◇◇新语丝(www.xys.org)(xys4.dxiong.com)(www.xysforum.org)(xys2.dropin.org)◇◇ 我对周志华教授被指控剽窃的看法 作者:Huimin Chen http://hchen2.spaces.live.com/blog/cns!76203919D10BFD99!3024.entry My take on Prof. Zhi-hua Zhou's alleged 剽窃 From xys, I noticed an ongoing debate on whether Prof. Zhou borrowed key ideas from others in his award winning conf. paper and subsequent publications related to ensemble learning under the term "many could be better than all" for combining weak learners. Some anonymous posters @ xys found an earlier paper with quite similar formulation of the ensemble method and the MSE analysis for combining multiple regressors. Obviously, Prof. Zhou and his coauthors did not cite this work in their subsequent publications. I have adequate background to understand ensemble learning and did a survey on estimation and decision fusion with several colleagues a few years ago including various methods for combining classifiers/regressors with guaranteed performance improvement. My take on Prof. Zhou's case is as follows. 1. Prof. Zhou and his coauthors were aware of the work by Perrone and Cooper and it is very dishonest to ignore its existence in their literature review. The extension from optimal weighting on the full set of weak learners to the best subset invoking genetic algorithm for this combinatorial optimization problem is trivial. The conf. paper probably should not get any award and the journal version does not deserve to be published. 2. It is hard to tell whether Prof. Zhou intentionally copied some materials from Perrone and Cooper's paper as being accused. However, Prof. Zhou's response took too much credit for the contributions to ensemble learning area for his "many could be better than all" argument. For those familiar with Bayesian model averaging, the generalization error analysis based on the same assumed statistical model can be derived without reading Perrone and Cooper's paper, so even if Prof. Zhou independently came up with the same conclusions as those by Perrone and Cooper, I would not be surprised. Even Perrone and Cooper's paper is not that influencial in ensemble learning area. 3. It seems that Prof. Zhou later became famous in machine learning area (maybe only within China) and he does not want to tolerate any different opinion regarding the quality of his publications. From his writing, I can hardly tell that he is a decent and modest researcher. Even after many years and with the title of Cheung Kong professor, he fails to prove himeself in any significant work along ensemble learning or other machine learning areas. So I have to give benefit of doubt that his achievement has been exaggerated. (XYS20091106) ◇◇新语丝(www.xys.org)(xys4.dxiong.com)(www.xysforum.org)(xys2.dropin.org)◇◇