Elsevier

Expert Systems with Applications

Volume 83, 15 October 2017, Pages 326-332
Expert Systems with Applications

Early churn prediction with personalized targeting in mobile social games

https://doi.org/10.1016/j.eswa.2017.04.056Get rights and content

Highlights

  • A model for predicting churn one day after registration in games is proposed.

  • The model is generalized and applicable to various types of free-to-play games.

  • Two approaches for retaining churned users via push notifications are presented.

  • The proposed system is deployed and used in the production of a large-scale game.

Abstract

Customer churn is a widely known term in many industries, including banking, telecommunications and gaming. By definition, churn represents the act of a customer leaving a product for good. Most commonly, late customer churn is addressed. In the dynamics of free to play games, most of newly registered users abandon the game in the first few days, so the main focus is on early customer churn. Therefore, successful early churn prevention methodology is vital to having a successful business in free to play gaming industry. To tackle this problem, we introduce a two stage intelligent system. It employs early churn prediction, formulated as a binary classification task, followed by a churn prevention technique using personalized push notifications. For early churn prediction, common machine learning models are trained and compared using a data set obtained from two million players of Top Eleven - Be A Football Manager online mobile game. To prevent churn, we track user activity, identify the game features that are potentially interesting to the user and then use that data to tailor personalized push notifications with a purpose to attract users back into the game. Using this approach, we are able to reduce churn up to 28%, which, at the scale of millions of users, represents a significant positive impact to business.

Introduction

Social games are type of online games that use the potential of social networks for the purposes of obtaining new users - players, retaining current ones and creating meaningful social interactions among them. The social gaming industry has gained an edge in recent years and it is now arguably the biggest global mobile market, taking 85% of all mobile market app revenue in 2015. By 2020 games will reach the estimated revenue of $74.6 billion (Takahashi, 2016).

For the biggest part of their history, games were produced and thereby treated as any other product: they were initially bought and played to a certain extent. The introduction and development of a freemium model of games changed numerous things within the inner logic of the game and in monetization strategies. The key distinction of a freemium game model is that the freemium game is not bought by paying an initial price. It is usually downloaded for free, but there are optional microtransactions inside the game. This new type of approach to users (players or consumers) generates new problems for developers in the revenue strategies. For example: after the game is downloaded, how do you retain players, and how do you keep them engaged in such a way that it leads to increased revenue?

Naturally, these new types of problems have led to an extension of focus of game developers in two directions: one behavioral, and other technical/data driven. The former implies making more fun and addictive game based on user needs, while the latter relies on detecting and solving impediments playing the game, as well as recognizing when certain users need an additional incentive in order to continue playing. This paper follows the second direction, presenting a methodology/pipeline for retaining users longer in the game.

Any methodology that aims at preventing users from leaving the game, should consist of two natural parts. The first challenge is to detect which users are likely to leave the game in a short time period. Using that information, the second part should focus on finding a way to prevent those users from leaving the game. The methodology presented in this paper follows the same reasoning - it employs a churn detection technique, followed by a personalized churn prevention approach.

Churn is usually defined as the act of a player leaving the game permanently, while churn prediction represents a problem of identifying users who are likely to churn. In order to successfully apply any churn prevention method, one of the most important prerequisites is the ability to predict early which users have a tendency to churn. It is common for the game industry to expect that for the game to be successful, at least 40% of all registered users should return to the game the following day, as is written in Agell (2015). According to the industry report from the deltaDNA, real market numbers are lower, ranging from 24 to 30% for different genres (DNA, 2015). That means that more than 70% of all newly acquired users would play the game for only one day. In a freemium game, this is of crucial importance because the price of acquiring a new user is far greater than retaining an already registered one.

Realizing this, identifying churners grants the opportunity to approach them differently. As gamers’ willingness to play and freely pay is the source of a game’s monetization, churn prediction represents the opportunity to increase revenue by developing well-informed strategies based on behavioral data regarding the churners’ activities in the game. And this finally leads to behavioral considerations: the focus has shifted towards the maximum possible improvement of players’ first overall experience. By constructing a stimulating environment organized in meaningfully packed set of choices, developers improve the possibility to increase a player’s life in the game. And this completes the circle, making behavioral and technical ventures act jointly in the prospect of defining monetization strategies.

In this paper, we propose the following methodology for dealing with churn prevention, that consists of two major components:

  • 1.

    churn prediction algorithm based on binary classification using behavioral data from the first day of lifetime

  • 2.

    churn prevention system based on personalized push notifications

Running these two components very early on, after only one day of user’s activity, represents a rather challenging technical problem, due to extremely limited amount of data available. To illustrate, most of the users will play only a couple of minutes, thus the data produced from their activity has very limited predictive power. In this type of setting, it is of crucial importance to leverage every piece of available information, challenging the conventional approaches to classification and targeting.

The contributions of our work are two-fold: scientific and business. From scientific perspective, the main contributions are the following: to the best of our knowledge, our approach is the first one that (1) deals with early churn, by using only one day of users’ activity data, (2) explores the possibility of retaining users with custom tailored deep-linked notifications.

From business perspective, the contribution of this paper lies in both the improved player experience and increased profits. Used on a mature game, the methodology presented in this paper is likely to lead to increased profits, potentially making a difference between a game that generates profits or losses.

The structure of the paper is as follows. In Section 2, some of the existing work regarding similar topics is discussed. In Section 3, we describe the methodology for both churn prediction and push notification targeting. Section 4 presents the experiment and its results. For the evaluation, we check which fraction of users that were predicted to churn actually returned to the game after seeing the push notification. Section 5 discusses results and the impact that the proposed approach had to the game metrics. A simple profit quantification method is also presented. In Section 6, we discuss potential future work and possible improvements. Section 7 concludes the paper.

Section snippets

Related work

As the methodology presented in this paper contains two components, we will also divide the related work section into two parts, prediction and prevention.

Churn prediction is a common problem, not only in the gaming industry but also in many other industries as well. Dealing with churn prediction was addressed in the telecommunication industry (Dierkes, Bichler, & Krishnan, Huang, Kechadi, Buckley, 2012, Huang, Kechadi, 2013, Idris, Rizwan, Khan, 2012, Keramati, Jafari-Marandi, Aliannejadi,

Methodology

In the following section, we describe the methodology we used, including the churn prediction model, notification targeting strategy, as well as the methods used to evaluate the performance of the whole process.

Dataset

The work presented in this paper builds on the data collected in 2015 from 2,000,000 randomly selected players of the mobile free-to-play online social football manager game Top Eleven, developed and published by Nordeus. Top Eleven is an online multi-platform football manager video game. It covers all aspects of managing a team, from creating a squad to building a stadium and surrounding facilities. Users compete in various competitions and take on the challenges of an everyday football

Discussion and conclusion

We have shown that early one-day churn prediction is possible to a certain extent with decent performance. By using more complex and game specific features, a better classification performance could possibly be achieved. However, with more user activity oriented features, we obtained generality, which means that the model can be transferred to almost any other online social game with minimal effort. It is interesting to note that only behavioral data was used - no demographic data was included

Future work

The procedure described in this paper can be further developed. Other models, such as Neural Networks and Deep Neural Networks, could be used. Also, follow up notifications could be exploited. In other words, users could be notified on more than one consecutive day according to both how they act in their first session and how they responded to previous notifications. An additional psychological analysis could be introduced when engineering messages for notifications in order to achieve maximum

7.  Acknowledgments

We would like to thank Dejan Prokić for his role in implementing the notification system infrastructure. We would also like to thank our colleagues Milovan Dekić, Marko Jevremović, Benoit Detalle and Marko Knežević for their valuable suggestions.

References (36)

Cited by (63)

  • Predicting customer abandonment in recurrent neural networks using short-term memory

    2024, Journal of Open Innovation: Technology, Market, and Complexity
  • Customer churn prediction for web browsers

    2022, Expert Systems with Applications
    Citation Excerpt :

    Churn prediction is one of the most crucial stages in customer relationship management (CRM) (Ahn, Hwang, Kim, Choi, & Kang, 2020; Arivazhagan & Sankara, 2020) hence developing an accurate and effective client churn prediction model is key for customer retention. In recent decades, extensive researches have been conducted on this issue using static features, especially in industries with large numbers of customers, such as the telecommunication (Amin et al., 2019; Lu, Lin, Lu, & Zhang, 2012), banking (Bilal Zorić, 2016; Keramati, Ghaneei, & Mirmohammadi, 2016), insurance (Dela Llave, López, & Angulo, 2019; Sundarkumar & Ravi, 2015), social networking (Backiel, Verbinnen, Baesens, & Claeskens, 2015; Farquad, Ravi, & Raju, 2014; Óskarsdóttir et al., 2017), credit card (Keramati et al., 2016; Rajamohamed & Manokaran, 2018), and online game industries (Milošević, Živić, & Andjelković, 2017; Periáñez, Saas, Guitart, & Magne, 2016; Tamassia et al., 2016). In addition, many researchers build churn prediction models based on the dynamic feature.

  • Churn prediction in digital game-based learning using data mining techniques: Logistic regression, decision tree, and random forest

    2022, Applied Soft Computing
    Citation Excerpt :

    Because DGBL is a combination of these industries and there is no research in DGBL churn prediction, the common algorithms from these two industries were chosen to apply for the DGBL dataset. These algorithms are logistic regression [7,21,22,25,33,34], decision tree [21,22,25,31,34], and random forest [7,25,33,34]. Logistic regression (LR) is a probabilistic algorithm for binary classification [25].

View all citing articles on Scopus
View full text