报 告 人：何向南 教授
Research on machine learning (ML) has been intensively focused on model training, paying relatively little attention on the stage of model inference. When serving online, the same model evaluation procedure, e.g., the forward propagation of a neural network, is executed for the testing samples. However, the model may fail on hard samples such as adversarial samples and out-of-distribution samples. Thinking what human beings will do when facing hard examples --- they may hesitate, scrutinize, perform deep reasoning such as counterfactual thinking to make the final decision. Nevertheless, the inference stage of ML does little such causal reasoning. This talk introduces a new schema of ML inference that revises the prediction through conducting counterfactual inference. In particular, the schema is instantiated over recommender models, showing advantages in eliminating popularity biases and clickbait biases. Apart from user-item relations, the schema is also applied to graph neural network (GNN) that processes general graph data, enhancing the model in a wide spectral of graph analytic applications.
Dr. Xiangnan He is a professor at the University of Science and Technology of China (USTC), leading the USTC Lab for Data Science. His research interests span information retrieval, data mining, machine learning and multi-media analytics. He is in the Editorial Board of the AI Open journal, and has over 80 publications that appeared in several top conferences such as SIGIR, WWW, KDD, and MM, and journals including TKDE and TOIS. His Google Scholar citations is over 7000, and his work on recommender systems has received the Best Paper Award Honorable Mention in WWW 2018 and ACM SIGIR 2016. Moreover, he has served as the PC chair of CCIS 2019, the area chair of MM 2019-2020, ECML-PKDD 2020, and the (senior) PC member for several top conferences including SIGIR, WWW, KDD, IJCAI etc.