Knowledge-reinforced Automatic Machine Learning

Various machine learning (ML) models have been proposed during the past decades to augment existing automation and human-centric systems by providing superior prediction performance. However, due to the finite support and inductive bias of existing models, developing a one-size-fits-all model that can be used for any ML task has been a challenge. For each specific task (i.e., a dataset), current practice requires a data scientist to manually search for an appropriate ML pipeline (i.e., a sequence of ML method options). In such a search process, unique insights about data distributions and the support and bias of models are usually generated to achieve satisfactory performance. Though being valuable to guide other ML tasks, understanding, summary, and utilization of such searching process knowledge is lacking. As a result, given a new ML task, explicit pathways for the model-searching process are still unknown. Even the state-of-the-art automatic machine learning (AutoML) methods need to explore a large model space with guidance from greedy optimizers to learn quantitative knowledge from trials, which is often computationally prohibitive (e.g., model searching time varies from days to weeks). Therefore, we propose a reinforcement learning method to assist knowledge-reinforced AutoML (KR-AutoML) system to effectively and efficiently identify satisfactory ML pipeline(s) for a task. Such reinforcement learning model naturally learn from the sequential decision-making process to optimize the stepwise selection of ML pipelines. The proposed KR-AutoML is evaluated on several real datasets.