1343edb2739a61a6e20bd8764e814b50-Supplemental-Conference.pdf

We summarized the conclusion for weight-moving influenceintotheTheorem2. As demonstrated in the previous works [4, 41], the stochastic method is efficient and relatively accurate for sample-wise influence estimation. Step 2. For i {1,2,...,J}, recursively compute the inverse HVP estimation using a batch size B of randomly sampled a data point (xi For real-world dataset, estimating each sample separately using the LiSSA method is intolerable. Table 7: Spearman's ranking correlation coefficient ( 1) between actual effect of each LF and that estimated from source-aware IF. We highlight results that do not pass the significance test in underline.Dataset Census Mushroom PWSpambase IMDb Yelp Youtube DN-real DN-sketch DN-quickdraw DN-painting DN-infograph DN-clipart Avg.