Predicting Online Community Churners Using Gaussian Sequences

Abstract

Abstract. Knowing which users are likely to churn (i.e. leave) a service enables service providers to offer retention incentives for users to remain. To date, the prediction of churners has been largely performed through the examination of users' social network features; in order to see how churners and non-churners differ. In this paper we examine the social and lexical development of churners and non-churners and find that they exhibit visibly different signals over time. We present a prediction model that mines such development signals using Gaussian Sequences in the form of a joint probability model; under the assumption that the values of churners' and non-churners' social and lexical signals are normally distributed at a given time point. The evaluation of our approach, and its different permutations, demonstrates that we achieve significantly better performance than state of the art baselines for two of the datasets that we tested the approach on.

Mining User Development Signals for Online Community Churner Detection

“…We build on our other prior work [Rowe 2013] which presented an approach to model the lifecycles of users in online communities by examining different lifecycle fidelities, rather than the 20 stages inspected previously, and without the need to induce development trajectory functions. This paper also presents a more thorough, fine-grained, and deeper analysis of how churners and non-churners differ in their development; thereby expanding over our prior work [Rowe 2014;2013], and leading to the proposition of a theory for churner development. We hope that the framework and approach outlined in this paper, and the theory of churner development, will be tested in future work within the community on differing datasets and platforms.…”

“…), and then used this information to learn a decision tree classifier. In order to provide a comparison between our proposed detection models, and that of prior work, we use the approach from [Karnstedt et al 2011] as our baseline, along with a model from our own prior work based on a dual-gaussian sequence model [Rowe 2014]. We empirically demonstrate that our proposed linear detection model significantly outper-…”

“…-A detection model based upon linear models with elastic net regularisation trained using a stochastic coordinate descent learning approach -rather than average coordinate descent. -Evaluation of the detection model across the four platforms' datasets showing that we significantly outperform existing models including baselines from current work [Karnstedt et al 2011;Rowe 2014].…”

“…In this paper we present the results from examining both of these questions with an analysis of the development signals of churners and non-churners mined from four online community platforms, and a detection model that leverages features of these signals to identify who will churn and who will remain. In our prior work [Rowe 2014] we presented a detection model based upon Gaussian Sequences, in this paper we advance over this work by explaining how a linear model with elastic net regularisation offers significantly better detection accuracy (e.g. yielding ROC value of 0.706 in one instance) in certain cases, and examining the impact of smoothing on detection accuracy of the Gaussian Sequence model [Rowe 2014].…”

“…In our prior work [Rowe 2014] we presented a detection model based upon Gaussian Sequences, in this paper we advance over this work by explaining how a linear model with elastic net regularisation offers significantly better detection accuracy (e.g. yielding ROC value of 0.706 in one instance) in certain cases, and examining the impact of smoothing on detection accuracy of the Gaussian Sequence model [Rowe 2014]. We build on our other prior work [Rowe 2013] which presented an approach to model the lifecycles of users in online communities by examining different lifecycle fidelities, rather than the 20 stages inspected previously, and without the need to induce development trajectory functions.…”