1:- Latent feature models for dyadic prediction 2:- Probability estimation, ranking and friends
Aditya Menon (Computer Science and Engineering University of California)
NICTA SML SEMINARDATE: 2013-07-04
TIME: 11:15:00 - 12:15:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
1: - Dyadic prediction is the task of predicting a label for the interaction of a pair of entities. For example, we may wish to predict the rating a user gives to a movie (collaborative filtering), the click through rate of an ad on a webpage (response prediction), or the social relationship between a pair of individuals (link prediction). These and other instances of dyadic prediction have been modelled with different approaches, such as matrix factorization, a reduction to supervised learning, and graph-theoretic scoring. This talk presents a general model for dyadic prediction tasks based on the log-linear framework. The model learns latent features that represent implicit characteristics of entities, and combines these with explicit features for dyad members and/or their interactions, if available. We study the application of this model to collaborative filtering and response prediction, in both cases showing its adaptability to problem-specific challenges, and performance competitive with existing methods for these tasks.
2:- In this talk, I will present my perspective on some nascent interests of mine, namely (primarily binary) probability estimation and (primarily bipartite) ranking, and the related learning problems of imbalanced and cost-sensitive learning. I will discuss some of my work on these problems, but the focus will be on open questions, with the hope of fostering collaboration.
BIO:
http://cseweb.ucsd.edu/~akmenon/
