With the birth of XGBoost, additive tree models have been widely applied in industries (such as LightGBM, CatBoost) due to their state-of-the-art performance and well-engineered parallel acceleration computing. However the one general rule in machine learning field is more complex models that fit well to data are less interpretable. This article will talk about different approaches that let a model “speaks” and explains how the model generate outputs.

While deep learning models have been proved to be powerful to apply in industrial applications, they certainly cost us a lot more time to make them right compared to developing machine learning pipelines, especially when the data is a lot and complex, or model scale is large, we often encounter issue of underfitting or loss not converging. The article will talk a little bit my experience of how to solve such problem.

Embedding is a dense representation of tokens in the form of numeric vectors. The concept derived from word embeddings which can be learned using a variety of language models. Such technique is well-known, but beyond NLP problems embedding can be implemented to represent other instances and entities. I’ll briefly discuss such application through my experiences in healthcare industry.