We have demonstrated how to build an ad click predictor that learns from massive click logs using Spark. Thus far, we have been using one-hot encoding to employ the categorical inputs. In this section, we will be talking about two popular feature engineering techniques: feature hashing and feature interaction. One is an alternative to one-hot encoding, another is a variant of one-hot encoding. Feature engineering means generating new features based on domain knowledge or defined rules, in order to improve learning performance achieved with existing feature space.