1. Response Aware Model-Based Collaborative Filtering
Motivation
Rating data contain explicit scores and implicit response patterns
The rated items are not randomly selected
Goal: How to integrate users’ response patterns into a successful matrix factorization, Probabilistic Matrix Factorization, to avoid bias parameter estimation
Experiments
Three recommender protocols
Synthetic and Yahoo! datasets
RAPMF performs better on
randomly selected items
Items
Items 5 4 3 2 1 1
Users
Users
Rating matrix X
Response matrix R
Data model: probabilistic matrix factorization (θ=(U, V))
Response model: variants of soft assignment
of Bernoulli distribution with parameters μ

2. A Unified Framework for Reputation Estimation in Online Rating Systems
Motivation
Online Rating Systems are easily manipulated by malicious users for various reasons.
To promote sales of certain items
To build up personal image

3. A Unified Framework for Reputation Estimation in Online Rating Systems
Motivation
Online Rating Systems are easily manipulated by malicious users for various reasons.
To promote sales of certain items
To build up personal image
We propose a user reputation estimation framework
Three ingredients of the framework
Prediction Model ℋ
Given a CF model ℋ that can predict entries of 𝑅, assume
𝑟 𝑖𝑗 ~𝒩 ℋ 𝑖,𝑗 , 𝜎 2
The log-likelihood of observing 𝑟 𝑖𝑗 is
ℒ 𝑖𝑗 =𝐶− 1 2 𝜎 𝑟 𝑖𝑗 −ℋ 𝑖,𝑗 2
The unexpectedness is
𝑠 𝑖𝑗 = 𝑟 𝑖𝑗 −ℋ 𝑖,𝑗 2
Penalty Function
Summarize the unexpectedness { 𝑠 𝑖𝑗 } into 𝑠 𝑖 or 𝑠 𝑗
Link Function
Link the 𝑠 𝑖 or 𝑠 𝑗 to 𝑐 𝑖
Model
Prediction Model
Penalty Function
Link Function
Mizzaro's algorithm
ℋ 𝑖,𝑗 = 𝑖∈ 𝒰 𝑗 𝑐 𝑖 𝑟 𝑖𝑗 / 𝑖∈ 𝒰 𝑗 𝑐 𝑖
𝑠 𝑗 = 𝑖∈ 𝒰 𝑗 𝑐 𝑖
𝑐 𝑖 = 𝑗∈ ℐ 𝑖 𝑠 𝑗 (1− 𝑠 𝑖𝑗 / 𝑠 𝑚𝑎𝑥 𝑗∈ ℐ 𝑗 𝑠 𝑗
Laureti's algorithm
Same as above
𝑠 𝑖 = 1 ℐ 𝑖 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗
𝑐 𝑖 = 𝑠 𝑖 +𝜖 −𝛽
De Kerchove's algorithm
𝑐 𝑖 =1−𝑘× 𝑠 𝑖
Li's L1-AVG
ℋ 𝑖,𝑗 = 1 𝒰 𝑗 𝑖∈ 𝒰 𝑗 𝑟 𝑖𝑗 𝑐 𝑖
𝑠 𝑖 = 1 ℐ 𝑖 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗
𝑐 𝑖 =1−𝜆 𝑠 𝑖
Li's L2-AVG
𝑐 𝑖 =1− 𝜆 2 𝑠 𝑖
Li's L1-MAX
𝑠 𝑖 = max 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗
Li's L2-MAX
𝑠 𝑖 = max 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗
Li's L1-MIN
𝑠 𝑖 = min 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗
Li's L2-MIN
𝑠 𝑖 = min 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗

4. A Unified Framework for Reputation Estimation in Online Rating Systems
We propose a user reputation estimation framework
Previous methods assume an item has an intrinsic quality. Depending on the situation, taste view might be more suitable.
Personalized prediction model PMF is used, which assumes a low-rank Matrix Factorization structure
ℒ= 𝑖,𝑗,𝑟 ∈𝒬 𝑟− 𝑈 𝑖 𝑇 𝑉 𝑗 𝜆 𝑈 2 𝑈 + 𝜆 𝑉 2 𝑉
Penalty function is
𝑠 𝑖 = 1 ℐ 𝑖 𝑗∈ ℐ 𝑖 𝑠 𝑖𝑗
Link function is
𝑐 𝑖 =1 − 𝑠 𝑖
Intrinsic value view VS. Taste View