Loading...

A Solution to Exploration/Exploitation Trade-off in Recommender Systems

Feyzabadi Sani, Mohammad Javad | 2021

897 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54588 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiee, Hamid Reza; Hosseini, Abbas
  7. Abstract:
  8. The growing use of the Internet has led to the creation of new businesses around it. Traditional businesses have to use the Internet in order to maintain their competitive conditions. One of the most important strategies for developing sales on the Internet is the proper use of recommendation systems.With the advent of businesses in cyberspace, the way has been paved for the use of recommendation systems in this space.Recommendation systems should exploit their knowledge about users’ preferences and explore their new preferences simultaneously. Establish a balance between exploring users’ new interests and exploiting known interests is key to build a good recommendation system. Existing data for training recommendation systems are biased towards recommendation policies that gathered them. This is an important challenge. Also most previous work does not consider recommendation systems as interactive systems and model them in supervised learning paradigm.In this thesis, we formulate the problem through the contextual multi-arm bandit framework and propose a solution to this trade-off using uniform gathered data and bayesian neural networks. At last, we show our method’s superiority over similar basic methods through various experiments on synthetic and real data. To show this superiority we use AUC and IPS as evaluation metrics. In different experiments we have seen 2-3% increase in AUC and about 30% increase in accumulative IPS
  9. Keywords:
  10. Recommender System ; Reinforcement ; Reinforcement Learning ; Multi-Armed Bandit Problem ; Exploration/Exploitation Trade-Off ; Bayesian Neural Networks

 Digital Object List

 Bookmark

...see more