CN-121998018-A - Training and recommending method of self-supervision learning model for travel recommendation

CN121998018ACN 121998018 ACN121998018 ACN 121998018ACN-121998018-A

Abstract

The application provides a training and recommending method of a self-supervision learning model for travel recommendation, which relates to the technical field of data processing, and comprises the steps of constructing a heterogeneous information graph based on tourist data, carrying out edge random masking, inputting the obtained information graph into a target encoder to extract characteristics, and obtaining node embedding; the method comprises the steps of calculating node embedding of each two nodes through a cross-correlation decoder and splicing results 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 the node global embedding and the context information thereof to obtain node context embedding, and optimizing and training a model based on a target loss function consisting of the reconstructed heterogeneous information graph, the node embedding, the node global embedding and the node context embedding to obtain a target model qualified in training. The prediction accuracy of the travel recommendation model is improved.

Inventors

Zuo Enguang
LV XIAOYI
XIE XIA
CHEN CHEN
CHEN CHENG
ZHANG GUOTAI
LI MIN

Assignees

新疆大学

Dates

Publication Date: 20260508
Application Date: 20251203

Claims (10)

1. A method of training a self-supervised learning model for travel recommendation, the method comprising: Constructing a heterogeneous information graph based on acquired tourist travel data in a target area, wherein the travel data comprises travel object data and interaction data among travel objects, and the travel objects at least comprise tourists, scenic spots, historical relics, commodity created by texts, catering and accommodation; Carrying out edge random mask processing on the heterogeneous information graph to generate a mask heterogeneous information graph; Inputting the mask heterogeneous information graph to a target encoder of a model for processing, and extracting features through two feature extraction layers to obtain node embedding of each node extracted in different feature extraction layers; Performing element-by-element Hadamard product calculation on node embedding of every two nodes through a cross-correlation decoder of the model, and splicing calculation results to obtain edge embedding of the corresponding two nodes; Decoding the edge embedding through a target decoder of the model to reconstruct the masked edge in the heterogeneous information graph to obtain a corresponding reconstructed heterogeneous information graph; Inputting the heterogeneous information graph to a target generator of a model for processing, and extracting node global embedding of each node; the node global embedding and the context information thereof are aggregated to generate corresponding node context embedding; And optimizing the model based on a target loss function to obtain a target model which is qualified in corresponding training and is used for recommending a travel for a tourist, wherein the target loss function comprises a local loss function for evaluating deviation between the reconstructed heterogeneous information graph and the heterogeneous information graph, a global loss function for evaluating deviation between target node embedding and node global embedding, an additional loss function for evaluating deviation between target node embedding and node context embedding, and the target node embedding is node embedding extracted by a second-layer feature extraction layer in a target encoder.
2. A method of training a self-supervised learning model for travel recommendation as claimed in claim 1, further comprising: Aligning the dimension of the target node embedding with the node global embedding through a first dimension alignment module of the model to obtain a first node embedding corresponding to the target node embedding; the global loss function for evaluating the deviation between the target node embedding and the node global embedding comprises evaluating the global loss function for the deviation between the first node embedding and the node global embedding; Aligning the dimension of the target node embedding with the node context embedding through a second dimension alignment module of the model to obtain a second node embedding corresponding to the target node embedding; The evaluating the additional loss function of the deviation between the target node embedding and the node context embedding includes evaluating the additional loss function of the deviation between the second node embedding and the node context embedding.
3. The method for training a self-supervised learning model for travel recommendation as recited in claim 1, wherein performing edge stochastic masking processing on the heterogeneous information map to generate a masked heterogeneous information map comprises: Carrying out edge random sampling on the heterogeneous information graph through a uniform random sampling strategy without replacement to obtain a corresponding mask edge subset; and carrying out mask processing on each target edge in the heterogeneous information graph to obtain a corresponding mask heterogeneous information graph, wherein the target edges are edges in the mask edge subset.
4. The training method of a self-supervised learning model for travel recommendation as set forth in claim 1, wherein the node global embedding of the nodes and the context information thereof are aggregated to generate the corresponding node context embedding definition as follows: Wherein, the The node context for node v is embedded, Representing the number of relationships of the node v, For a set of neighbor nodes where node v is directly connected to node v by relationship type q, The node for node u output by the target generator is globally embedded.
5. A method of training a self-supervised learning model for travel recommendation as recited in claim 2, the method is characterized in that the expression of the local loss function is as follows: Wherein, the For the local loss function, z is the set of all neighbor nodes of node v, For estimating the probability of linking between node v and node u A subset of masked edges that are formed for the masked edges in the heterogeneous information map.
6. A method of training a self-supervised learning model for travel recommendation as recited in claim 2, the global loss function is characterized in that the expression of the global loss function is as follows: Wherein, tau is a temperature parameter, To calculate cosine similarity between the first node embedment and the target node global embedment in the potential space, And The target nodes for the i-th and j-th nodes respectively are globally embedded, And The first node, i and j nodes respectively, is embedded and N represents the number of negative samples.
7. A method of training a self-supervised learning model for travel recommendation as recited in claim 2, wherein the expression of the additional loss function is: Wherein V is a node set of the heterogeneous information graph, A second node embedded for the v-th node, The target node context for the v-th node is embedded.
8. A method of training a self-supervised learning model for travel recommendation as recited in claim 2, the target loss function is characterized in that the expression of the target loss function is as follows: Wherein, the As a function of the local loss, As a function of the global loss, As an additional loss function, α and β represent non-negative super-parameters that balance the corresponding losses, respectively.
9. The method of claim 2, wherein the target decoder, the first dimension alignment module, and the second dimension alignment module are multi-layered perceptrons.
10. A travel recommendation method of a travel recommendation self-supervised learning model, applied to a training method of a travel recommendation self-supervised learning model as recited in any one of claims 1 to 9, the method comprising: constructing a heterogeneous information graph corresponding to the target tourist based on acquired travel data of the target tourist in a target area; inputting the heterogeneous information graph corresponding to the target tourist into the target model to perform inter-node edge link prediction processing to obtain a corresponding target heterogeneous information graph, wherein in the actual application process, the structure of the target model comprises a target encoder, a cross-correlation decoder and a target decoder; And recommending the travel of the target tourist according to the target heterogeneous information map obtained through prediction.

Description

Training and recommending method of self-supervision learning model for travel recommendation Technical Field The application relates to the technical field of data processing, in particular to a training and recommending method of a self-supervision learning model for travel recommendation. Background In the current intelligent travel technology context, personalized recommendation and services mainly rely on traditional collaborative filtering, content-based recommendation algorithms and elementary graph neural network applications. The technologies construct entities such as tourists, scenic spots, hotels, commodities and the like into a relationship map, and the relationship map is analyzed to recommend the entities. The main recommendation method is a recommendation method based on a graph self-encoder (GAE Graph Autoencoder) and graph contrast learning (GCL Graph Contrastive Learning). These schemes represent the dominant approach in the current graph data mining field, but expose significant limitations in the complex scenario of intelligent travel. The core idea of the solution based on generative learning (e.g. GAE) is to learn node representations by reconstructing existing connections in the graph (e.g. attractions visited by guests). Such a method captures "local neighborhood information" well, e.g., it can accurately recommend restaurants that are nearby or functionally similar to the guest's current location. However, the problem is that these direct, short-range connections are overstressed and fitted, sacrificing insight into global structural information. This makes it difficult for the system to find potential long-term interests or thematic preferences of the guest, and to recommend content that is geographically remote but highly relevant to cultural connotation. The limitation is due to the model mechanism, namely shallow layer design of the GNN encoder and direct reconstruction of the adjacency matrix, so that the field of view is limited to a few-hop neighbor, the problem of 'over-smoothing' occurs, and long-range dependence is difficult to learn. On the other hand, contrast learning (e.g., GCL) based schemes learn by maximizing mutual information between the local and global representations of the graph, with the aim of capturing the "global structural information" of the graph. In travel scenarios, such methods are good at making large categories of recommendations based on the overall representation of the guest (e.g., "historical lovers"). However, the method has the disadvantage that the perception capability of local details is weak, and the slight interest change of tourists in specific situations cannot be accurately understood. For example, a "historic fan" may be focusing on the military defenses of a certain dynasty while visiting the north court Gucheng, while a comparison learning model may not be able to capture this specific, dynamic point of interest due to its learning paradigm focusing more on invariance between different "views" of the graph, thereby making a generalized, redundant, and inaccurate recommendation. The root of this problem is that contrast learning tends to sacrifice sensitivity to local topology changes in order to learn global invariance. Therefore, current recommendation methods cannot compromise the local context and global preferences that are critical in the travel experience, resulting in fragmentation and shallowness of the service experience. Disclosure of Invention In view of the above, the present application provides a training and recommending method for a self-supervised learning model for travel recommendation. Aims to solve or partially solve the problems existing in the background art. The first aspect of the present application provides a training method for a self-supervised learning model for travel recommendation, the method comprising: Constructing a heterogeneous information graph based on acquired tourist travel data in a target area, wherein the travel data comprises travel object data and interaction data among travel objects, and the travel objects at least comprise tourists, scenic spots, historical relics, commodity created by texts, catering and accommodation; Carrying out edge random mask processing on the heterogeneous information graph to generate a mask heterogeneous information graph; Inputting the mask heterogeneous information graph to a target encoder of a model for processing, and extracting features through two feature extraction layers to obtain node embedding of each node extracted in different feature extraction layers; Performing element-by-element Hadamard product calculation on node embedding of every two nodes through a cross-correlation decoder of the model, and splicing calculation results to obtain edge embedding of the corresponding two nodes; Decoding the edge embedding through a target decoder of the model to reconstruct the masked edge in the heterogeneous information graph to obtain a correspondin