Dissertation
Incorporating textual information with recommender systems
Doctor of Philosophy (Ph.D.), Drexel University
Dec 2019
DOI:
https://doi.org/10.17918/a7as-4p49
Abstract
Collaborative filtering-based approaches typically use structured signals, such as likes, clicks, and ratings, and predict such signals via matrix factorizations. Collaborative filtering-based approaches have shown great performance with large datasets, but also suffer from the "cold-start" problem with small datasets, and are not easily interpretable and explainable. In the Meanwhile, more and more unstructured textual information becomes available nowadays, such as reviews, comments, and tags. How to incorporate textual information into collaborative filtering-based recommender systems is a non-trivial research question, presenting new opportunities and challenges. This thesis focuses on the overarching research question concerning how to incorporate textual information into recommender systems to improve system performance and enhance model explainability. More specifically, this thesis addresses four specific research questions to answer the overarching research question. The four research questions center around the effects of using (1) textual information alone; (2) structured textual information and ratings; (3) unstructured textual information and ratings, to improve recommendation performance. The fourth research question subsequently focuses on how visualization tools can assist in gaining insights and intuitions for recommendation tasks. Three recommendation frameworks are developed to tackle the first three research questions respectively. The first framework casts the problem as a joint-ranking problem between emotion and topic concentration in text streams. The second framework uses uncertainty signals extracted from text context as additional supervised signals for link prediction. The third framework uses both quantum-like language models and a mutual-attention layer to boost the performance of rating prediction. Extensive experiments show that proposed frameworks outperform corresponding baselines.
Metrics
26 File views/ downloads
79 Record Views
Details
- Title
- Incorporating textual information with recommender systems
- Creators
- Qing Ping - DU
- Contributors
- Chaomei Chen (Advisor) - Drexel University (1970-)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xi, 114 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- Information Science (Informatics) (2013-2026); College of Computing and Informatics (2013-2026); Drexel University
- Other Identifier
- 11282; 991014632542504721