Upload
keaton-oneal
View
57
Download
1
Tags:
Embed Size (px)
DESCRIPTION
Item-Based Collaborative Filtering Recommendation Algorithms. Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl GroupLens Research Group/ Army HPC Research Center Department of Computer Science and Engineering University of Minnesota, Minneapolis, 2001 2008. Nov. 05 - PowerPoint PPT Presentation
Citation preview
Item-Based Collaborative Filtering Item-Based Collaborative Filtering Recommendation AlgorithmsRecommendation Algorithms
Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl
GroupLens Research Group/ Army HPC Research Center
Department of Computer Science and Engineering
University of Minnesota, Minneapolis, 2001
2008. Nov. 05
Presented by Eun-gyeong Kim, IDS Lab.
Copyright 2008 by CEBT
ContentsContents
Introduction
Collaborative Filtering Based Recommender Systems
Overview of the Collaborative Filtering Process
Challenges of User-based Collaborative Filtering Algorithms
Item-based Collaborative Filtering Algorithm
Item Similarity Computation
Prediction Computation
Performance Implications
Experimental Evaluation
Contributions
Discussion & Conclusion
IDS Lab. Seminar - 2Center for E-Business Technology
Copyright 2008 by CEBT
Introduction Introduction (What is Collaborative (What is Collaborative filtering?)filtering?)
Now it is time to create the technologies that can help us sift through all the available information to find that which is most valuable to us.
One of the most promising such technologies is collaborative filtering
Collaborative filtering (by Wikipedia)
The process of filtering for information or patterns using techniques involving collaboration among multiple agents, viewpoints, data sources, etc.
The underlying assumption of CF approach is that those who agreed in the past tend to agree again in the future
CF systems usually take two steps
Look for users who share the same rating patterns with the active user
Use the ratings from those like-minded users found in step 1 to calculate a prediction for the active user
IDS Lab. Seminar - 3Center for E-Business Technology
Copyright 2008 by CEBT
Two main Categories of CF algorithmsTwo main Categories of CF algorithms
Memory-based CF Algorithms
Utilize the entire user-item database to generate a prediction
Employ statistical techniques to find the neighbors
Model-based CF Algorithms
First developing a model of user ratings.
Computing the expected value of a user prediction , given his/her ratings on other items.
To build the model
– Bayesian network (probabilistic)
– clustering (classification)
– rule-based approaches (association rules between co-purchased items)
IDS Lab. Seminar - 4Center for E-Business Technology
Copyright 2008 by CEBT
Recommendation AlgorithmsRecommendation Algorithms
User-based collaborative filtering
Traditional Collaborative Filtering
Cluster Models
Item-based collaborative filtering
Search-based Methods
Item-to-item collaborative filtering
Amazon.com Recommendations: Item-to-Item Collaborative Filtering http://www.win.tue.nl/~laroyo/2L340/resources/Amazon-Recommendations.pdf
IDS Lab. Seminar - 5Center for E-Business Technology
Copyright 2008 by CEBT
CF Based Recommender SystemsCF Based Recommender Systems
provide item recommendations or predictions based on the opinions of other like-minded users
IDS Lab. Seminar - 6
32 4
Center for E-Business Technology
Copyright 2008 by CEBT
Traditional Collaborative Filtering (1)Traditional Collaborative Filtering (1)
Represents a customer as an N-dimensional vector of items, where N is the number of distinct catalog items
For almost all customers, this vector is extremely sparse
Generates recommendations based on a few customers(neighbors) who are most similar to the user
Measure the similarity of two customers, A and B
IDS Lab. Seminar - 7Center for E-Business Technology
Copyright 2008 by CEBT
Traditional Collaborative Filtering (2)Traditional Collaborative Filtering (2) Generate recommendations
A common technique is to rank each item according to how many similar customers purchased it
O(MN) in the worst case
Performance tends to be closer to O(M+N) because the average customer vector is extremely sparse
Scaling issues
Reduce the data size
– Reduce M by randomly sampling the customers or discarding customers with few purchases
– Reduce N by discarding very popular or unpopular items
Reduce recommendation quality
We need better algorithms to scale to large data sets and at the same time produce high-quality recommendations
IDS Lab. Seminar - 8Center for E-Business Technology
Copyright 2008 by CEBT
Challenges of User-based CF Challenges of User-based CF AlgorithmsAlgorithms
Challenges
Sparsity
– A person may have purchased well under 1% of the items
– (1% of 2 million books is 20,000 books)
– The accuracy of recommendations may be poor
Scalability
– Computation grows with both the number of users and the number of items
– Traditional CF does little or no offline computation, and its online computation scales with the number of customers and catalog items.
=> The key to item-to-item CF’s scalability and performance is that it creates the expensive similar-items table offline
IDS Lab. Seminar - 9Center for E-Business Technology
Copyright 2008 by CEBT
Item-based CF AlgorithmItem-based CF Algorithm
Similarity computation between two item i and j
First isolate the users who have rated both of these items
Then apply a similarity computation technique to determine the similarity
Prediction generation
Take a weighted average of the target user’s ratings on these similar items
IDS Lab. Seminar - 10Center for E-Business Technology
Copyright 2008 by CEBT
Item Similarity ComputationItem Similarity Computation
IDS Lab. Seminar - 11Center for E-Business Technology
Copyright 2008 by CEBT
Item Similarity ComputationItem Similarity Computation
IDS Lab. Seminar - 12
i1 i2 i3 i4 Ave
u1(out of 5)
3 5 3 3.67
u2(out of 5)
1 2 1.5
u3(out of 5)
4 2 4 2 3
u4(out of 5)
5 4 2 3.67
average 2.67 4 3.25
2
(1,2) (1,3) (1,4) (2,3) (2,4) (3,4)
0.61 0.79 0.55 0.87 0.67 0.84
-0.76 0.94 0 -0.57 0 0
-0.94 0.70 -1 -0.54 -0.38 -0.76
Center for E-Business Technology
Copyright 2008 by CEBT
Prediction ComputationPrediction Computation
IDS Lab. Seminar - 13Center for E-Business Technology
Copyright 2008 by CEBT
Prediction ComputationPrediction Computation
Weighted Sum Compute the sum of the ratings given by the user on the items
similar to I
Each ratings is weighted by the corresponding similarity
Regression Similarities computed using cosine or correlation measures
may be misleading
Approximated values based on a linear regression model are used (Instead of using the similar item N’s “raw” ratings values )
IDS Lab. Seminar - 14Center for E-Business Technology
Copyright 2008 by CEBT
Weighted Sum ExampleWeighted Sum Example
Let’s predict the value of item i1 for u4
IDS Lab. Seminar - 15
i1 i2 i3 i4 Ave
u1(out of 5)
3 5 3 3.67
u2(out of 5)
1 2 1.5
u3(out of 5)
4 2 4 2 3
u4(out of 5)
5 4 2 3.67
average 2.67 4 3.25
2Pu4,i1 Pu2,i2 Pu1,i4 Pu2,i4
3.75 1.59 3.65 1.60
4 - - -
4 - - -
(1,2) (1,3) (1,4) (2,3) (2,4) (3,4)0.61 0.79 0.55 0.87 0.67 0.84
-0.76 0.94 0 -0.57 0 0
-0.94 0.70 -1 -0.54 -0.38 -0.76
Center for E-Business Technology
Copyright 2008 by CEBT
Item-to-item CF in Amazon.comItem-to-item CF in Amazon.com
We could build a product-to-product matrix by iterating through all item pairs and computing a similarity metric for each pair.
However, many product pairs have no common customers, thus the approach is inefficient in terms of processing time and memory usage
Better approach by calculating the similarity between a single product and all related products
in the worst case
in practical
IDS Lab. Seminar - 16Center for E-Business Technology
Copyright 2008 by CEBT
Performance ImplicationsPerformance Implications
Precompute item-item similarity scores In a typical E-Commerce scenario, we usually have a set of
item that is static compared to the number of users that changes most often
Compute all-to-all similarity and then performing a quick table look-up to retrieve the required similarity values
Generating predictions for a user u on item i Retrieves the precomputed k most similar items
corresponding to the target item i
Then intersect between those k items and items purchased by the user u
The prediction is computed using basic item-based CF algorithm
IDS Lab. Seminar - 17Center for E-Business Technology
Copyright 2008 by CEBT
Experimental Evaluation: Data setExperimental Evaluation: Data set
Movie data
Data from MovieLens
– 943 users (among 43,000 users )
– 1682 movies (among over 3,500 different movies)
– 100,000 ratings (only considered users that had rated 20 or more movies)
Divided the DB into a training set and a test set.
– X=0.8 (80% of the data is used as training set)
Sparsity level:
IDS Lab. Seminar - 18Center for E-Business Technology
Copyright 2008 by CEBT
Experimental Evaluation: Evaluation Experimental Evaluation: Evaluation MetricsMetrics
Statistical accuracy metrics Mean Absolute Error (MAE) is a measure of the deviation of
recommendations from their true user-specified values.
The lower the MAE, the more accurately the recommendation engine predicts user ratings.
Decision support accuracy metrics
IDS Lab. Seminar - 19Center for E-Business Technology
Copyright 2008 by CEBT
Experimental Results (1)Experimental Results (1)
Effect of Similarity Algorithms
IDS Lab. Seminar - 20Center for E-Business Technology
Copyright 2008 by CEBT
Experimental Results (2)Experimental Results (2)
Sensitivity of Training/Test Ratio
Experiments with neighborhood size
IDS Lab. Seminar - 21Center for E-Business Technology
Copyright 2008 by CEBT
Experimental Results (3)Experimental Results (3)
Quality Experiments
IDS Lab. Seminar - 22Center for E-Business Technology
Copyright 2008 by CEBT
Sensitivity of the Model SizeSensitivity of the Model Size
The High accuracy that can be achieved using only a fraction of items
It is useful to precompute the item similarities using only a fraction of items and yet possible to obtain good prediction quality
IDS Lab. Seminar - 23Center for E-Business Technology
100%
98.3%
96%
Copyright 2008 by CEBT
Impact of the model size on run-time and Impact of the model size on run-time and throughputthroughput
IDS Lab. Seminar - 24Center for E-Business Technology
Copyright 2008 by CEBT
ContributionsContributions
Analysis of the item-based prediction algorithms and identification of different ways to implement its subtasks
Formulation of a precomputed model of item similarity to increase the online scalability of item-based recommendations
An experimental comparison of the quality of several different item-based algorithms to the classic user-based (nearest neighbor) algorithms
IDS Lab. Seminar - 25Center for E-Business Technology
Copyright 2008 by CEBT
Discussion & ConclusionDiscussion & Conclusion
Discussion
Item-item scheme provides better quality of predictions than the user-user scheme
Item neighborhood is fairly static, which can be pre-computed, which results in very high online performance
Possible to retain only a small subset of items and produce reasonably good prediction quality
Conclusion
Item-based techniques allow CF-based algorithms to scale to large data sets and at the same time produce high-quality recommendations
IDS Lab. Seminar - 26Center for E-Business Technology
Copyright 2008 by CEBT
My commentsMy comments
Lack of explanations about recommendation process
Does the calculated similarity really represent the similarity of items?
Lack of explanations about the range of similarity value
Can’t we precompute the similarity of users?
IDS Lab. Seminar - 27Center for E-Business Technology
Copyright 2008 by CEBT
ReferencesReferences
Amazon.com Recommendations: Item-to-Item Collaborative Filtering http://www.win.tue.nl/~laroyo/2L340/resources/Amazon-Recommendations.pdf
Item-based Collaborative Filtering Recommendation Algorithms
http://www.grouplens.org/papers/pdf/www10_sarwar.pdf
IDS Lab. Seminar - 28Center for E-Business Technology