MCW: A new weighting method for linear combination of regressors in WSNs
Hadi Shakibian, Nasrollah Moghadam Charkari · 2010
The ultimate goal in a multiple classifier system (MCS) is to obtain a global and more accurate model through the combination of several base learners. Among the popular combining rules, averaging has been emphasized as a well qualified option. The averaging rule can be applied with equal (simple averaging) or non-equal (weighted averaging) weights vector for the linear combination. When the formed ensemble includes unbalanced base learners, the efficiency of simple method might undesirably be decreased. On the other hand, conventional weighting methods usually weigh each learner according to only a single criterion leading to unfair combination. So, when each ensemble's member is learned based on different metrics/criteria, a more efficient weighting method will be needed to fairly assign the weight of each learner. Accordingly a new multi criteria based weighting (MCW) method based on multi criteria decision making concept (MCDM) has been proposed in this paper. Particularly, distributed regression in wireless sensor networks (WSNs) is addressed in which different regressors can be achieved from network's clusters (one for which), and the final network's model can be obtained through applying the weighted averaging rule on the clusters' regressors by the fusion center. The results show when the regressors are weighted according to different criteria by the proposed method, a better prediction accuracy can be reached in comparison with the simple and variance-based weighting methods. The influence of the number of the clusters/regressors has also been analyzed which indicate the proposed method behaves more stable than two other rules.