Transcript Document

Collaborative filtering with ordinal scale-based
implicit ratings for mobile music recommendations
S.-K. Lee et al., KAIST, Information Sciences, Vol. 180, Issue 11, pp. 2142-2155, 2010.
이시혁
[email protected]
Introduction
• Increasing variety of content in mobile web environment
– Music
– Graphics
– Games
– Other mobile content
• Searching for the music on mobile web devices
– Inefficiencies of searching sequentially
– Log on to a site to download music : best selling or newest music
– Pages through the list and selects items to inspect
– Customer : buy or repeats the steps
• Compared to PCs
– Smaller screens
– Fewer input keys
– Less sophisticated browsers
S
FT COMPUTING @ YONSEI UNIV . KOREA
1
16
Recommender system
• Collaborative filtering(CF)
– One of the variety of recommendation techniques
– Identify customers : similar to target customer and recommend
items(customers have liked)
• CF based recommender systems
– Customer profile : identify preferences and make recommendations
– Explicit ratings
• Well-understood and fairly precise, but some problems in mobile domain
• User interface of mobile devices is typically poor
• The cost of using the mobile web through these devices is high
– Implicit ratings
• Used cardinal scales to increase the accuracy of estimation
• Uncertain whether cardinal scales are also better in implicit ratings
2
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Proposed system
• CoFoSIM
– COllaborative Filtering with Ordinal Scale-based IMplicit ratings
– CF recommendation methodology for the mobile music
• mWUM
– Mobile Web Usage Mining
– Capture implicit preference information
– Apply data mining techniques to discover customer behavior
patterns by using mobile web log data
– All recorded transactions in mobile web logs are individually
analyzed
3
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Scenario of searching for music
4
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
General behavior pattern in the mobile web
• General behavior patterns in the mobile web enviornment
– Ignore : not clicking on the title
– Click-through : clicking on a certain title, viewing the detail
information
– Pre-listen : a sample of the music
– Purchase : buying the music(clicked-through or pre-listened)
• Preference order
– {music ignored(never clicked)} < {music clicked-through}
< {music pre-listened} < {music purchased}
– Greatest weight : purchased music
5
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
Phase 1 : mobile web usage mining(mWUM)
• Creating customer action
• Step1-1. data preprocessing
– including data cleaning, user identification, session identification
– Most web pages contain numerous irrelevant items(gif, jpg, swf…)
– Creating customer’s session file
• Step 1-2. customer behavior mining
– Creating specific matrix : the customer actions set
– The customer action set C : matrix
– Containing the numerical weights of the target customer’s shopping
behaviors for music items
6
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
Phase 2 : Ordinal scale-based customer profile creation
• Customer’s product interests or preferences : the customer
profile
• Requires three sequential steps
• Step2-1. preference intensity matrix creation
– Customer action set for each customer : L rows
– Limits on the number of music items(they are capable of browsing
through)
– Individual rows of customer action sets contain a part of the
preferences information
– DEF) The preference intensity matrix
if matrix for which
7
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
Phase 2 : Ordinal scale-based customer profile creation
• Step 2-2. optimal preference intensity matrix creation
^
– An optimal preference intensity matrix X
– DEF) the optimal preference intensity matrix : preference intensity
matrix
☞
8
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
Phase 2 : Ordinal scale-based customer profile creation
• Step 2-3. Ordinal scale-based customer profile creation
– Creating the ordinal scale-based customer profile for recommender
system
Requires a series of transformations(optimal preference intensity
matrix)
– Sorted as
9
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
Phase 3 : neighborhood formation and recommendation generation
• Given the customer profile
• Perform - the CF-based recommendation procedure for a target
customer
• Step 3-1. neighborhood formation
– Computes the similarity between customers and forms
– A neighborhood between a target customer and a group of likeminded customers
– Similarity : between the target customer a and other customers b
• Example 4) RAB=0.63, RAC=0.30, RAD=0.81, RAE=0.70, RAF=0.43
10
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Methodology:
Phase 3 : neighborhood formation and recommendation generation
• Step 3-2. recommendation generation
– Top-N recommendation
– Recommendation list of N music items
– Previously purchased music items : excluded, each customer’s
purchase patterns or coverage
– Music j, target customer a
• Exam6) result in exam5.
11
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Experimental environment
• Experiment design
– Live user experiment
– Benchmark system
• CoFoSIM running PC (same interface- mobile)
• cardinal scale-based recommender system (CS-RS)
• ordinal scale-based recommender system (OS-RS)
– Common factor for systems
• Fixed neighborhood size : 10
• Recommendation lists(Top-n) : 9
• Data collection
– Between May 1 and June 18, 2007
– 317 real mobile Web customers
– Previous experience purchasing music from real mobile Web sites
12
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Experimental results:
Variation of error by rating scales
• Compared the accuracy of CS-RS and OS-RS
– OS-RS average : 0.6677, higher 29% than CS-RS (during 7-weeks)
– T-test(OS-RS, CS-RS) : -4.309(d.f=138, p<0.01)
• Different mean of the correlation metric between the two systems
• OS-RS : smaller estimation error than CS-RS
13
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Experimental results:
Variation of estimation error by consensus models
• Compared CoFoSIM with OS-RS (Used the ordinal scale)
– CoFoSIM 11% higher than OS-RS
– T-test (OS-RS, CoFoSIM) : -2.822(d.f=138, p<0.01)
• Different mean of the correlation metric between the two systems
• CoFoSIM : smaller estimation error than OS-RS
14
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Experimental results:
relationship between the estimation error and recommendation quality
• Performance (precision, recall, and F1)
– OS-RS > CS-RS : 60%, 15%, and 44%
– CoFoSIM > OS-RS : 16%, 12%, and 15%
• T-test
– OS-RS and CoFoSIM- differences
– t={3.96, 16.25, and 5.43}
• One-way ANOVA test (p<0.01)
– F(precision)=32.2
– F(recall)=9.5
– F(F1)=17.9
15
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Conclusion
• CoFoSIM
– viable CF-based recommender system for the mobile web
– Enhance the quality of recommendations while mitigating customers’
burden of explicit ratings
• Customers will be able to purchase content with much lower
connection time on the mobile Web because they will be able to
easily find the desired items
• mobile content providers will be able to improve their profitability and
revenues because their purchase conversion rate will be improved
through increased customer satisfaction.
16
S
FT COMPUTING @ YONSEI UNIV . KOREA
16
Discussion
• CF-based recommender + LBS
• Drawbacks
• 분석방법
– T-test
– ANOVA
– MAE(mean absolute error)
17
S
FT COMPUTING @ YONSEI UNIV . KOREA
16