Stacking Collaborative Filtering for Implicit Feedback

MPhil Thesis Defence


Title: "Stacking Collaborative Filtering for Implicit Feedback"

By

Miss Lin Huang


Abstract

Stacking is a successful ensemble learning method which has been used widely in 
machine learning area. There are also many hybrid systems which can combine 
recommender algorithms using stacking technique. However, most of them focus on 
combining algorithms based on explicit feedback datasets, which are not always 
available. Implicit feedback, which expresses preferences implicitly in the 
format of users? behaviors such as click log or purchase log, is more common 
and easier to get in the most situations. So how to combine implicit 
recommender algorithms becomes a hot topic. Collaborative Filtering (CF), a 
central methodology in recommender systems, makes predictions of unknown 
preferences for a particular user based on other users? preferences. In this 
thesis, our work focuses on combining implicit CF algorithms by stacking 
framework. The procedures are arranged as follows. Firstly, we study many 
existing CF methods for implicit feedback. Secondly, we build a hybrid 
recommender system that can combine implicit CF algorithms. Our hybrid system 
aims to improve rank-based performance metrics, MAP and NDCG, which are more 
relevant and useful. In addition, we find that the popularity of items 
correlate well with many CF models in our experimental observations and propose 
to add the popularity of items as an additional feature in our combination. 
Empirical results show that our hybrid system outperforms the component model 
significantly. However, performance decreases when adding items’ popularity as 
an additional feature into meta-learning algorithm directly.


Date:			Friday, 18 June 2010

Time:			9:00am – 11:00am

Venue:			Room 3501
 			Lifts 25/26

Committee Members:	Prof. Qiang Yang (Supervisor)
 			Dr. Lei Chen(Chairperson)
 			Dr. Raymond Wong


**** ALL are Welcome ****