STACKING ENSEMBLE STRATEGY FOR CLICK THROUGH RATE PREDICTION

Abstract

In online advertising and recommender systems, CTR prediction, which seeks to forecast the likelihood that a user will click an item, is critical. The two most prominent strategies for CTR prediction are feature interaction modelling and user interest mining. Both have been extensively investigated for many years and have made significant progress. The main focus of this study will be on examining user interest mining-based models using sequence-based CTR models. In current days, most sequential CTR models employ only recent user behaviour and do not take into account a user's past historical data. Directly using historical data can make model training activities complex and time-consuming. The model that uses historical data cannot do so directly and must rely on the user behaviour extraction module to extract the most important user interactions. To address the aforementioned issue, a stacking ensemble based CTR model is suggested, which stacks or combines the prediction results for each of the long sequence and short sequence based CTR models and uses a meta-learner to train over the forecasted data to obtain the final prediction. Tmall, a real-world industrial dataset, is used to conduct extensive experiments. The experimental results reveal that our suggested strategy outperforms current CTR prediction models significantly.