Competition-aware task routing for contest based crowdsourced software development

Abstract

In crowdsourced software development, routing a task to right developers is a critical issue that largely affects the productivity and quality of software. In particular, crowdsourced software development platforms (e.g. Topcoder and Kaggle) usually adopt the competition-based crowdsourcing model. Given an incoming task, most of existing efforts focus on using the historical data to learn the probability that a developer may take the task and recommending developers accordingly. However, existing work ignores the locality characteristics of the developer-task dataset and the competition among developers. In this work, we propose a novel recommendation approach for task routing in competitive crowdsourced software development. First, we cluster tasks on the basis of content similarity. Second, for a given task, with the most similar task cluster, we utilize machine learning based classification to recommend a list of candidate developers. Third, we consider the competitive relationship among developers and re-rank the candidates by incorporating the competition network among them. Experiments conducted on 3 datasets (totally 7,481 tasks) crawled from Topcoder show that our approach delivers promising recommendation accuracy and outperforms the two comparing methods by 5.5% and 25.4% on average respectively.