online evolutionary collaborative filtering recsys 2010 intelligent database systems lab. school of...

20
Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science & Engineering Seoul National University Center for E-Business Technology Seoul National University Seoul, Korea Presented by Sung Eun, Park 3/25/2011 Nathan N. Liu, Min Zhao, Evan Xiang, Qiang Yang Hong Kong University of Science and Technology, Hong Kong,

Upload: kathlyn-richardson

Post on 18-Jan-2018

216 views

Category:

Documents


0 download

DESCRIPTION

Copyright  2010 by CEBT Introduction  User’s preference changes over a long period of time  Online evolutionary collaborative filtering Tracks user interests over time in order to make timely recommendations  Extension of neighborhood based algorithms 3 user1 user2 PopPop JazzJazzClassic PopPop JazzJazzClassic

TRANSCRIPT

Page 1: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Online Evolutionary Collaborative Filtering

RECSYS 2010

Intelligent Database Systems Lab.School of Computer Science & Engineering

Seoul National UniversityCenter for E-Business TechnologySeoul National UniversitySeoul, Korea

Presented by Sung Eun, Park3/25/2011

Nathan N. Liu, Min Zhao, Evan Xiang, Qiang YangHong Kong University of Science and Technology, Hong Kong,

Page 2: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Contents

Introduction Evolutionary Collaborative Filtering Online evolutionary Collaborative Filtering

Incremental Similarity Computation Experiments Conclusion

2

Page 3: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Introduction

User’s preference changes over a long period of time Online evolutionary collaborative filtering

Tracks user interests over time in order to make timely recommendations

Extension of neighborhood based algorithms

3

user1

user2

Pop Pop Jazz Jazz Classic

PopPop JazzJazzClassic

Page 4: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Typical Item-based Collaborative Filtering

1. Similarity Computation: Compute the item-item similarities (Cosine Similarity)

2. Neighborhood Computation: Find the most similar k-items

3. Score Prediction: Predict the unobserved ratings

4

For all users who rated both i and j

For all similar items

Page 5: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evolutionary Collaborative Filtering

Temporal Relevance Weight of rating on item i of user u at time t on parameter

α Should decrease with the amount of time that has passed Based on the assumption that older ratings are generally

less correlated with a user’s current interests or an item’s current characteristic

A time gab between the current time and the time user rated the item i

Emphasizing currently rated items

Page 6: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evolutionary Collaborative Filtering

Page 7: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evolutionary Collaborative Filtering

Similarity Computation More emphasis on the recent rating of both items Inclined to identify nearest neighbors

user1

user2

user3

Page 8: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evolutionary Collaborative Filtering

user3

user3

Page 9: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Incremental Similarity Computation The problem of efficiently updating the model as new data

arrives over time in large volumes

Online Evolutionary Collaborative Filtering

where

Page 10: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Online Evolutionary Collaborative Filtering

Incremental Similarity Computation A set of users who newly rated i in time step t

Page 11: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Online Evolutionary Collaborative Filtering

Page 12: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Experiments

Dataset Analysis Early ratings were often

high – probably because they are

most voted by the most enthusiastic fans

Slow increase over time– Very old movies that has

watched were often classics and therefore more likely to receive high ratings

The variance of users’ ratings tended to increase over time– Better to user old age

users to catch a explicit preference

Movie age

User age

Page 13: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Dataset Analysis

A user tended to rate many more movies when he joined and became less and less active over time

Experiments

Movie age

user age

Page 14: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evaluation Measure

RMSE(Root mean square error) The rating prediction accuracy How close their predicted ratings are to the true ratings

MAP(Mean Average Precision) Choice Prediction

Page 15: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evaluation Results

Effect of temporal relevance weighting

RMSE MAP

1. In the item-based algorithm, the predicted scores are obtained by averaging a target user’s very few observed ratings 2. temporal relevance weighting’s effect to further reduce the contribution of old ratings would make the prediction less robust

Page 16: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evaluation Results

Effect of temporal relevance weighting

Why are improvements on new items important?1. the cosine similarity tends to favor old movies

Old movies get more ratings and their cosine similarity with other movies tend to be higher

2. Reflects users’ current interests

RMSE

Movie Age Group User Age Group

Better at new items

Better at old users

Page 17: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evaluation Results

Effect of temporal relevance weighting

Better at old users Consistent with the intuition that it is more likely for the

taste of old users to have drifted over time

MAP

Movie Age Group User Age Group

Page 18: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Evaluation Results

Effect of incremental computation Incremental algorithm is 15 -20 times faster than the non

incremental algorithm

Page 19: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Copyright 2010 by CEBT

Conclusion

The use of temporal relevance weighting could lead to more significant improvements for the choice prediction task than for the rating prediction task.

A detailed analysis reveals that our algorithm can most effectively improve predictions for older users and newer items.

The proposed algorithm is simple and fast enough to cope with frequent data updates

Page 20: Online Evolutionary Collaborative Filtering RECSYS 2010 Intelligent Database Systems Lab. School of Computer Science  Engineering Seoul National University

Q&A

Thank you

20