=Paper= {{Paper |id=Vol-1627/paper8 |storemode=property |title=Churn Prediction for Game Industry Based on Cohort Classification Ensemble |pdfUrl=https://ceur-ws.org/Vol-1627/paper8.pdf |volume=Vol-1627 |authors=Evgenii Tsymbalov |dblpUrl=https://dblp.org/rec/conf/cla/Tsymbalov16 }} ==Churn Prediction for Game Industry Based on Cohort Classification Ensemble== https://ceur-ws.org/Vol-1627/paper8.pdf

Churn Prediction for Game Industry Based on
Cohort Classification Ensemble

Evgenii Tsymbalov1,2
1
National Research University Higher School of Economics, Moscow, Russia
2
Webgames, Moscow, Russia
etsymbalov@gmail.com
http://corpwebgames.com/en/

Abstract. In this paper, we present a cohort-based classification ap-
proach to the churn prediction for social on-line games. The original
metric is proposed and tested on real data showing a good increase in
revenue by churn preventing. The core of the approach contains such
components as tree-based ensemble classifiers and threshold optimiza-
tion by decision boundary.

Keywords: Churn prediction, ensemble classification, cohort-based pre-
diction, on-line games, game analytics

1 Introduction

The churn prediction is a real problem, which can be found in businesses that deal
with a permanent stream of customers using subscription services: banking [1,2],
telecommunication [3], and entertainment industries, with increasing popularity
of game analytics (e.g. [4]). The focus of business development shifts from the
attraction of new customers to retention of the old ones. Therefore modeling
users’ outflow can be used to plan company’s tactics and strategies. However,
it is often important to not simply know the outflow indicators on a macro
level, but to predict the churn probability for every customer to use the spot
interventions.
The definition of churn and the churners varies depending on a specific prob-
lem. Here we define the churners as users who were absent for 30 days and more.
Moreover, in this research we are interested in users who were absent for at least
three days and made at least one transaction. Such restrictions are motivated by
the low revenue of the late returners and recommendations from social platforms,
e.g., Facebook does not recommend to send out notifications to the players with
more than 28 days of absence.
In this paper, we propose a cohort-based classification approach to the churn
prediction for social on-line games. Our data processing pipeline includes feature
engineering and selection along with optimization of specific metrics. Thus, we
introduce a meta-metric, which penalizes undesired outcomes of the cohort test
procedure; it is designed to reflect real-life experience of using the prediction
Title Suppressed Due to Excessive Length 95

models. All the steps of the model training pipeline include optimization of
classifiers; the final step optimizes the whole ensemble on meta-metric values.
In Section 2, we introduce our cohort-based ensemble method for churn pre-
diction. In Section 3, we provide the reader with the results of experimental
evaluation. Section 4 concludes the paper.

2 Proposed Method

In a few words, the churn prediction problem can be stated as follows: given the
data about users, who recently stopped playing, predict whether a particular
user will abandon the game or not. This information is used for sending app-
to-user notifications to those users, who are likely to stop playing our game at
all. These notifications also offer some in-game goods to the user, usually some
in-game currency, rare valuable items or discounts.

2.1 Meta-metric

This problem is rather different from the standard data mining classification
problem due to several evaluation requirements, which result in so-called “cohort
test”. It is defined as follows (see Fig. 1 for short description):

1. We take the cohort of players who had been absent for 3 days by the given
date D.
2. A model, that solves the classification problem for a given day of absence,
predicts for every player, whether or not he or she becomes a churner.
3. Those, who were predicted as churners (both true and false churners), are
eliminated from the next steps.
4. We take the cohort of players who had been absent have been absent for 4
days by the given date (D + 1) and had not been eliminated in the previous
part.
5. Steps 2, 3, 4 are repeated until the date (D + 29) or no more users left in
cohort.
6. We evaluate the meta-metric, using the information about successful and
unsuccessful predictions at each step as well as users left after Step 5 (if
there are any).

Final challenges are connected with the result of the cohort test and its
influence on the game balance:

– We do not want to send prizes (bonuses) to people who are likely to return
by themselves, because it may ruin the game balance.
– We want the players to be returned as soon as possible. The probability of
the player’s future transactions as well as the average revenue dramatically
decreases with the growing number of days the player is absent for.
– We assume that we should not send notifications to the same player twice.
96 Evgenii Tsymbalov

Fig. 1. The cohort test evaluation scheme. The users, defined by a classification algo-
rithm as churners, are provided with app-to-user notification, while those users who
remain go to the next step of the algorithm according to Eq. (2).

To answer these requirements, the cohort-based meta-metric is introduced:

X
MM = (γ day−4 T Pday − αF Pday ) − β(T P27 + F P27 ), 0 < γ ≤ 1 (1)
day=4...26

This metric can be interpreted as a weighted number of returned players,
with the reward for early return and the penalty for type I error (marking a user
who will return as a churner). Note that different goods offered to a user with
notifications, as well as various social platforms and projects require different
parameters for the meta-metric.
The question of determining coefficients α, β, γ for a given project should be
discussed at both levels, analytical and managerial. While α could be estimated
by approximating the probability of a player being a payer after he or she returns,
β and γ could be estimated by experts (like we do in Webgames now) or based on
the data from the previous experiments (which is unavailable for this research).

2.2 Problem statement
D
Let Cdate = (P1 , . . . , PNdate ) be data of D-th day cohort of players on date date,
where D = 4 . . . 26, Pj ∈ R|F |×D are all player’s features up to day (D + date),
F is the set of player’s features at a particular day.
For every user in a cohort, our algorithm A makes a decision S: on which
day D = 4 . . . 26 send out the notification (or not to send, then D = 27):
j
A : (Cdate )j=4...26 7−→ S = S(D1 , . . . , DN ); (2)
Title Suppressed Due to Excessive Length 97

Our goal is to find the decision, which maximizes the meta-metric (1):
j
S = arg max
0
M M ((Cdate )j=4...26 , S 0 ) (3)
S

2.3 Ensemble of classifiers
To classify a user on a given day, we use the set of classifiers {C4 , C5 , . . . , C26 },
one for the data for each day the user absents. So the classifier C4 is trained
on the users who have been absent for 3 days and evaluated on the 4-th day
of absence; the classifier C5 is trained on the users who have been absent for 4
days, etc. This results in total of 23 classifiers.

2.4 Optimization and greedy solution
Since the main problem requires the complex elimination procedure, therefore it
seems to be complicated (if even possible) to optimize the meta-metric analyt-
ically. Instead, we use a greedy approach, with models for solving Problem (3)
optimized by the cohort ensemble classifiers parameters in terms of ROC AUC
metric, and these classifiers thresholds (decision boundaries) are optimized by
the meta-metric.

3 Experiments
3.1 Data description
We used data from the Ghost Tales project on Facebook during the period
11.2015 – 03.2016, with the last month as a test subset. Hereby the features
described below are mostly game-independent (for free-to-play games), so this
feature set could be used for other projects as well.
Here we treat a player on his/her different days of absence as different players,
which allows us to increase the number of objects in the dataset by more than
10 times, resulting in 2.7M records in total.
Typically, features could be divided into three main groups:
– Personal features: year of birth, country, etc.
– Behavioral features: the total number of active days for a player, the number
of days a player is absent, the number of completed quests, etc.
– Transactional features (based on users payments).

3.2 Data processing
Our data processing includes the following steps:
1. Feature engineering. We have calculated pairwise ratios between all the fea-
tures with the same measurement units (e.g. number of days with payments
divided by total days for a given player) and added some other feature com-
binations based on expert knowledge.
98 Evgenii Tsymbalov

2. Deletion of low variance features.
3. Selection of the best features based on F -test.
The number of the best features selected may vary for different base classi-
fiers. Here we used Extra Random Forest classifier [5] and logistic regression.
We have also tried to use Gradient Boosting algorithms (XGBoost [6] and
AdaBoost), but these classifiers shows similar or even worse results and do
not support effective parallelization. We do not use algorithms based on
neural nets due to their long training time, which is undesirable for weekly
model retraining.

3.3 Single classifier optimization
We have performed 10-fold cross-validation procedure on the training set with
optimization by ROC AUC score. The parameters used during the grid search
are summarised below.
The list of ensemble trees’ parameters (taken according to the real-life re-
source constraints):
– number of trees in ensemble: 50 – 500;
– tree depth: 5 – 20;
– minimal sample split: 2 – 15.
Logistic regression’s parameters:
– penalty type: l1 or l2;
– C (regularization constant): 0.1 – 300.
The performance summary for selected values of the parameters of Extra
Random Forest classifier is given in Table 1.

Table 1. Optimized parameters and classification metrics on the test set for the extra
trees based classifier for selected days.

Day Number of trees Tree depth Minimal sample split Precision Recall F1-score
4 500 10 6 0.73 0.75 0.74
7 500 10 10 0.66 0.66 0.66
10 300 15 6 0.64 0.65 0
14 200 15 6 0.68 0.7 0.69
18 500 15 10 0.63 0.79 0.7

3.4 Ensemble optimization
The simplest method to continue optimization is to set decision boundaries
(thresholds) for all the classifiers to the same value (which can be found via
cross-validation, for example). However, it is easy to see that we should reduce
Title Suppressed Due to Excessive Length 99

the decision boundaries of classifiers for several first days, because the users,
who returned on the first days, have more impact on the meta-metric. Since
23-parameter optimization (for every classifier) is a computationally exhaustive
problem, we reduced the number of parameters by considering the threshold as a
function of day. We optimized the thresholds for different types of such functions:

– linear: threshold = A · day + B;
– exponential: threshold = A · exp(B · day);
– quadratic: threshold = A · day 2 + B · day + C (where A, B, and C are
constants)

using differential evolution as a global optimization algorithm. However, the
optimization on various cohorts shows that exponential and quadratic approxi-
mations tends to reduce to linear, see Table 2.

Table 2. Optimal parameters found by differential evolution algorithm for the cohort
of users absent from 2016-02-21.

Approximation function A B C
Linear 0.4921 0.0141 -
Exponential 0.5416 0.0207 -
Quadratic 0.0002 0.0138 0.5344

The optimal parameters found by linear optimization results in 3-7% increase
of the meta-metric, compared to the optimized parameters for the constant de-
cision boundary, see Fig. 2.

4 Conclusion

In this research, we have proposed a cohort-based ensemble model for the churn
prediction. The final part of this model includes optimization of the original
meta-metric, which is designed to reflect the real-life experience in usage of
prediction models that rely on a cohort test; other steps are also constructed by
taking real-life resource constraints into account. Various numerical experiments
show the importance of the steps used in the model training pipeline.
During this research, the algorithm for the weekly model evaluation has been
developed and implemented in the Webgames company. Mostly automated, this
algorithm requires human assistance only at the step of the meta-metric param-
eters’ choice.
We assume it can be used in various areas for churn prediction. Thus, the
obtained results form the groundwork for model improvement and generalization
in other areas such as telecommunication and banking industries.
100 Evgenii Tsymbalov
Sheet1
180 ETC
ETC Logi
sti
cregr
.
170
Logi
sti
cregr
essi
on Random

160

150

140

130 Random cl
assi
fi
cat
ion

120
r
Metc
i

110

100

40
0,
0 0,
1 0,
2 0,
3 0,
4 0,
5 0,6 0,
7 0,
8 0,
9 1,
0
Deci
sionboundar
y

Fig. 2. Meta-metric values on α=0.8, β=1.2, γ=0.87. (The Ghost Tales on Facebook
with diamonds prize) for the cohort absent from 2016-02-27. The solid line shows the
meta-metric value for the case of the constant threshold, the dashed lines correspond
to linearly optimized values of the metric.

Acknowledgment. We would like to thank Webgames’ Analytics department
for the provided data and feedback. We are also grateful to Mehdi Kaytoue from
INSA-Lyon (France), and Evgeniy Sokolov and Peter Romov from Yandex Data
Factory (Moscow, Russia) for their advice. Last but not least we would like to
thank Dmitry Ignatov (HSE, Moscow) for his supervision.

References
1. Nie, G., Rowe, W., Zhang, L., Tian, Y., Shi, Y.: Credit card churn forecasting
by logistic regression and decision tree. Expert Systems with Applications 38(12)
(2011) 15273–15285
2. Mutanen, T., Ahola, J., Nousiainen, S.: Customer churn prediction-a case study
in retail banking. In: Proc. of ECML/PKDD Workshop on Practical Data Mining.
(2006) 13–19
3. Ahn, J.H., Han, S.P., Lee, Y.S.: Customer churn analysis: Churn determinants
and mediation effects of partial defection in the Korean mobile telecommunications
service industry. Telecommunications policy 30(10) (2006) 552–568
4. Bosc, G., Kaytoue-Uberall, M., Raı̈ssi, C., Boulicaut, J., Tan, P.: Mining balanced
sequential patterns in RTS games. In: ECAI 2014 - 21st European Conference on
Artificial Intelligence, 18-22 August 2014, Prague, Czech Republic. (2014) 975–976
5. Geurts, P., Ernst, D., Wehenkel, L.: Extremely randomized trees. Machine learning
63(1) (2006) 3–42
6. Chen, T., He, T.: XGBoost: eXtreme Gradient Boosting. R package version 0.4-2
(2015)