AT&T Labs Research - Au, Siu-tong T.

This version of the work is reprinted here with permission of IEEE for your personal use. Not for redistribution. The definitive version was published in IEEE Computer Society's Conference Publishing Services (CPS) . , 2012-12-01

{In telecommunications service industry, a group of customers may be targeted for a set of marketing interests, and these interests are usually inter-correlated. For example, churn, upselling and appetency are often considered together, and decisions on how to retain customers, and to promote or to upgrade services are associated. Instead of predicting them separately as univariate models, we propose an iterative procedure to model multiple responses prediction into correlated multivariate predicting scheme. Our correlation factor derivations show that the exclusive case has more negative correlation factors, which is always favorable for responses separations in our multivariate prediction. We also point out that non-exclusive responses case can be reformed as another exclusive case via adding the overlapped positive response areas as new exclusive responses. This proposed method combines partial least squares (PLS) method and logistic regressions, in which the former is used to extract the mutual information from correlations, while the latter is utilized to refine every single response prediction through auxiliary information from PLS predictions. In other words, not only with the given predictor matrix, but the predicted probability information from other correlated responses are also inserted to help every single response prediction. This hybrid regression modeling is implemented iteratively to refine the prediction gradually. More importantly, before every round of iteration, all the positive predictions from different responses compete each other and the highest values are kept for the only positive prediction and the others are changed to negative. Here we exploit the positive exclusive property (i.e., positive for one response means the negative for others) between multivariate responses. Numerical results show that with the aid of mutual information from other responses and the positive exclusion adjustment, our proposed scheme can improve the conventional regression models significantly.}

The definitive version was published in The International MultiConference of Engineers and Computer Scientists 2011. , 2012-03-01

{This paper presents an expository study of temporal data mining for prediction of a future response sequence via mining large number of highly correlated concurrent time series. In the study, we investigate a two dimensional search scheme over time domain weighting and feature space selection. The weighting of observation records over time domain is used to exploit the time dependency structure and feature space selection is enforced to avoid the over-fitting issue. For a specific temporal and spatial selection, its area under ROC curve (AUC) is used to evaluate the prediction performance over the training and testing data. By varying the weighting scheme and feature selection, AUC contour maps on both training and testing data are generated. The contour maps can suggest us to apply the optimal allocations with highest AUC for future responses prediction in training, testing , and possible validation data. Numerical results over two sets of temporal data with different applications have shown that the proposed scheme can improve the prediction performance of conventional data mining methods significantly.}

AT&T Labs Research — Leading Invention, Driving Innovation

site map