Methods and Systems for Establishing Tourism Intent Prediction Models

Abstract

A system for building a travel intention prediction model includes a computing device and a server device. For each user's training data, the computing device filters out multiple candidate multi-source information items according to a predetermined time interval, and filters out multiple travel-related multi-source information items related to travel activities according to the annotations of these candidate multi-source information items. Furthermore, it aggregates these travel-related multi-source information items along a time dimension according to a predetermined time granularity to obtain multiple data blocks. The server device uses a large language model to generate multiple dense embedding vectors from these data blocks. The computing device trains a machine learning model based on these dense embedding vectors to obtain a travel intention prediction model for predicting whether there is travel intention in at least a specific time interval.

Claims

1. A method for establishing a tourism intention prediction model, executed by a system storing multiple sets of user training data corresponding to multiple users, each set of user training data including multiple sets of multi-source information, such as multiple sets of foreign currency account information, multiple sets of bank product browsing records information, and multiple sets of credit consumption merchant information, each set of multi-source information having a timestamp and at least one label indicating whether it is related to tourism activities, the system running a large language model, the method comprising the following steps: (A) for each set of user training data, according to a predetermined time interval, filtering out multiple sets of candidate multi-source information with timestamps within the predetermined time interval from the multi-source information of the user training data; (B) for each set of user training data, according to the labels of the candidate multi-source information of the user training data, filtering out multiple sets of tourism-related multi-source information related to tourism activities from the candidate multi-source information; (C) for each set of user training data, according to a predetermined time granularity, aggregating the tourism-related multi-source information of the user training data along the time dimension to obtain multiple data blocks corresponding to the predetermined time granularity. (D) For each user training data, the large language model generates multiple dense embedding vectors corresponding to the data blocks corresponding to the user training data; and (E) For each user training data, a machine learning model is trained based on the dense embedding vectors corresponding to the user training data to obtain a travel intention prediction model for predicting whether there is a travel intention in at least one specific time interval.

2. The method for establishing a tourism intention prediction model as described in claim 1 further includes the following steps between steps (C) and (D): (F) for each user training data, filtering out multiple pieces of tourism-irrelevant multi-source information that are unrelated to tourism activities from the candidate multi-source information based on the annotations of the candidate multi-source information in the user training data; and (G) for each user training data, incorporating the tourism-irrelevant multi-source information in the user training data into the embedding representation of the data blocks.

3. The method for establishing a travel intention prediction model as described in claim 1, wherein each user's training data further includes a travel record of the user's travel time, wherein, Step (E) includes the following sub-steps: (E-1) For each user training data, using the machine learning model to generate a prediction result based on the dense embedding vectors corresponding to the user training data; and (E-2) For each user training data, adjusting the parameters of the machine learning model based on the travel records of the user training data and the prediction result.

4. The method for establishing a tourism intention prediction model as described in claim 3, wherein, In step (E-1), the dense embedding vectors are concatenated sequentially to generate a concatenated dense embedding vector, which is then input into the machine learning model to produce the prediction result.

5. A system for building a tourism intention prediction model, comprising: a computing device storing multiple sets of user training data corresponding to multiple users, each set of user training data including multiple sets of multi-source information, such as multiple sets of foreign currency account information, multiple sets of bank product browsing records information, and multiple sets of credit consumption merchant information, each set of multi-source information having a timestamp and at least one label indicating whether it is related to tourism activities; and a server device communicatively connected to the computing device, running a large-scale language model; wherein... For each user training data set, the computing device filters multiple candidate multi-source information items with timestamps within the predetermined time interval from the multi-source information in the user training data set. Based on the annotations of the candidate multi-source information in the user training data set, it filters multiple tourism-related multi-source information items related to tourism activities from the candidate multi-source information. Based on a predetermined time granularity, it aggregates these tourism-related multi-source information items along the time dimension to obtain multiple data blocks corresponding to the predetermined time granularity. These data blocks are then transmitted to the server device. The server device uses the large language model to generate multiple dense embedding vectors corresponding to the data blocks in the user training data set. For each user training data set, the computing device trains a machine learning model based on the dense embedding vectors corresponding to the user training data set to obtain a tourism intention prediction model for predicting whether there is a tourism intention in at least a specific time interval.

6. The tourism intention prediction model building system as described in claim 5, wherein, The computing device also filters out multiple pieces of tourism-irrelevant multi-source information that are unrelated to tourism activities from the candidate multi-source information based on the annotations of the candidate multi-source information in the user training data, and integrates the tourism-irrelevant multi-source information into the embedded representation of the data blocks.

7. The tourism intention prediction model building system as described in claim 5, wherein, Each user training data also includes a travel record of the corresponding user's travel time. For each user training data, the computing device generates a prediction result using the machine learning model based on the dense embedding vectors corresponding to the user training data, and adjusts the parameters of the machine learning model based on the travel record of the user training data and the prediction result.

8. The tourism intention prediction model building system as described in claim 7, wherein, The computing device concatenates the dense embedding vectors sequentially to generate a concatenated dense embedding vector, and then inputs the concatenated dense embedding vector into the machine learning model to generate the prediction result.

Images