Trash Papers

Not the first time that I notice the existing of low-quality papers, but it seems to be the first time I have to study a paper of that low-quality in details.

In a coursework group project, our group needed to select a paper to build upon that. In fact, at the beginning, I thought this paper was decent, it experimented with quite similar idea that I have for expandable AI. Since the paper’s authors didn’t release the code, we had to re-implement it. We made a joke that maybe the authors made up their result as the evaluation tasks are all k-way classification but the accuracy is all 98%, thus they did’t release the code. Turns out, it seems to be the truth. Our experiment result was far worse than the result reported in the paper. You can doubt our code base, but we have checked that several times and the probability of wrong implementation is small.

Not just the result, the writing of the paper is also a shame! It repetitively described (easy) concepts, especially abused the only novelty of the paper - triplet data - to describe the usual training procedure AGAIN with triplet data.

However, the paper is accepted to KDD’22 conference. What a shame!