摘要： Bayesian Target Encoding is a feature engineering technique used to map categorical variables into numeric variables. The Bayesian framework requires only minimal updates as new data is acquired and is thus well-suited for online learning. Furthermore, the Bayesian approach makes choosing and interpreting hyperparameters intuitive. I developed this technique in the recent Avito Kaggle Competition, where my team and I took 14th place out of 1,917 teams. We found that the Bayesian target encoding outperforms the built-in categorical encoding provided by the LightGBM package.
Target-encoded variables are inherently leaky; that is, their construction requires information that we will not have when we make predictions.
By adopting a Bayesian framework, we can handle regularization in a systematic yet intuitive way. Furthermore, the Bayesian framework is well-suited for online learning requiring only minimal updates as new data is acquired. In the next section, we will discuss target encoding in the context of binary classification.
...
Full Text： mattmotoki.github
若喜歡本文，請關注我們的臉書 Please Like our Facebook Page：
留下你的回應
以訪客張貼回應