Upload
willa-hines
View
215
Download
2
Embed Size (px)
Citation preview
CollaborativeCollaborative Filtering versus Personal Filtering versus Personal Log based Filtering:Log based Filtering:Experimental Comparison for Hotel Experimental Comparison for Hotel Room SelectionRoom Selection
Ryosuke Saga and Hiroshi Tsuji
Osaka Prefecture University
----
Dongmin Shin
IDS., SNU
2008.07.24.
Copyright 2006 by CEBT
Paper ChoosingPaper Choosing
The reason why I chose this paper
The title of paper is interesting
– The title of paper is in quite straight style
A vs B
The author should pick one method as winner
– How to utilize personal log?
– How to implement CF?
– Why is one method chosen as winner?
Center for E-Business Technology
Copyright 2006 by CEBT
IndexIndex
Introduction
Features of TPO-goods
Consideration of recommender system
Personal Log based filtering
Collaborative filtering
Simulation
Conclusion
Center for E-Business Technology
Copyright 2006 by CEBT
IntroductionIntroduction
Recommender system
Personal Log based Filtering
– Content-based
– Good for TPO-goods
Collaborative Filtering
– Good for non-TPO-goods (ex. CD and books, etc)
– Applicability to TPO-goods has not been known yet
Center for E-Business Technology
Copyright 2006 by CEBT
Features of TPO-goodsFeatures of TPO-goods
Sensitive to external factors
Season, location and event related goods
Three features
The number of attribute is high
Multiformity
– derived from several combinations of the attributes
High-frequency update
– The external factors force to update attributes of TPO-goods
Center for E-Business Technology
Copyright 2006 by CEBT
Consideration of recommender Consideration of recommender systemsystem
Rating– In order to recommend goods/services, recommender system
should rate user’s preferences
Explicit rating– Consciously rated by users
Implicit rating– Not expressed by users
– Recorded in database as log Ex. Web visiting log, sales records, etc
Rates for TPO goods..– Often time-variant
– Implicit rating is preferred An explicit rating for goods at one TPO is not the same as for the same
goods at different TPO
Center for E-Business Technology
Copyright 2006 by CEBT
Personal Log based filteringPersonal Log based filtering
Sales records work statistics analysis Pattern resulted from the analysis is expressed as
distribution– Preference distribution
– pj(x) : preference value of the attribute j on item x
Range is from 0 to 1
Three search patterns High-angle search
– from the most preferable area for user
Low-angle search– from the selected goods to the preferable area
Neighbor search– Around the selected goods without preference distribution
Center for E-Business Technology
Copyright 2006 by CEBT
Collaborative filteringCollaborative filtering
The basic premise
Similar users might like similar things
The basic processes
1. To identify the similar users on their preference
2. To recommend items witch they preferred
Sales records as Venn diagrams
Center for E-Business Technology
Copyright 2006 by CEBT
Collaborative filteringCollaborative filtering
F-measure
Used for the measurement of retrieval performance
Same tendency of the correlation in Venn diagram
Incidentally, the recall for user a is regarded as the precision for user b
Center for E-Business Technology
Copyright 2006 by CEBT
SimulationSimulation
Goal of simulation Comparing log based filtering with collaborative filtering
Simulation environment Actual data of business hotel
Provided by BestReserve Co.,Ltd
– 10,000 users
– 400,000 sales records
– 160,000 room plans
Criteria Goods fitness
– Evaluated value based on the preference extracted sales records
– K : set of attributes (price, room size, distance from mass transit and breakfast service)
Center for E-Business Technology
Copyright 2006 by CEBT
SimulationSimulation
Simulation of CF
Recommend items are not changed
– Because collaborative filtering depends on the items which are bought and evaluated by other person in spite of changing the attributes
Assume three cases
– On season, off-season, and the other season
Three price patterns As corresponding to each case
– The case of highest price, the case of lowest price, the case of average price
Center for E-Business Technology
Copyright 2006 by CEBT
SimulationSimulation
Center for E-Business Technology
Copyright 2006 by CEBT
ConclusionConclusion
TPO-goods as hotel rooms have three features Many attributes
Multiformity
High-frequency update
We could not use explicit rating for recommendation on TPO-goods
Personal log based filtering is more appropriate for the hotel room selection than collaborative filtering The accuracy of log based filtering except neighbor
search kept high performance
The accuracy of collaborative filtering was lowe3r than log based filtering and changed by TPO
Center for E-Business Technology
Copyright 2006 by CEBT
Paper EvaluationPaper Evaluation
Good Point Interesting subject & motive
Simple & easy construction and development
Clear conclusion– They made conclusion such as formula form
Actual data of business web-site
Bad point Frequent mistyping
– Even in formula
Not fully explained– Possibly explained in other paper they wrote (access impossible)
Appropriateness of criteria
Center for E-Business Technology