CLiMF.md

CLiMF: Collaborative Less-is-More Filtering

Overview

less-is-more: provide users with only few but valuable recommendations

Feature

• focused on improving top-k recommendation (making a few but relevant recommenddations)
• is tailored to recommendation domains where only binary relevance data is available

Approach

• Directly maximizing the Mean Reciprocal Rank (MRR)

Dataset

• Epinions
• Tuenti

Reciprocal Rank (RR)

In this paper, they introduce a way to significantly reducing the computational complexity of RR optimization

CLiMF Algorithm

Smoothing the RR

Lower Bound of Smooth RR

objective funciton of CLiMF

$$ F(U, V) = \sum^M_{i=1}\sum^N_{j=1}Y_{ij}[\ln g(U_i^T V_j) + \sum^N_{k=1}\ln(1-Y_{ik}g(U_i^T V_k - U_i^T V_j))] - \frac{\lambda}{2}(||U||^2 + ||V||^2) $$

Optimization

Use stochastic gradient ascent to maximize the deriviatives of objective function with respect to $U_i$ and $V_j$

Resources

Source Code

• gamboviol/climf
• coreylynch/pyCLiMF