Recommendation systems (RSs) have become one of the most essential tools for many e-commerces, streaming services, and social network websites. By utilizing users’ past interactions (such as purchase history, clicks, or ratings) with the systems, RSs aim to deliver personalized recommendation lists of products or services that would best match each individual user’s personal preference. However, many RSs suffer from the data sparsity problem where many users have only a few interactions, which is insufficient to produce accurate recommendation results. Because of this, many RSs try to incorporate the additional information into consideration when making recommendations. Among all available information, user-generated reviews are one of the most popular resources frequently used in many RS research. The main reason is that there is much useful information that can be extracted from reviews, such as user sentiments or contextual information. Moreover, many researchers have been developing recommendation algorithms that learn user and item representations from review data using recent deep learning techniques. However, there are still challenges in utilizing review data in RSs, such as how to extract useful information from reviews efficiently, or how to detect and develop robust RS algorithms against the presence of spam reviews.
Possible topics that interns can join:
- Representation learning based on language models (e.g., Transformers, BERT)
- Text-based information extraction (e.g., sentiment analysis, context-aware recommendations)
- Spam review detection/robust recommendations
- Advance in deep learning-based recommendation systems