Smoothing is a method of removing the rough (the error or noise component in data) and retaining the smooth (the predictable component in data) by averaging within neighborhoods of similar data values. Its utility is self-evident: No data analyst wants to model noise, producing a model that yields unreliable (large error variance) and inaccurate (large prediction bias) results. The concept behind smoothing, and CHAID as a data smoother is  in my book (Ratner, B., Statistical Modeling and Analysis for Database Marketing: Effective Techniques for Mining Big Data, pages 74 – 84).