Training and recommending method of self-supervised learning model for tourism recommendation

Abstract

The invention provides a training and recommendation method for a self-supervised learning model for tourism recommendation, and relates to the technical field of data processing, and the method comprises the steps: constructing a heterogeneous information graph based on tourist text tourism data, carrying out the edge random mask, inputting a target encoder, extracting features, and obtaining node embedding; performing calculation and result splicing on node embedding of every two nodes through a cross-correlation decoder to obtain corresponding edge embedding; decoding the edge embedding through a target decoder to obtain a reconstructed heterogeneous information graph; inputting the heterogeneous information graph into a target generator to extract node global embedding; aggregating node global embedding and context information thereof to obtain node context embedding; and performing optimization training on the model based on a target loss function consisting of a reconstructed heterogeneous information graph, a heterogeneous information graph, node embedding, node global embedding and node context embedding to obtain a target model which is qualified in training. The objective of the invention is to improve prediction accuracy of a text travel recommendation model.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a training and recommendation method for a self-supervised learning model for tourism recommendation. Background Technology

[0002] In the current context of smart tourism technology, personalized recommendations and services primarily rely on traditional collaborative filtering, content-based recommendation algorithms, and rudimentary graph neural network applications. These technologies construct a relationship graph of entities such as tourists, attractions, hotels, and goods, and make recommendations by analyzing the connections within the graph. The main recommendation methods are based on graph autoencoders (GAE) and graph contrastive learning (GCL). These approaches represent the mainstream methods in the current field of graph data mining, but they have revealed significant limitations in the complex scenario of smart tourism.

[0003] Generative learning-based approaches (such as GAE) learn node representations by reconstructing existing connections in the graph (e.g., attractions visited by tourists). These methods are adept at capturing "local neighborhood information," accurately recommending restaurants near or functionally similar to a tourist's current location. However, their problem lies in overemphasizing and fitting these direct, short-range connections at the expense of insights into global structural information. This makes it difficult for the system to discover tourists' potential long-term interests or thematic preferences, and to recommend content that is geographically remote but culturally relevant. This limitation stems from the model mechanism itself—the shallow design of the GNN encoder and the direct reconstruction of the adjacency matrix limit its vision to a few hops of neighbors, resulting in an "oversmoothing" problem and difficulty in learning long-range dependencies.

[0004] On the other hand, contrastive learning-based approaches (such as Global Logical Learning) learn by maximizing the mutual information between the local and global representations of a graph, aiming to capture the graph's "global structural information." In tourism scenarios, these methods excel at making broad category recommendations based on the overall profile of tourists (e.g., "history enthusiasts"). However, their drawback lies in their weak ability to perceive local details, failing to accurately understand subtle changes in tourists' interests within specific contexts. For example, a "history enthusiast" visiting the Beiting Ancient City might be focused on the military defense system of a particular dynasty. Contrastive learning models, due to their learning paradigm's emphasis on invariance between different graph "views," might fail to capture this specific, dynamic point of interest, resulting in recommendations that are overly generalized but lack precision. The root of this problem is that contrastive learning, in order to learn global invariance, often sacrifices sensitivity to changes in local topological structure. Therefore, current recommendation methods cannot simultaneously consider the crucial local context and global preferences in the tourism experience, leading to fragmented and superficial service experiences. Summary of the Invention

[0005] In view of this, this application provides a training and recommendation method for a self-supervised learning model for tourism recommendation. It aims to solve or partially solve the problems existing in the background technology.

[0006] The first aspect of this application provides a method for training a self-supervised learning model for tourism recommendation, the method comprising: Based on the collected tourist and cultural tourism data within the target area, a heterogeneous information map is constructed. The cultural tourism data includes data on cultural tourism objects and interaction data between cultural tourism objects. The cultural tourism objects include at least tourists, scenic spots, historical relics, cultural and creative products, catering, and accommodation. Randomly mask the edges of the heterogeneous information graph to generate a masked heterogeneous information graph; The mask heterogeneous information map is input into the target encoder of the model for processing. Feature extraction is performed through two feature extraction layers to obtain the node embedding of each node extracted in different feature extraction layers. The Hadamard product is calculated element-wise for the node embeddings of every two nodes by using the cross-correlation decoder of the model, and the results are concatenated to obtain the edge embeddings of the corresponding two nodes. The edge embedding is decoded by the target decoder of the model to reconstruct the masked edges in the heterogeneous information graph and obtain the corresponding reconstructed heterogeneous information graph. The heterogeneous information graph is input into the target generator of the model for processing, and the global embedding of each node is extracted. The global embedding of the node and its context information are aggregated to generate the corresponding node context embedding. Based on the target loss function, the model is optimized and trained to obtain a qualified target model for recommending cultural and tourism products to tourists. The target loss function includes: a local loss function that evaluates the deviation between the reconstructed heterogeneous information graph and the heterogeneous information graph; a global loss function that evaluates the deviation between the target node embedding and the global node embedding; and an additional loss function that evaluates the deviation between the target node embedding and the node context embedding. The target node embedding is the node embedding extracted by the second feature extraction layer in the target encoder.

[0007] The second aspect of this application provides a tourism recommendation method based on a self-supervised learning model for tourism recommendation, applied to a target model trained using the training method for a self-supervised learning model for tourism recommendation described in the first aspect of this application. The method includes: Based on the acquired cultural and tourism data of the target tourists in the target area, a heterogeneous information map corresponding to the target tourists is constructed. The heterogeneous information graph corresponding to the target tourist is input into the target model for edge link prediction processing between nodes to obtain the corresponding target heterogeneous information graph. In practical applications, the structure of the target model includes a target encoder, a cross-correlation decoder, and a target decoder. Based on the target heterogeneous information map obtained from the prediction, travel recommendations are made for the target tourists.

[0008] The training method for a self-supervised learning model for tourism recommendation provided in this application has the following advantages: This application provides a training method for a self-supervised learning model for tourism recommendation. First, based on collected tourist and cultural tourism data within a target area, a heterogeneous information graph is constructed. This cultural tourism data includes data on cultural tourism objects and interaction data between these objects. Cultural tourism objects include at least tourists, scenic spots, historical relics, cultural and creative products, catering, and accommodation. The heterogeneous information graph is then subjected to random edge masking to generate a masked heterogeneous information graph. This masked heterogeneous information graph is input into the model's target encoder for processing. Feature extraction is performed through two feature extraction layers to obtain the node embeddings extracted from each node at different feature extraction layers. The model's cross-correlation decoder performs element-wise Hadamard product calculations on the node embeddings of every two nodes, and the results are concatenated to obtain the edge embeddings of the corresponding two nodes. Finally, the model's target decoder processes the edge embeddings... The process involves decoding the input to reconstruct the masked edges in the heterogeneous information graph, obtaining the corresponding reconstructed heterogeneous information graph. This heterogeneous information graph is then input into the model's target generator for processing, extracting the global embedding of each node. The global embeddings and their contextual information are aggregated to generate the corresponding contextual embeddings. Based on the target loss function, the model is optimized and trained to obtain a qualified target model for recommending cultural and tourism products to tourists. The target loss function includes: a local loss function evaluating the deviation between the reconstructed heterogeneous information graph and the heterogeneous information graph itself; a global loss function evaluating the deviation between the target node embedding and the global node embedding; and an additional loss function evaluating the deviation between the target node embedding and the contextual node embedding. The target node embedding is the node embedding extracted by the second feature extraction layer in the target encoder.

[0009] The core of the training method for the self-supervised learning model for tourism recommendation provided in this application lies in multi-scale collaborative decoding and reconstruction, specifically including three aspects: Cross-correlation decoding at the local neighborhood scale: This task aims to accurately reconstruct local edge relationships in a masked heterogeneous information graph. This application designs a cross-correlation decoder that receives the node embeddings of any two nodes (e.g., a tourist and a scenic spot) obtained by the target encoder. It calculates the Hadamard product between the node embeddings extracted by each feature extraction layer and concatenates them to form a rich set of interaction features (i.e., corresponding edge embeddings) corresponding to the two nodes. This interaction feature can capture the relationships between nodes at different distance scales. This cross-layer interaction design allows the model to predict direct connections not only by relying on one-hop neighbors but also by referencing information from more distant neighbors, thus effectively mitigating overfitting to the direct neighborhood while maintaining local perception accuracy.

[0010] Prediction of latent global representations at the global structural scale: To compensate for the shortcomings of the target decoder (the GNN encoder chosen in this application) in capturing long-range dependencies, this application introduces a parallel global information learning task. This application uses an external model (i.e., the target generator, Node2Vec chosen in this application) adept at capturing global structural information to encode an undisturbed heterogeneous information graph, generating an ideal global embedding for each node. Then, this application forces the node embeddings generated by the target encoder on the masked heterogeneous information graph, after transformation by a projection head (MLP Multilayer Perceptron structure), to predict and match this ideal global embedding as closely as possible. This process is akin to hiring a "teacher" (Node2Vec, the target generator) with a global perspective for a "student" (the GNN encoder, which only focuses on local information), forcing the model to learn global topological structures and long-range dependencies beyond the local neighborhood.

[0011] Additional Reconstruction for Context Stability: To address the randomness introduced by edge masks in heterogeneous information graphs and enhance the robustness of the node representations learned by the model, this application introduces a third task. This task requires the model to reconstruct the stable neighborhood context features (e.g., an aggregation of features from all its neighbors) of a node in the undisturbed heterogeneous information graph, based on the perturbed neighborhood of the node. This task forces the model to ignore noise introduced by random perturbations and learn the invariant features of the node in the real neighborhood, thereby obtaining a more stable and generalizable node representation.

[0012] Finally, this application weights and sums the loss functions of the three tasks (local link reconstruction loss, global representation prediction loss, and context reconstruction loss) to form a unified joint learning objective function for end-to-end optimization. After training, the resulting GNN encoder can generate stable and information-rich embedding vectors with both local context awareness and global perspective for any entity in a heterogeneous information graph related to cultural tourism. These vectors can be directly used for downstream tasks, such as calculating the matching degree between tourists and attractions for recommendations, discovering potentially related attractions to form new travel routes, or generating personalized interactive content for tourists' digital avatars in virtual reality scenarios. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 The flowchart illustrates a training method for a self-supervised learning model for tourism recommendation, as shown in one embodiment of this application. Figure 2 This is a structural diagram of a self-supervised learning model in a training method for a self-supervised learning model for tourism recommendation, as shown in one embodiment of this application. Figure 3 This is a flowchart illustrating a self-supervised learning model for tourism recommendation, as shown in one embodiment of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] refer to Figure 1 , Figure 1 This is a flowchart illustrating a training method for a self-supervised learning model for tourism recommendation, as shown in one embodiment of this application. Figure 1 As shown, the method includes: Step S01: Based on the collected tourist cultural and tourism data in the target area, construct a heterogeneous information map, wherein the cultural and tourism data includes cultural and tourism object data and interaction data between cultural and tourism objects, and the cultural and tourism objects include at least tourists, scenic spots, historical relics, cultural and creative products, catering and accommodation.

[0017] In this embodiment, the target area to be applied is first defined, such as by city or province, i.e., a city or province is designated as the target area. Then, tourist cultural and tourism data generated in the target area is collected. This data includes two types: one is cultural and tourism object data, i.e., cultural and tourism objects existing within the target area, which at least include tourists, scenic spots, historical relics, cultural and creative products, catering (i.e., restaurants, etc.), and accommodation (bed and breakfasts, hotels, etc.); the other is interaction data between cultural and tourism objects, such as tourist 1 visiting scenic spots a1 and a2 and purchasing cultural and creative product b1, and tourist 2 visiting scenic spots a1 and a3 and staying at hotel c1. Based on the collected tourist cultural and tourism data generated in the target area, a heterogeneous information graph is constructed. This heterogeneous information graph is an undirected, unweighted graph, where nodes are composed of cultural and tourism objects, and the edges between nodes are composed of the interaction behaviors generated between these objects.

[0018] In this embodiment, the heterogeneous information graph is defined as follows: Represents an undirected, unweighted graph, where The number of representatives is The set of nodes, Represents the edge set. It is the adjacency matrix of a heterogeneous information graph, containing elements Indicate node and nodes Edges exist; the input node feature matrix is ,in For nodes of 3D eigenvectors.

[0019] Step S02: Perform random edge masking on the heterogeneous information graph to generate a masked heterogeneous information graph.

[0020] In this embodiment, for the constructed heterogeneous information graph, traditional GAE (Graph Autoencoder) methods often use the entire heterogeneous information graph as input to learn and reconstruct the entire graph network structure of the input graph (i.e., the heterogeneous information graph). However, this design of reconstructing the entire graph network structure often sacrifices global structural information, overemphasizing local neighborhood information that is not always beneficial for link prediction tasks. To solve this problem, this application proposes to perturb the input heterogeneous information graph through a random masking strategy, and only use the undisturbed part as the input of the target encoder of the model. Specifically, based on the heterogeneous information graph obtained in step S01, a portion of the edges are selected from the observable edges in the heterogeneous information graph obtained in step S01 at a certain masking ratio (e.g., 10%, 20%, etc.) for masking processing, resulting in a masked heterogeneous information graph after mask perturbation. Figure 2 As shown, Figure 2 The Random Mask in this context refers to the masking operation performed on heterogeneous information graphs.

[0021] Step S03: Input the mask heterogeneous information map into the target encoder of the model for processing. Perform feature extraction through two feature extraction layers to obtain the node embedding of each node extracted in different feature extraction layers.

[0022] In this embodiment, to effectively map nodes in the masked heterogeneous information graph to hidden representations, a Graph Neural Network (GNN) is typically used as the encoder backbone. Its goal is to update the representation of a node using the representations of the node itself and its neighboring nodes. This application uses GraphSAGE as the network layer of the encoder. This application also allows the use of other different encoder architectures. Specifically, the target encoder of the model in this application is preferably a GNN encoder, which includes two feature extraction layers. These two feature extraction layers extract node embeddings for each node in the masked heterogeneous information graph, obtaining the node embeddings extracted for each node from different feature extraction layers. The specific representation is as follows:

[0023] in, This represents the node embedding of dimension d extracted from node v at the k-th layer in the masked heterogeneous information graph; Let v be the set of all directly connected neighboring nodes of node v in the masked heterogeneous information graph, and let the initial embedding of node v be set to 0. The AGG function is used to aggregate features from neighboring nodes; the COM function is used to combine the aggregated neighbor information from the previous layer with the node's own node embedding; a The layer-1 GNN encoder generates the K-layer node embeddings of node v. Due to the neighbor aggregation mechanism for information transmission, node v has node embedding at layer k. It can capture the neighborhood structure within a k-hop range; by inputting the masked heterogeneous information map into the target encoder, it solves the problem of overfitting the direct neighborhood information caused by the traditional GAE input and reconstruction of the entire heterogeneous information map.

[0024] In this embodiment, as Figure 2 As shown, Figure 2 This is a structural diagram of a self-supervised learning model in a training method for a self-supervised learning model for tourism recommendation, as shown in one embodiment of this application. Figure 2 In this context, GNN Encoder is the target encoder, Layer 1 is the first feature extraction layer of the target encoder, and Layer 2 is the second feature extraction layer of the target encoder.

[0025] Step S04: Calculate the element-wise Hadamard product of the node embeddings of every two nodes using the cross-correlation decoder of the model, and concatenate the calculation results to obtain the edge embeddings of the corresponding two nodes.

[0026] In this embodiment, to address the noise issue in the node embeddings encoded by the target encoder while ensuring accurate perception of local structures, this application designs a cross-correlation decoder to capture the cross-correlation similarity between two endpoints at different granularities, thereby improving the model's reconstruction capability. The cross-correlation decoder receives the node embeddings of any two nodes generated by the target encoder (e.g., k sets of node embeddings for node v and k sets of node embeddings for node u), denoted as... The Hadamard product is then performed element-wise on the K sets of node embeddings of the two nodes extracted by the feature extraction layer to obtain the final edge embeddings of the two nodes. For example, when the feature extraction layer in the target encoder is 2, and K is 2, then the element-wise Hadamard product of the two sets of node embeddings of node v and node u represents: the Hadamard product of the first-layer node embedding of node v with the first-layer node embedding of node v; the Hadamard product of the first-layer node embedding of node v with the second-layer node embedding of node v; the Hadamard product of the second-layer node embedding of node v with the first-layer node embedding of node v; and the Hadamard product of the second-layer node embedding of node v with the second-layer node embedding of node v. The specific expression is as follows:

[0027]

[0028] in, This represents the element-wise product of the k-th layer node embedding of node v and the j-th layer node embedding of node u at corresponding positions, which can be viewed as a model of the relationships between nodes. The final edge embedding obtained for every two nodes... Most elements in the edge embedding will become zero or close to zero; only elements are retained when two nodes are highly correlated. This design allows the prediction model to accurately perceive local structure. Simultaneously, thanks to the cross-layer cross-correlation design, the prediction model can flexibly learn more global structural information from different nearby neighbors when the most direct and local links are broken, effectively mitigating the prediction model's over-reliance on local information. Finally, edge embedding... The edge embedding will be decoded by the target decoder (in this application, an MLP-structured decoder is selected) to obtain the reconstruction score y(v,u)∈R.

[0029] In this embodiment, as Figure 2 As shown, Figure 2 China and The corresponding part is the process of calculating the Hadamard product of the node embeddings by the cross-correlation decoder, as shown in the figure. and The MLPs on the right are the target decoders.

[0030] Step S05: Decode the edge embeddings using the target decoder of the model to reconstruct the masked edges in the heterogeneous information graph and obtain the corresponding reconstructed heterogeneous information graph.

[0031] In this embodiment, after obtaining the edge embeddings between every two nodes through the cross-correlation decoder of the model in step S04, the target decoder of the model (the decoder selected in this application uses an MLP structure) decodes each edge embedding to obtain a reconstruction score for each edge embedding. Based on the reconstruction score of each edge embedding, the masked heterogeneous information graph is reconstructed, that is, the edges that were masked in the original heterogeneous information graph are reconstructed as the target, thereby obtaining the corresponding reconstructed heterogeneous information graph. Specifically, the goal of this application is to learn an encoder network. This encoder network is able to connect each node The feature vector is mapped as dimensional embedding vector There is also a decoding network. Based on the embedding vector encoded by the encoder network, the entire graph network structure is reconstructed. The core objective of the reconstruction is to predict the potential edges in the heterogeneous information graph.

[0032] Step S06: Input the heterogeneous information graph into the target generator of the model for processing, and extract the global embedding of each node.

[0033] In this embodiment, current Graph Masking Autoencoder (GMAE) methods typically include an encoder (such as the target encoder in this application) and a decoder (such as the target decoder in this application). The encoder is set as a Graph Neural Network (GNN) model, which mainly calculates node embeddings from a perturbed graph structure through message passing of neighborhood information, where the node features or edges of the perturbed graph structure are masked. The decoder is often set as a Graph Neural Network (GNN) model or a Multilayer Perceptron (MLP) model, whose task is to reconstruct the masked node features or edges using the node embeddings output by the encoder. Although the Graph Masking Autoencoder method has achieved high accuracy in various graph link prediction tasks, this application finds that it still has certain limitations. The main limitation is that when the Graph Masking Autoencoder method uses a GNN model as the encoder or decoder, it aggregates the neighbor information of the current node when calculating the embedding for each node. The design of a multi-layered GNN means that the reconstructed features of adjacent nodes tend to be similar, resulting in the "oversmoothing" problem that is common in GNN models. This also determines that the encoder and decoder in the GMAE method use a GNN model with fewer layers. However, a k-layer GNN model can only consider the k-hop neighbors of each node. This inherently shallow design leads to the encoder and decoder focusing only on local information and losing global information. Therefore, considering the inherent shallow design of the target encoder, which causes existing prediction models to lose global information during reconstruction learning, this application argues that the single learning method of only reconstructing masking edges still limits the prediction model to learning only local graph information. To address this, without changing the encoder architecture (i.e., the target encoder in this application), this application proposes to construct an information-rich additional prediction target in a higher-order latent representation space to extend the learning content to global information. This target aims to capture the global structural information of the graph to the greatest extent.

[0034] Based on this concept, this application finds that Node2Vec can effectively capture global structural information beyond local neighborhoods. Therefore, this application preferably uses Node2Vec as the target generator to directly generate potential predicted targets containing long-range dependencies and global structural information from unmasked heterogeneous information graphs. This application formally represents the Node2Vec encoder as a function... Meanwhile, this application defines a projection head. This corresponds to the decoder at the global scale, whose function is to map the node embeddings encoded by the target encoder into a representation space for prediction, where... These are learnable weights.

[0035] During pre-training, the undisturbed heterogeneous information graph is first fed into the target generator to generate a result that includes the global embedding of each node in the heterogeneous information graph. Then, the encoding result of the masked heterogeneous information graph. (The encoded result, which is the node embedding extracted by the second feature extraction layer in the target encoder,) is projected onto the representation space to obtain the node embedding used for latent prediction. The specific expression is:

[0036]

[0037] In other words, the undisturbed heterogeneous information graph is fed into the target generator for processing, generating the global embedding of each node in the heterogeneous information graph. The node embeddings extracted by the second feature extraction layer of the target encoder are then projected onto the representation space through a defined projection head to obtain the corresponding node embeddings used for potential prediction. The obtained global node embeddings and the projected node embeddings are then used to calculate the subsequent global loss function.

[0038] In this embodiment, as Figure 2 As shown, Figure 2 In this context, Node2Vec is the target generator. This represents the global embedding of nodes obtained after processing by the target generator.

[0039] Step S07: Aggregate the global embedding of the node and its context information to generate the corresponding node context embedding.

[0040] In this embodiment, within the Graph Masked Autoencoder (GMAE) framework, the masked heterogeneous information graph is input into the target encoder for message passing. The embedding representation of the latent space is obtained by aggregating the neighborhood information of nodes in the masked heterogeneous information graph. However, in real heterogeneous graphs, a current node often has complex neighborhood relationships with various types of nodes through multiple semantic links. The strategy of randomly masking edge structures introduces random graph topological perturbations into these complex neighborhood relationships. This randomness causes the prediction model to observe an unstable neighborhood instance in each training iteration. This means that two nodes with fundamentally different original neighborhoods may exhibit similar or even identical neighborhood instances after random topological perturbation, resulting in structural confusion. Although this application has demonstrated that its current prediction model possesses the ability to accurately perceive local structures and accurately capture global information, the perturbation of random graph topology still limits its ability to disentangle structural confusion caused by observed unstable neighborhood instances, thus failing to stably represent the deterministic true representation of nodes in the original graph. To address this problem, this application establishes stable neighborhood information as an additional prediction target, aiming to fundamentally solve the problem of node representation entanglement caused by random topological perturbations introduced by random masks.

[0041] Specifically, this application aims to ensure that an ideal node embedding exhibits stability under transformations that enhance the model's generalization ability. A prediction task that uses stable neighborhood information as the reconstruction target essentially forces the prediction model to decouple and ignore random noise caused by perturbations, while capturing more stable neighborhood structures and information. Under this strong structural consistency regularization constraint, the prediction model can learn stable node representations that are invariant to local topological perturbations and can realistically represent neighborhood structural information. This additional reconstruction design and introduction enables the prediction model to learn more complex link relationships, thereby ensuring that the obtained node representation is highly robust and information-rich, laying the foundation for solving complex graph link prediction tasks. Specifically, the global node embedding of the node obtained in step S07 is aggregated with the node's context information (such as the node embeddings of all its neighboring nodes) to generate the corresponding node context embedding, such as... Figure 2 As shown, Figure 2 In The node context embedding represents the node. Based on the generated node context embedding, and the node embedding extracted by the second feature extraction layer of the target encoder, the node embedding for potential prediction is projected onto the representation space by the projector head. This forces the model to reconstruct the stable neighborhood context features of the node in the undisturbed heterogeneous information graph according to the perturbed neighborhood information of the node. This forces the model to ignore the noise brought by random perturbation, learn the invariant features of the node in the real neighborhood, and thus obtain a more stable and more generalizable node representation.

[0042] Step S08: Based on the target loss function, optimize and train the model to obtain a qualified target model for making cultural and tourism recommendations for tourists; wherein, the target loss function includes: a local loss function that evaluates the deviation between the reconstructed heterogeneous information graph and the heterogeneous information graph; a global loss function that evaluates the deviation between the target node embedding and the global node embedding; and an additional loss function that evaluates the deviation between the target node embedding and the node context embedding; the target node embedding is the node embedding extracted by the second feature extraction layer in the target encoder.

[0043] In this embodiment, a local loss function is constructed based on the reconstructed heterogeneous information graph obtained in step S05 and the original heterogeneous information graph. This local loss function is used to evaluate the deviation between the reconstructed heterogeneous information graph and the original heterogeneous information graph.

[0044] In this embodiment, a global loss function is constructed based on the target node embedding (i.e., the node embedding extracted by the second feature extraction layer of the target encoder) and the global node embedding obtained in step S06. This global loss function is used to evaluate the deviation between the node embedding after the target node embedding is projected onto the representation space by the projection head and the global node embedding. The purpose of this global loss function is to force the node embedding generated by the target encoder on the masked heterogeneous information graph to predict and match the ideal global node embedding as much as possible after transformation by a projection head (structured as an MLP Multilayer Perceptron).

[0045] In this embodiment, an additional loss function is constructed based on the target node embedding (i.e., the node embedding extracted by the second feature extraction layer of the target encoder) and the node context embedding obtained in step S07. This additional loss function is used to evaluate the deviation between the node embedding after the target node embedding is projected onto the representation space by the projection head and the node context embedding. The purpose of this additional loss function is to require the model to reconstruct the stable neighborhood context features (e.g., the aggregation of features of all its neighbors) of the node in the undisturbed heterogeneous information graph based on the perturbed neighborhood of the node. That is, it forces the model to ignore the noise caused by random perturbations and learn the invariant features of the node in the real neighborhood, thereby obtaining a more stable and more generalizable node representation.

[0046] In this embodiment, a target loss function that integrates the three loss functions is constructed based on the three loss functions. The model is optimized and trained using this target loss function. During the optimization and training process, when the calculation result of the target loss function meets the set conditions, a qualified target model that can be used for tourism recommendations for tourists is determined. The set conditions are: the target loss function gradually stabilizes with increasing training epochs (e.g., the gradient approaches zero, the range of loss value fluctuations narrows), indicating that the model training is qualified and the corresponding target model is obtained; and / or, the loss function or related indicators (e.g., accuracy, F1 score) no longer improve on the validation set, or an "inflection point" occurs (e.g., the validation loss begins to rise), indicating that the model training is qualified and the corresponding target model is obtained.

[0047] This application provides a training method for a self-supervised learning model for tourism recommendation. First, based on collected tourist and cultural tourism data within a target area, a heterogeneous information graph is constructed. This cultural tourism data includes data on cultural tourism objects and interaction data between these objects. Cultural tourism objects include at least tourists, scenic spots, historical relics, cultural and creative products, catering, and accommodation. The heterogeneous information graph is then subjected to random edge masking to generate a masked heterogeneous information graph. This masked heterogeneous information graph is input into the model's target encoder for processing. Feature extraction is performed through two feature extraction layers to obtain the node embeddings extracted from each node at different feature extraction layers. The model's cross-correlation decoder performs element-wise Hadamard product calculations on the node embeddings of every two nodes, and the results are concatenated to obtain the edge embeddings of the corresponding two nodes. Finally, the model's target decoder processes the edge embeddings... The process involves decoding the input to reconstruct the masked edges in the heterogeneous information graph, obtaining the corresponding reconstructed heterogeneous information graph. This heterogeneous information graph is then input into the model's target generator for processing, extracting the global embedding of each node. The global embeddings and their contextual information are aggregated to generate the corresponding contextual embeddings. Based on the target loss function, the model is optimized and trained to obtain a qualified target model for recommending cultural and tourism products to tourists. The target loss function includes: a local loss function evaluating the deviation between the reconstructed heterogeneous information graph and the heterogeneous information graph itself; a global loss function evaluating the deviation between the target node embedding and the global node embedding; and an additional loss function evaluating the deviation between the target node embedding and the contextual node embedding. The target node embedding is the node embedding extracted by the second feature extraction layer in the target encoder.

[0048] The core of the training method for the self-supervised learning model for tourism recommendation provided in this application lies in multi-scale collaborative decoding and reconstruction, specifically including three aspects: Cross-correlation decoding at the local neighborhood scale: This task aims to accurately reconstruct local edge relationships in a masked heterogeneous information graph. This application designs a cross-correlation decoder that receives the node embeddings of any two nodes (e.g., a tourist and a scenic spot) obtained by the target encoder. It calculates the Hadamard product between the node embeddings extracted by each feature extraction layer and concatenates them to form a rich set of interaction features (i.e., corresponding edge embeddings) corresponding to the two nodes. This interaction feature can capture the relationships between nodes at different distance scales. This cross-layer interaction design allows the model to predict direct connections not only by relying on one-hop neighbors but also by referencing information from more distant neighbors, thus effectively mitigating overfitting to the direct neighborhood while maintaining local perception accuracy.

[0049] Prediction of latent global representations at the global structural scale: To compensate for the shortcomings of the target decoder (the GNN encoder chosen in this application) in capturing long-range dependencies, this application introduces a parallel global information learning task. This application uses an external model (i.e., the target generator, Node2Vec chosen in this application) adept at capturing global structural information to encode an undisturbed heterogeneous information graph, generating an ideal global embedding for each node. Then, this application forces the node embeddings generated by the target encoder on the masked heterogeneous information graph, after transformation by a projection head (MLP Multilayer Perceptron structure), to predict and match this ideal global embedding as closely as possible. This process is akin to hiring a "teacher" (Node2Vec, the target generator) with a global perspective for a "student" (the GNN encoder, which only focuses on local information), forcing the model to learn global topological structures and long-range dependencies beyond the local neighborhood.

[0050] Additional Reconstruction for Context Stability: To address the randomness introduced by edge masks in heterogeneous information graphs and enhance the robustness of the node representations learned by the model, this application introduces a third task. This task requires the model to reconstruct the stable neighborhood context features (e.g., an aggregation of features from all its neighbors) of a node in the undisturbed heterogeneous information graph, based on the perturbed neighborhood of the node. This task forces the model to ignore noise introduced by random perturbations and learn the invariant features of the node in the real neighborhood, thereby obtaining a more stable and generalizable node representation.

[0051] Finally, this application weights and sums the loss functions of the three tasks (local link reconstruction loss, global representation prediction loss, and context reconstruction loss) to form a unified joint learning objective function for end-to-end optimization. After training, the resulting GNN encoder can generate stable and information-rich embedding vectors with both local context awareness and global perspective for any entity in a heterogeneous information graph related to cultural tourism. These vectors can be directly used for downstream tasks, such as calculating the matching degree between tourists and attractions for recommendations, discovering potentially related attractions to form new travel routes, or generating personalized interactive content for tourists' digital avatars in virtual reality scenarios.

[0052] In conjunction with the above embodiments, in one implementation, this application also provides a method for training a self-supervised learning model for tourism recommendation. In this method for training a self-supervised learning model for tourism recommendation, step S02 may include steps S021 to S022: Step S021: Randomly sample the edges of the heterogeneous information graph using a uniform random sampling strategy without replacement to obtain the corresponding masked edge subset.

[0053] In this embodiment, the specific implementation method of random masking in this application is as follows: To prevent the masked edges from being mainly composed of nodes with high influence, this application uses uniform random sampling without replacement with a certain probability to generate a subset of masked edges. First, through a uniform random sampling strategy without replacement, random sampling is performed on the observable edges in the heterogeneous information graph at a certain masking ratio. After sampling the number of edges corresponding to the set masking ratio, the corresponding subset of masked edges is obtained.

[0054] Step S022: Perform masking processing on each target edge in the heterogeneous information graph to obtain the corresponding masked heterogeneous information graph, wherein the target edge is an edge in the subset of the masked edges.

[0055] In this embodiment, each target edge in the original heterogeneous information graph is masked to obtain the corresponding masked heterogeneous information graph, wherein the target edge is the edge in the subset of masked edges obtained in step S021.

[0056] In conjunction with the above embodiments, in one implementation, this application also provides a method for training a self-supervised learning model for tourism recommendation. In this method for training a self-supervised learning model for tourism recommendation, the method further includes: Step S06_1: Align the dimension of the target node embedding with the global node embedding using the first dimension alignment module of the model to obtain the first node embedding corresponding to the target node embedding.

[0057] In this embodiment, to facilitate loss calculation, a first dimension alignment module is constructed in the model. The first dimension alignment module aligns the dimension of the target node embedding (i.e., the node embedding extracted by the second feature extraction layer of the target encoder) with the dimension of the global node embedding obtained by the target generator, thereby obtaining the first node embedding corresponding to the target node embedding.

[0058] Step S07_1: Align the dimension of the target node embedding with the node context embedding using the second dimension alignment module of the model to obtain the second node embedding corresponding to the target node embedding.

[0059] In this embodiment, to facilitate loss calculation, a second dimension alignment module is constructed in the model. The second dimension alignment module aligns the dimension of the target node embedding (i.e., the node embedding extracted by the second feature extraction layer of the target encoder) with the dimension of the aggregated node context embedding to obtain the second node embedding corresponding to the target node embedding.

[0060] In this application, when the training method for a self-supervised learning model for tourism recommendation provided in the embodiments of this application includes steps S06_1 and S07_1, the global loss function for evaluating the deviation between the target node embedding and the global node embedding includes: a global loss function for evaluating the deviation between the first node embedding and the global node embedding; the additional loss function for evaluating the deviation between the target node embedding and the node context embedding includes: an additional loss function for evaluating the deviation between the second node embedding and the node context embedding.

[0061] In this embodiment, when the application includes dimensional alignment of the target node embedding to obtain the corresponding first node embedding and second node embedding, the global loss function is a loss function used to evaluate the deviation between the node embedding after the first node embedding is projected onto the representation space by the projection head and the global node embedding, while the additional loss function is a loss function used to evaluate the deviation between the node embedding after the second node embedding is projected onto the representation space by the projection head and the node context embedding.

[0062] In conjunction with the above embodiments, in one implementation, this application also provides a training method for a self-supervised learning model for tourism recommendation. In this training method for a self-supervised learning model for tourism recommendation, the global embedding of a node and its context information are aggregated to generate the corresponding node context embedding, defined as follows:

[0063] in, Embed the node context for node v. The number of relations to node v. Let q be the set of neighboring nodes of node v that are directly connected to node v through relation type q. Global embedding of node u output by the target generator.

[0064] In this embodiment, formally, for any node v, the prediction target of its neighborhood information is defined as (i.e., the node global embedding and its context information are aggregated to generate the corresponding node context embedding as defined):

[0065] in, The number of relations to node v. Let q be the set of neighboring nodes of node v that are directly connected to node v through relation type q. The node embedding representation of node u, obtained from the output of the Node2Vec encoder, is used to finally obtain the prediction target on the neighborhood information of node v, namely the node context embedding of node v obtained by aggregating the global embedding of node v and its context information. At this scale of learning objectives, the obtained... As "pseudo-labels" during the training process, they guide the target encoder to learn invariance to topological perturbations.

[0066] In conjunction with the above embodiments, in one implementation, this application also provides a training method for a self-supervised learning model for tourism recommendation. In this training method for a self-supervised learning model for tourism recommendation, the expression for the local loss function is:

[0067] in, Let z be the local loss function, and z be the set of all neighboring nodes of node v. Estimated link probability between node v and node u It is a subset of masked edges formed by the edges that are masked out in the heterogeneous information graph.

[0068] In this embodiment, during actual training, this application uses a subset of masked edges in the predicted mask heterogeneous information graph as the training target. This learning paradigm at this scale forces the prediction model to understand the topological structure between nodes, effectively alleviating the shortcomings of traditional GAE work that overemphasizes proximity information.

[0069] Specifically, let y(v,u)=y(f(v),f(u)) be the estimated link probability between node v and node u. In this application, the training objective of the prediction model at this scale is to reconstruct the masked edges in the masked heterogeneous information graph. The defined local loss function expression is:

[0070] in, Let z be the local loss function, and z be the set of all neighboring nodes of node v. Estimated link probability between node v and node u The masked edge subset, which consists of the masked edges in the heterogeneous information graph, is usually approximated by negative sampling in the denominator of the local loss function to accelerate training.

[0071] In conjunction with the above embodiments, in one implementation, this application also provides a training method for a self-supervised learning model for tourism recommendation. In this training method for a self-supervised learning model for tourism recommendation, the expression for the global loss function is:

[0072] Where τ is the temperature parameter, To calculate the cosine similarity between the first node embedding in the latent space and the global embedding of the target node, and The target nodes are globally embedded for the i-th and j-th nodes, respectively. and The first node embeddings of the i-th and j-th nodes are respectively, and N represents the number of negative samples.

[0073] In this embodiment, at this scale, the global node embedding output by the Node2Vec encoder, which acts as the target generator, is used as the target embedding. Given a heterogeneous information graph as input, the Node2Vec encoder, acting as the target generator, generates target embeddings (i.e., global node embeddings) that preserve path-aware topology information through two traversal strategies: breadth-first sampling (BFS) and depth-first sampling (DFS). To accurately capture the variable dependency between the node embeddings of the masked heterogeneous information graph (i.e., the node embeddings extracted by the second feature extraction layer in the target encoder) and the target embedding output by the Node2Vec encoder, the InfoNCE method is used as the global loss function, expressed as:

[0074] Where τ is the temperature parameter, To calculate the cosine similarity between the first node embedding in the latent space and the global embedding of the target node, and The target nodes are globally embedded for the i-th and j-th nodes, respectively. and The first node embeddings of the i-th and j-th nodes are respectively, and N represents the number of negative samples.

[0075] In conjunction with the above embodiments, in one implementation, this application also provides a training method for a self-supervised learning model for travel recommendation. In this training method for a self-supervised learning model for travel recommendation, the expression for the additional loss function is:

[0076] Where V is the set of nodes in the heterogeneous information graph. Embed the second node of the v-th node. Embed the target node context for the v-th node.

[0077] In this embodiment, for all nodes in the heterogeneous information graph, this application minimizes the node embedding generated by the model in the masked heterogeneous information graph. The prediction target is its neighborhood information (i.e., node context embedding). Using the difference between them as the training objective, an additional loss function is designed, as follows:

[0078] By optimizing the additional loss function, the model is explicitly incentivized to decouple from the unstable structural noise caused by random masks, and instead focus on learning more fundamental and generalizable patterns in the neighborhood, thereby endowing the model with stronger robustness and representational ability.

[0079] In conjunction with the above embodiments, in one implementation, this application also provides a training method for a self-supervised learning model for tourism recommendation. In this training method for a self-supervised learning model for tourism recommendation, the expression for the target loss function is:

[0080] in, For local loss functions, For the global loss function, For the additional loss function, α and β represent the non-negative hyperparameters that balance the corresponding loss.

[0081] In this embodiment, this application effectively solves three problems in graph structure learning—accurate perception of local neighborhoods, effective capture of global structural information, and representation invariance to graph topological perturbations—by carefully designing this multi-scale collaborative learning framework. In actual training, the objectives of all proposed strategies are complementary. Therefore, this application constructs the final learning objective (i.e., the objective loss function) by weighted summation of the above three losses to jointly optimize the model's performance, thereby guiding the model to learn and capture stable representations that aggregate local neighborhood and global structural information. Formally, this application defines the objective function as follows:

[0082] in, For local loss functions, For the global loss function, For the additional loss function, α and β represent the non-negative hyperparameters that balance the corresponding loss.

[0083] In conjunction with the above embodiments, in one implementation, this application also provides a training method for a self-supervised learning model for tourism recommendation. In this training method for a self-supervised learning model for tourism recommendation, the target decoder, the first dimension alignment module, and the second dimension alignment module are multilayer perceptrons.

[0084] In this embodiment, the target decoder, the first dimension alignment module, and the second dimension alignment module in the training method of the self-supervised learning model for tourism recommendation provided in this application are preferably multilayer perceptrons (MLPs).

[0085] Based on the same inventive concept, this application provides a tourism recommendation method using a self-supervised learning model, applied to the target model trained by the training method for a self-supervised learning model for tourism recommendation described in the first aspect of this application, such as... Figure 3 As shown, the method includes: Step S001: Based on the acquired cultural and tourism data of the target tourists in the target area, construct a heterogeneous information map corresponding to the target tourists.

[0086] In this embodiment, the cultural and tourism data of the target tourists in the target area is first acquired. The target tourist is a specific tourist object, or a group of tourists with similar user attributes (e.g., a group of tourists aged 45 to 50 with the same occupation). The cultural and tourism data of the target tourists in the target area includes all other cultural and tourism objects in the target area besides the tourist-type cultural and tourism objects (e.g., scenic spots, historical relics, cultural and creative products, catering, and accommodation), as well as interaction data showing interactions between these cultural and tourism objects by the target tourists. Based on the acquired cultural and tourism data of the target tourists in the target area, a heterogeneous information map targeting the target tourists is constructed.

[0087] Step S002: Input the heterogeneous information graph corresponding to the target tourist into the target model to perform edge link prediction processing between nodes to obtain the corresponding target heterogeneous information graph. In actual application, the structure of the target model includes a target encoder, a cross-correlation decoder, and a target decoder.

[0088] In this embodiment, after obtaining the heterogeneous information graph corresponding to the target tourist through step S1, the heterogeneous information graph is input into the target model trained by the training method of the self-supervised learning model for tourism recommendation provided in the first aspect of this application to perform edge link prediction processing between nodes, thereby obtaining the corresponding target heterogeneous information graph. The target heterogeneous information graph includes which cultural and tourism objects the target tourist will interact with, as predicted by the target model. For example, if tourist a has not visited attraction 1, but the target model predicts that tourist a will interact with attraction 1 based on the heterogeneous information graph corresponding to the target tourist to which tourist a belongs, then an edge connecting tourist a and attraction 1 will be created in the target heterogeneous information graph.

[0089] In this embodiment, during practical application, the target generator, the first dimension alignment model and the second dimension alignment module in the prediction model of this application are only used during the model training phase. After training to obtain a qualified target model, the structure of the target model includes a target encoder, a cross-correlation decoder and a target decoder.

[0090] Step S003: Based on the target heterogeneous information map obtained from the prediction, make travel recommendations to the target tourists.

[0091] In this embodiment, after obtaining the target heterogeneous information graph corresponding to the target tourist through step S2, tourism recommendations are made to the target tourist based on the edges between the predicted nodes in the target heterogeneous information graph. For example, if it is predicted that tourist a will interact with attraction 1, attraction 1 is recommended to tourist a.

[0092] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of this application.

[0093] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0099] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal 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 process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0100] The above provides a detailed description of the training and recommendation method for a self-supervised learning model for tourism recommendation provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A training method for a self-supervised learning model for tourism recommendation, characterized in that, The method includes: Based on the collected tourist and cultural tourism data within the target area, a heterogeneous information map is constructed. The cultural tourism data includes data on cultural tourism objects and interaction data between cultural tourism objects. The cultural tourism objects include at least tourists, scenic spots, historical relics, cultural and creative products, catering, and accommodation. Randomly mask the edges of the heterogeneous information graph to generate a masked heterogeneous information graph; The mask heterogeneous information map is input into the target encoder of the model for processing. Feature extraction is performed through two feature extraction layers to obtain the node embedding of each node extracted in different feature extraction layers. The Hadamard product is calculated element-wise for the node embeddings of every two nodes by using the cross-correlation decoder of the model, and the results are concatenated to obtain the edge embeddings of the corresponding two nodes. The edge embedding is decoded by the target decoder of the model to reconstruct the masked edges in the heterogeneous information graph and obtain the corresponding reconstructed heterogeneous information graph. The heterogeneous information graph is input into the target generator of the model for processing, and the global embedding of each node is extracted. The global embedding of the node and its context information are aggregated to generate the corresponding node context embedding. Based on the target loss function, the model is optimized and trained to obtain a qualified target model for recommending cultural and tourism products to tourists. The target loss function includes: a local loss function that evaluates the deviation between the reconstructed heterogeneous information graph and the heterogeneous information graph; a global loss function that evaluates the deviation between the target node embedding and the global node embedding; and an additional loss function that evaluates the deviation between the target node embedding and the node context embedding. The target node embedding is the node embedding extracted by the second feature extraction layer in the target encoder.

2. The training method for a self-supervised learning model for tourism recommendation according to claim 1, characterized in that, The method further includes: The first dimension alignment module of the model aligns the dimension of the target node embedding with the global node embedding to obtain the first node embedding corresponding to the target node embedding. The global loss function for evaluating the deviation between the target node embedding and the global node embedding includes: a global loss function for evaluating the deviation between the first node embedding and the global node embedding; By using the second dimension alignment module of the model, the dimension of the target node embedding is aligned with the node context embedding to obtain the second node embedding corresponding to the target node embedding; The additional loss function for evaluating the deviation between the target node embedding and the node context embedding includes: an additional loss function for evaluating the deviation between the second node embedding and the node context embedding.

3. The training method for a self-supervised learning model for tourism recommendation according to claim 1, characterized in that, Randomly mask the edges of the heterogeneous information graph to generate a masked heterogeneous information graph, including: The heterogeneous information graph is randomly sampled by a uniform random sampling strategy without replacement to obtain the corresponding masked edge subset; Each target edge in the heterogeneous information graph is masked to obtain a corresponding masked heterogeneous information graph, wherein the target edge is an edge in the subset of the masked edges.

4. The training method for a self-supervised learning model for tourism recommendation according to claim 1, characterized in that, The global node embedding and its context information of a node are aggregated to generate the corresponding node context embedding, defined as follows: in, Embed the node context for node v. The number of relations to node v. Let q be the set of neighboring nodes of node v that are directly connected to node v through relation type q. Global embedding of node u output by the target generator.

5. The training method for a self-supervised learning model for tourism recommendation according to claim 2, characterized in that, The expression for the local loss function is: in, Let z be the local loss function, and z be the set of all neighboring nodes of node v. Estimated link probability between node v and node u It is a subset of masked edges formed by the edges that are masked out in the heterogeneous information graph.

6. The training method for a self-supervised learning model for tourism recommendation according to claim 2, characterized in that, The expression for the global loss function is: Where τ is the temperature parameter, To calculate the cosine similarity between the first node embedding in the latent space and the global embedding of the target node, and The target nodes are globally embedded for the i-th and j-th nodes, respectively. and The first node embeddings of the i-th and j-th nodes are respectively, and N represents the number of negative samples.

7. The training method for a self-supervised learning model for tourism recommendation according to claim 2, characterized in that, The expression for the additional loss function is: Where V is the set of nodes in the heterogeneous information graph. Embed the second node of the v-th node. Embed the target node context for the v-th node.

8. The training method for a self-supervised learning model for tourism recommendation according to claim 2, characterized in that, The expression for the target loss function is: in, For local loss functions, For the global loss function, For the additional loss function, α and β represent the non-negative hyperparameters that balance the corresponding loss.

9. The training method for a self-supervised learning model for tourism recommendation according to claim 2, characterized in that, The target decoder, the first dimension alignment module, and the second dimension alignment module are multilayer perceptrons.

10. A tourism recommendation method based on a self-supervised learning model, characterized in that, The target model obtained by training the self-supervised learning model for tourism recommendation according to any one of claims 1 to 9, the method comprising: Based on the acquired cultural and tourism data of the target tourists in the target area, a heterogeneous information map corresponding to the target tourists is constructed. The heterogeneous information graph corresponding to the target tourist is input into the target model for edge link prediction processing between nodes to obtain the corresponding target heterogeneous information graph. In practical applications, the structure of the target model includes a target encoder, a cross-correlation decoder, and a target decoder. Based on the target heterogeneous information map obtained from the prediction, travel recommendations are made for the target tourists.

Images