*** 1

协同过滤中两个问题探讨
Related papers :
1. Modeling Evolutionary Behaviors for Community-based Dynamic
Recommendation
--SIAM
2. Learning Preferences of New Users in Recommender System : An
Information Theoretic Approach
--SIGKDD
演讲人:余海洋
2011:12:10
Will user U like item X?
1. Content-based filtering
2. Collaborative filtering(协同过滤)
3.Hybrid approaches
Cold start
Dynamic Recommendation
Cold Start
How to choose items?
1.
2.
3.
4.
Popularity
Entropy
Entropy0
HELF : harmonic mean of Entropy
and logarithm of Frequency
5. IGCN : Information Gain through
Clustered Neighbors
1. Popularity
Indicates how frequently users rated the item
1. Uninformative
2. Prefix bias
2. Entropy
Represents the dispersion of opinion of users on the item
1. It often selects very obscure
items
2. Suggests “garbage” patterns to be useful
as well . Do not considers the frequency .
3. Entropy0
1. Bias frequently-rated too
much
4. HELF
Harmonic mean of Entropy and logarithm of Frequency
4. HELF
Harmonic mean of Entropy and logarithm of Frequency
5. IGCN
Information Gain through Clustered Neighbors
IGCN works by repeatedly computing
information gain of items where the
necessary rating data is considered only
from those users who match best with the
target users profile so far
5. IGCN
Information Gain through Clustered Neighbors
1. Missing values
2. Inadequate goals
5. 1 Missing values
5. 2 Inadequate goals
Find the final best k neighbors
5. IGCN
Information Gain through Clustered Neighbors
Offline Experiments
Offline Experiments
Dynamic Recommendation
1. 用户兴趣的变化
2. 物品流行度的变化
3. 季节效应
CBDB
Community-based Dynamic
Recommendation
CBDB
Formal community
construction
People from the same department of same
company tend to have similar interests ,
since they tend to have similar background
and are working on similar projects.
Dynamic pattern analysis
1. Freshness of documents
2. Short-term of long-term type of
document
3. Popularity of documents
4. User intention
Short-term of long-term type
of document
Short-term of long-term type
of document
CBDB
Adaptive community model
Time-Sensitive Adaboost
Time-Sensitive Adaboost
Experimental Results
Thank You