Crowdsourcing Data Science for Innovation

In recent years, companies and non-profit organizations are increasingly leveraging crowdsourcing platforms to seek novel solutions to data-driven problems. Crowdsourcing innovation problems to the data science community frequently take the form of public contests. Existing literature studied the mechanism and design of open innovation contests. However, the data science contests have a number of unique characteristics that drive the need for extending existing literature. For example, a contributor is allowed to make multiple submissions, each submission takes considerable efforts, and participation in a contest requires some expertise. Moreover, the seeker can utilize multiple solutions to gain higher utility. In this paper, we identify the optimal strategies from the contributor's perspective and the solution seeker's perspective. We find that contributors with more than minimum expertise will participate in the data science contests and contributors with higher expertise will make more submissions. For the seeker, there exists an optimal award amount to obtain the highest utility. Meanwhile, compared with "winner-takes-all", using the ensemble method with "Top K" award structure could generate higher utility. Empirical evidence from a large real-world dataset supported theories derived from our model.