Abstract Details

<< Back to Schedule

10/14/2015  |   10:45 AM - 11:30 AM   |  Atlantic II

Learning in Rule-based Recommendation Systems

Rule-based systems are pervasive and are used in multiple domains. Rule-based systems rely on codifying knowledge of experts in the form of rules. A rule engine then fires applicable rules based on a user context. We consider an example of a course recommender system that is rule-based. Such systems have been built earlier. We enhance rule-based systems by introducing machine learning based on the principle of reinforcement learning. The recommender system relies on reinforcement learning, where the system is rewarded for "good" recommendations and penalized for "bad" recommendations. As a result the rule-base is dynamic. We also allow users to extend rule-base using domain-specific language constructs. Our main contribution is to bring reinforcement learning to a multi-criteria recommendation system. We use a rule-based system to make recommendations. Rules are fired based on salience. We externalize the salience to outside the rule engine so that salience can be determined by application of learning techniques. Our recommendation algorithm has the following main steps: Compute Salience: We externalize salience computation. Salience is not a rule attribute in our system. Salience computation involves learning and collaboration amongst users based on their feedback of earlier recommendations. Apply multi-criteria filtering: Based on user profile and user input, we apply multi-criteria to preselect rules that could be potentially fired. This is a key step in rule selection. We use criteria such as pre-requisites, prior grades in similar courses, and student's interest areas. Fire Rules and Match: Based on the rules selected, in the previous step, we fire rules according to the computed salience. We match rules according to the user goals specified in he user input. We choose all matching rules. The consequence of each matched rule will be a recommendation. Apply overrides: A set of overrides are applied. These override rules could come from experts in the domain. The overrides will ensure conflict resolutions between recommendations or ensure that certain courses are either recommended or not recommended based on configurable parameters. This allows for the recommender system to be customizable. Make Recommendations: Make final recommendations based on matched rules and application of overrides. The recommendations are segregated based on similarity of courses so that students can pick one among similar courses. Get User Feedback: Users can provide feedback on course recommendations at any time in future. Apply reinforcement learning: The Course Recommendation System learns using a Reinforcement Learning Methodology[6]. Reinforcement Learning trains the system to take actions of course recommendations in the environment based on a reward systems. The idea is to maximize cumulative rewards for successive good recommendations. Similarly not favorable recommendations are penalized. In a rule-based system, rewards and penalties translate to saliences of rules. When a recomendation is favorable, the parameters that influence salience are bumped up, and for unfavorable recommendations, the salience factors are reduced. Collaborate across users: Reinforcement learning is applied accross users in a collaborative manner making feedback available to all users. This collaborative learning enhances the efficacy of learning.

Presentation:
This presentation has not yet been uploaded.

Handouts:
No handouts have been uploaded.

Arvind Hudli (Co-Presenter,Author,Co-Author), M.S. Ramaiah Institute of Technology, arvind@hudli.com;
Arvind is a senior Computer Science student at M.S. Ramaiah Institute of Technology. Bangalore India. His research interests are Artificial Intelligence and Big Data.

Shrinidhi Hudli (Primary Presenter,Author), UCLA , shrinidhi@hudli.com;
Shrinidhi has a B.E in Computer Science from VTU India and an MS in Computer Science from UCLA, Los Angeles. He has completed an internship at Yellow Pages, a big data startup company in Glendale, CA. He is currently at Apple, Inc.

2013 Sponsors: IEEE and IEEE Computer Society