2023-01-012024-05-17https://scholars.lib.ntu.edu.tw/handle/123456789/674167Most real-world recommender systems affect their own behavior if they update over time. For example, in an advertising system, at any given time period only some Ads were recommended (i.e., presented) to users. For these displayed events we can observe if they are clicked or not. A model based on click/not-click information is then trained for recommending Ads to be displayed in the next time period. Clearly, the model is biased toward Ads that were displayed in the past. Counterfactual learning has recently emerged as an active area to study approaches for removing the bias. The goal of this sub-project is to develop techniques to impute labels of non-displayed events and efficiently train a model.recommender systems; counterfactual learning;recommender systems; counterfactual learning