Learning to predict closed questions on Stack Overflow

G. Lezina^ab, A. Kuznetsov^ab, P. Braslavski<sup>cb

a</sup> Ural Federal University, Ekaterinburg, Russia
^b SKB Kontur, Ekaterinburg, Russia
^c Ural Federal University, Ekaterinburg, Russia

Abstract: The paper deals with the problem of predicting whether the user's question will be closed by the moderator on Stack Overflow, a popular question answering service devoted to software programming. The task along with data and evaluation metrics was offered as an open machine learning competition on Kaggle platform. To solve this problem, we employed a wide range of classification features related to users, their interactions, and post content. Classification was carried out using several machine learning methods. According to the results of the experiment, the most important features are characteristics of the user and topical features of the question. The best results were obtained using Vowpal Wabbit – an implementation of online learning based on stochastic gradient descent. Our results are among the best ones in overall ranking, although they were obtained after the official competition was over.

Keywords: community question answering systems, large-scale classification, question classification.

UDC: 004.852

Received: 10.09.2013

Language: English