There is no absolute guidance on what influence numbers are harmful / helpful to the model to make a single prediction. Rather, those are in relation to one another meaning you're able to rank them.

I have the same question. I'm not sure whether in implementation you calculate I_{up,loss} or -1/n I_{up,loss} in original paper. And I use it in MINIST and find out that the avg contribution is very negative if the chosen train set and test set have the same number. I'm not sure whether something wrong occurs in my experiments. But hope you can check this part..

I have the same doubts as @QingXuTHU . It seems from

the points with negative I_{up, loss} (z, z_test) should be helpful, and those positive I_{up, loss} (z, z_test) should be harmful (refer to the title of Page 3, Figure 1 of the Koh and Liang paper)

I agree @chengrunyang. On page 6 of the paper, the authors also clarify that the most helpful images are those with the "most positive -I_{up, loss}" (or equivalently the most negative I_{up, loss}). Since np.argsort sorts the indices in ascending order, the harmful and helpful arrays should be switched in this code snippet:

I stumbled upon this issue while testing how often a training sample is its own most influential sample. Most of the time, the training sample came out as its own most harmful sample, which is counter-intuitive. This bug in the code explains that behavior.