Latent Inducing Sparse Gaussian Processes Regression followed by Social Collaborative filtering

Van Trung Nguyen followed by Suvash Sedhain (NICTA)

NICTA SML SEMINAR

DATE: 2013-06-06
TIME: 11:15:00 - 12:15:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
Latent Inducing Sparse Gaussian Processes Regression: Gaussian process (GP) has been widely used in various Bayesian models in machine learning, particularly for non-linear regression thanks to its simple analytical solution. Previous methods have been proposed to bring down the complexity of GP from O(N^3) to O(NB^2) for computation and O(N^2) to O(NB) for memory, where N is the problem size and B < N. However, this cost is still prohibitive for very large datasets. We present a new mixture of GP experts and an efficient inference procedure that has O(NM^2) cost in computation and O(NM/K) in memory, where K is the number of experts and M = B / K << N. Each expert is a GP with augmented inputs which induce sparsity. The experts are independent and specialize on localized regions of the data where its inducing inputs lie. The model achieves predictive performance comparable with sparse GP approximation while taking much less time to train. Further experiments with a dataset of size 100000 show favourable results compared to other baselines, and only takes 9 hours to train on a standalone computer.

Social Collaborative filtering: Content recommendation in social networks poses the complex problem of learning user preferences from a rich and complex set of interactions (e.g., likes, comments and tags for posts, photos and videos) and activities (e.g., favourites, group memberships, interests).While many social collaborative filtering approaches learn from aggregate statistics over this social information, we propose a different approach: we first define social affinity groups (SAGs) of a target user by analysing their fine-grained interactions (e.g., users who have been tagged in the target useras video) and activities (e.g., users who have joined the same special interest group that the target user has joined). Then we learn which SAGs are most predictive of the target useras preferences in a method we term social affinity filtering (SAF). Furthermore, we explore the idea of social affinity filtering under one-class collaborative filtering setting. We compare and contrast the rank list generated by standard item-item neighbourhood models and social affinity based neighbourhood models.