Developing machine learning products better and faster at startups

Abstract

There are many uncertainties in developing machine-learning (ML)-based products. Due to the gap between research and development, the overall progress becomes slow, and experiences many failures and learnings only to see an initial idea not working or generating no significant revenue. To minimize these drastic effects, there are continuing studies to make a balanced handshake between research and development to shorten the time span of ML-based products from idea generation to deployment. This paper demonstrates a three-phase ML product development workflow at OneClass. The workflow is a combination of multiple best-practices in the innovation-based startup industry. The first phase of the workflow considers the pivotal idea generation for products that involves data reliability assessment, idea prioritization, expectation setting, and building trust among users. The second phase concentrates on several state-of-the-art strategies for planning and future re-use of several product components. Finally, the actual research and development phase describes the fail-fast method practiced by OneClass to learn quickly from failures and act accordingly. The workflow is followed by the company to develop many sophisticated ML-based products successfully within a very short period of time.