There has been a surge in the use of consumer grade wearable Electroencephalogram (EEG) devices for emotion discrimination tasks in various research laboratories in recent times. The obvious advantages carried by these compared to medical grade equipment are reduced costs and portability, which enable monitoring for a longer term and in more natural environment. Different manufacturers of consumer grade EEG devices place the electrodes at different locations. In this paper we present a novel method for determining locations of the fewest electrodes with the most emotion valence discriminative power. It starts with feature generation and selection for identifying positional features for the classification task, followed by channel selection that minimizes the feature reconstruction error. To evaluate the proposed methods, benchmarking analysis was done using leave out one subject cross validation with various machine learning models, using three public datasets. Results show with 8 electrodes AUC scores of 0.78, 0.8 and 0.67 are obtained for AMIGOS, DREAMER and DEAP datasets, respectively on emotion valence classification task. It is further observed that out of the best 8 channels selected, frontal (F8), parietal (P7), and temporal (T8 and T7) are common brain areas which are active during emotion processing across all the three datasets.