Aggregating contribution scores

[berkuva]

Hello,

Since DeepLift outputs different contribution scores for each input data, what is the suggested way to aggregate the DeepLift outputs? For example, with input neurons A and B, samples s1 and s2, DeepLift outputs contribution scores c11, c12 for s1 and c21, c22 for s2. One way to aggregate is to calculate the mean (contribution of A = (c11+c21)/2) and (contribution of B = (c12+c22)/2). Maybe median is another option. Or is it not plausible to aggregate DeepLift outputs for different samples? Thanks.