Efficient inference with accuracy-preserved integer encoding on tree-based model compiler

In the ever expanding world of machine learning, where advancements continually push the boundaries of what's possible, the efficiency of deploying tree-based ensembles is often a point of interest. Renowned for their interpretability and minimal training data requirements, models such as random forests and extremely random forests have been found as effective tools across many different industries. In this paper, we integrate FLInt, a promising accuracy-preserving technique for accelerating inference speed, into the established TL2cgen framework. Through experimental analysis conducted on both x86 and ARM devices, we empirically demonstrate that FLInt consistently outperforms the baseline across a range of maximum tree depths. Additionally, we examine an established optimization technique, Quantization. However, our findings reveal that Quantization yielded mediocre results, often resulting in slowed inference times across most scenarios.