A safer approach to build recommendation systems on unidentifiable data
Download as PPTX, PDF0 likes96 views
The document discusses a study on creating a personalized movie recommendation system using unidentifiable data, avoiding the collection of identifiable personal information. It presents a comparative analysis of various deep reinforcement learning algorithms, concluding that the Dueling Double Deep Q-Learning (D3QN) algorithm performs best. The research underscores the feasibility of building effective recommendation systems without compromising user privacy.
A safer approach to build recommendation systems on unidentifiable data
- 1. A Safer Approach to build Recommendation Systems on Unidentifiable Data
- 2. Kishor Datta Gupta* Akib Sadmanee+ Nafiz Sadman+ Authors * University of Memphis, TN, USA + Silicon Orchard Ltd, Bangladesh
- 4. Introduction
- 5. 01/08 • Privacy is an illusion • People have become products • Identifiable information is shared knowingly/unknowingly • Unaware about identifiable/unidentifiable data • Recommendation systems exploit user privacy
- 6. 02/08 How can we build a personalized recommendation system without personal information? Identifiable Data Unidentifiable Data
- 7. 03/08 Our Contribution A comprehensive and comparative study of deep reinforcement learning algorithms (DQN, DDQN, D3QN) for devising a movie recommendation system using unidentifiable data
- 8. 04/08 It is possible to build a good enough recommendation system without using trackers and collecting identifiable sensitive personal data from the everyday user of the internet. Hypothesis
- 9. 04/08 A RL agent can learn to recommend a set of movies to a user based on only the user’s movie watch-time. Hypothesis
- 10. Data preparation
- 11. 05/08 • Pre-existing datasets comprise identifiable data like User ID, rating, clicks etc. • 20 human annotators • Each annotator was asked to watch 50 movies • We record the watch time of an annotator watching a particular movie • Why not consider movie watching frequency?
- 13. 06/08 Q-learning (QN) Deep Q-learning (DQN) Double Deep Q-learning (DDQN) Dueling Double Deep Q-learning (D3QN) Reward Reward Reward Episodes Episodes Episodes (a) DQN (b) DDQN (c) D3QN
- 14. 07/08
- 15. Conclusion
- 16. 08/08 D3QN performs the best among the three Deep RL algorithms It is indeed possible to build a recommendation system without using any identifiable/sellable data