A Modified PPM Algorithm for Online Sequence Prediction Using Short Data Records

Abstract

Discrete sequence prediction using source encoding techniques, generally involves two steps—(a) building frequency trees and (b) computing distributions using frequency trees to perform prediction. The second step is usually performed by a technique called Prediction by Partial Match (PPM) and its variants. The implicit assumption in PPM is that using frequency trees of greater depth results in better prediction. In this paper, we question that assumption especially when one has access only to small sequence lengths, since extracting information from longer contexts typically involves estimating a higher number of parameters. We propose a modified PPM algorithm, where, the different context based predictors are weighed according to their prediction accuracy and prediction is performed based on a combined model. We finally apply the algorithms on a well-known location prediction data-set and prove the efficacy of the algorithm proposed by us and its utility in location prediction.