Abstract:
Predicting product purchases is one of the important tasks of a broker in the barter
trade exchange. This work introduces the utilization of aspect models - a latent class
statistical mixture model used for soft-clustering of co-occurrence data - for generating
future purchase predictions for existing and new members in a barter exchange
from transaction data. Three aspect models are investigated. Expectation Maximization
(EM) algorithm and Annealed Expectation Maximization algorithm are used to fit
the models with the data. A system is implemented to train and evaluate the performance
of the proposed models and algorithms. Several experiments are carried out to
determine the optimal number of states for aspects for each model and to determine
which model performs better. The experimental results show that aspect models work
well in predicting product purchases from transaction data.