survey of recommendation systems
TRANSCRIPT
Survey of Recommendation Systems
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
Introduction
• What is recommendation system?
– Recommend related items
– Personalized experiences
• How to build a recommendation system?
– Content-Based
– Collaborative Filtering Algorithm
• Examples
– Amazon
– Youa
Examples
Browsing a book
Recommendations
Rating?
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
CF Algorithm
• Memory-Based User-Based
Item-Based
• Model-Based Bayes
Clustering
User-Based CF Algorithm
User-Based CF Algorithm
User by Item Matrix:
Table 1: An example of user-item matrix
Table 2: A simple example of ratings matrix
User-Based CF Algorithm
Voting : vi,j corresponding to the vote for user i on item j.
Mean Vote :
where Ii is the set of items on which user i voted.
Predicted vote:
weights of n similar users normalizer
Similarity Computation
Vector Cosine-Based Similarity
Correlation-Based Similarity (Pearson)
Other Similarities
Vector Cosine-Based Similarity
Vector cosine similarity:
Uu ujuUu uiu
Uu ujuuiu
BA
rrrr
rrrrw
2
,
2
,
,,
,
)()(
))((
Adjusted cosine similarity:
different rating scale?
Correlation-Based Similarity
Pearson correlation:
Thus in the example in Table 2, we have w1,5 = 0.756.
Prediction Computation
Weighted Sum of Others’ Ratings:
For the simple example in Table 4, using the user-based CF algorithm, to
predict the rating for U1 on I2, we have
Recommendations I
Rating Prediction Algorithm:
a) Calculate Pa,i for each item i with prediction
computation formulation.
b) Recommend the top-N highest rating items
that the active user a has not purchased.
Recommendations II
K Nearest Neighbors Algorithm:
a) Find k most similar users (KNN).
b) Identify a set of items, C, purchased by the
group together with their frequency.
c) Recommend the top-N most frequent items in
C that the active user has not purchased.
Item-Based CF Algorithm
Correlation-Based Similarity:
where ru,i is the rating of user u on item i, is the average rating of the ith item by
those users.
User-Item
Matrix
ir
Prediction Computation
Simple Weighted Average:
where wi,n is the weight between items i and n, ru,n is the rating for
user u on item n.
Extensions
• Default Voting
• Inverse User Frequency
• Case Amplification
Default Voting
Problem:
• pair-wise similarity is computed only from the ratings in
the intersection of the items both users have rated.
• too few votes at the beginning
Solution: Assuming some default voting values for the missing
ratings can improve the CF prediction performance.
Dimension Reduction, such as SVD, PCA etc.
Inverse User Frequency
Definition:
)/log( ji nnf
where nj is the number of users who have rated item j and
n is the total number of users.
Case Amplification
where ρ is the case amplification power, ρ ≥ 1, and
typical choice of ρ is 2.5. Case amplification reduces
noise in the data.
It tends to favor high weights as small values raised to a
power become negligible.
For example, wi,j = 0.9, then it remains high (0.92.5 ≈ 0.8);
if wi,j = 0.1, then it be negligible (0.12.5 ≈ 0.003).
Model-Based CF Algorithm
• Simple Bayesian CF Algorithm
• Clustering CF Algorithm
Simple Bayesian CF Algorithm
Simple Bayesian:
Laplace Estimator:
Simple Bayesian CF Algorithm
Example in Table 4, to produce the rating for U1 on I2 using the
Simple Bayesian CF algorithm and the Laplace Estimator:
Clustering CF Algorithm
For two data objects, X = (x1, x2, …, xn) and Y = (y1,
y2, …, yn), the popular Minkowski distance is defined as,
where n is the dimension number of the object, and q is a positive integer.
Obviously, when q = 1, d is Manhattan distance; when
q = 2, d is Euclidian distance.
Evaluation Metrics
Mean Absolute Error and Normalized Mean Absolute Error:
where rmax and rmin are the upper and lower bounds of the ratings.
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
Challenges
• Data sparsity
• Scalability
• Synonymy
• Gray Sheep
• Shilling Attacks
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
Demo
• Tools:Mahout - Scalable machine learning and data
mining library,http://mahout.apache.org/
• Data: MovieLens, http://www.movielens.org/
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
Conclusions
CF categories Memory-based CF
Representative techniques Item-based/user-based top-N
recommendations
Main advantages 1. easy implementation
2. new data can be added easily and
incrementally
3. need not consider the content of the
items being recommended
4. scale well with co-rated items
Main shortcomings 1. are dependent on human ratings
2. performance decrease when data
are sparse
3. cannot recommend for new users
and items
4. have limited scalability for large
Conclusions
CF categories Model-based CF
Representative techniques 1. Bayesian belief nets CF
2. Clustering CF
3. CF using dimensionality reduction
techniques, SVD, PCA
Main advantages 1. better address the sparsity,
scalability and other problems
2. improve prediction performance
3. give an intuitive rationale for
recommendations
Main shortcomings 1. expensive model-building
2. trade-off between prediction
performance and scalability
3. lose useful information for
dimensionality reduction techniques
Outline
• Introduction
• Collaborative Filtering Algorithm
• Challenges
• Experiments (demo)
• Summary
• Future work
Future work
Scalability Real-time
Q & A
References
J. Breese, D. Heckerman, and C. Kadie, “Empirical analysis of predictive
algorithms for collaborative filtering,” in Proceedings of the 4th
Conference on Uncertainty in Artificial Intelligence (UAI ’98), 1998.
B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Item-based collaborative
filtering recommendation algorithms,” in Proc. of the WWW Conference,
2001.
K. Miyahara and M. J. Pazzani, “Collaborative filtering with the simple
Bayesian classifier,” in Proceedings of the 6th Pacific Rim International
Conference on Artificial Intelligence, pp. 679–689, 2000.
L. H. Ungar and D. P. Foster, “Clustering methods for collaborative
filtering,” in Proceedings of the Workshop on Recommendation Systems,
AAAI Press, 1998.
Xiaoyuan Su and Taghi M. Khoshgoftaar, “A Survey of Collaborative
Filtering Techniques,” in Advances in Artificial Intelligence Volume 2009,
Article ID 421425, 19 pages.