level codeword index is used as a hierarchical semantic identifier of the same granularity to obtain the multi-level semantic identifier of the check-in point. After generating multi-level semantic identifiers for all check-in points, the method further includes: Extract conflicting check-in points that have the same multi-level semantic identifiers, and add conflict suffixes to the multi-level semantic identifiers of the conflicting check-in points to distinguish each conflicting check-in point based on the conflict suffixes.

8. A generative point of interest recommendation system based on hierarchical semantic tag alignment, characterized in that, include: The feature encoding module is used to construct sample data based on the multidimensional feature data of the expected check-in points, perform multi-granularity index encoding on the multidimensional features of each sample check-in point in the sample data, and use the encoded indices of different granularities as semantic identifiers of the corresponding levels of the sample check-in points to obtain multi-level semantic identifiers for each sample check-in point; and perform multi-granularity semantic feature extraction on the multidimensional features of each sample check-in point to obtain a structured semantic profile of each sample check-in point. The sample construction module is used to construct a bidirectional alignment task based on the multi-level semantic identifiers and the structured semantic profile, and to construct interest point recommendation training samples based on the multi-level semantic identifiers of each check-in point in the sample user check-in point access sequence. The bidirectional alignment task is used to enable the large model to map the multi-level semantic identifiers and the structured semantic profile to each other. The model training module is used to fine-tune the large model sequentially based on the bidirectional alignment task and the interest point recommendation training samples to obtain an interest point recommendation model with multi-level semantic labeling and semantic feature mutual recognition capabilities. The recommendation generation module is used to respond to the point of interest recommendation request by predicting the target point of interest based on the multi-level semantic identifier sequence of the target user's historical check-in point access sequence through the point of interest recommendation model, and generating the target recommended point of interest.

9. A storage medium, characterized in that, The storage medium stores at least one executable instruction that causes the processor to perform the operation corresponding to the generative point of interest recommendation method based on hierarchical semantic identifier alignment as described in any one of claims 1-7.

10. A terminal, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the generative point of interest recommendation method based on hierarchical semantic identifier alignment as described in any one of claims 1-7.

Images