Sentiment analysis algorithm using contrastive learning and adversarial training for POI recommendation.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Data Science

Background:

• Recommending Points of Interest (POI) is complex, often hindered by vast, unstructured social media data.
• Traditional methods struggle to interpret user sentiment from social media, requiring extensive tagged data.

Purpose of the Study:

• To develop an aspect-based sentiment analysis model for accurate POI recommendation.
• To effectively utilize social media data with limited tagged information for travel recommendations.

Main Methods:

• Utilized BERT for word embeddings, integrating semantic text information.
• Employed contrastive learning to cluster similar sentiment aspects and separate dissimilar ones.
• Analyzed comment ratings' impact on user perception to optimize the loss function.

Main Results:

• Achieved a 13.03% increase in test accuracy and a 12.23% improvement in F1-Score compared to the BERT base model.
• Demonstrated that incorporating aspect-based sentiment attributes significantly enhances POI recommendation accuracy.

Conclusions:

• The proposed aspect-based sentiment analysis model effectively captures user sentiment from social media.
• This approach improves the accuracy and efficiency of POI recommendation systems, especially with limited labeled data.