dynamic generation of personalized hybrid recommender systems

Post on 20-Aug-2015

199 Views

Category:

Science

3 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Dynamic Generation of Personalized Hybrid Recommender Systems

Simon Dooms

About …

Simon Dooms PhD Student Ghent University Belgium

FeedbackResultsLearningFrameworkIntro

2009 2010 2014

Research Personal PhD Grant (4 years)

October, 2013

RatingsDatasetsOnline FeedbackOnline experimentsDistributed recsysHybrid recsysUser-centric

Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013 2

Information Overload

FeedbackResultsLearningFrameworkIntro

Collaborative Filtering

Content-based FilteringItemAttributeKNN

Content-based Filtering

FactorWiseMatrixFactorization

BiasedMatrixFactorization

MatrixfactorizationItemKNN

Random Items

Popular Items

SigmoidSVDPlusPlus

SigmoidCombinedAsymmetricFactorModel

SigmoidItemAsymmetricFactorModel

SigmoidUserAsymmetricFactorModel GlobalAverage

ItemAverage

SVDPlusPlus

TimeAwareBaselineWithFrequencies

CoClusteringSlopeOne

UserItemBaseline

UserKNN

BiPolarSlopeOne

NaiveBayes TimeAwareBaseline

Probability-based Extended Profile Filtering

LatentFeatureLogLinearModelSVD

3Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Recommendation Algorithm Overload

What about Hybrids? Combine the merits!BUT Typical Hybrid: CB + CF Manual algorithm selection, static configuration

FeedbackResultsLearningFrameworkIntro

4Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Our Goal

Generate personalized hybrid recommender systems

Research Questions– Do all users benefit from personalized hybrid recommenders?– How automatically adapt a hybrid recommender?– How evaluate the system?– How respond to real-time online user feedback?

FeedbackResultsLearningFrameworkIntro

Automatically

Different for every user

Involving many different algorithms

5Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

A Recommender FrameworkFeedbackResultsLearningFrameworkIntro

MovieTweetings

MyMediaLite

Python Code

HTML front-end

6Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Learning module –an optimization problem–

FeedbackResultsLearningFrameworkIntro

Optimize weights such that is minimized.

How evaluate ? On what data?

Objective function: Evaluation metric (e.g., RMSE, MAE, …)

7Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Optimization parametersFeedbackResultsLearningFrameworkIntro

Training TestFold 1

Training TestFold 2

Training TestFold 3

Ratings for 1 user

Recommendation algorithms:

8

User algorithm weights:

Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Optimize

FeedbackResultsLearningFrameworkIntro

Slow (hours) Fast (seconds)

Fol

d da

tase

tsA

ll da

ta

9

Results (offline)

Experimented with Switching strategies A genetic algorithm Binary search tree

Statistically significant results, journal article: “Offline Optimization for User-specific hybrid Recommender Systems”

FeedbackResultsLearningFrameworkIntro

10Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Demo

Online TOUR Offline TOUR

FeedbackResultsLearningFrameworkIntro

11Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

FeedbackResultsLearningFrameworkIntro

FeedbackResultsLearningFrameworkIntro

FeedbackResultsLearningFrameworkIntro

FeedbackResultsLearningFrameworkIntro

FeedbackResultsLearningFrameworkIntro

CFF: Call For Feedback

Interesting topics include, but are not limited to Optimization suggestions Offline evaluation strategy Online evaluation strategy Interesting integration ideas

FeedbackResultsLearningFrameworkIntro

17Oct. 12, 2013 Simon Dooms - Ghent University - RecSys 2013

Dynamic Generation of Personalized Hybrid Recommender Systems

Simon Dooms

top related