1. On Prophesying Online Gamer Departure
Pin-Yun Tarng1, Kuan-Ta Chen2 and Polly Huang1
1
Department of Electrical Engineering, National Taiwan University
2
Institute of Information Science, Academia Sinica
3. Evaluate the average daily playtime and playing density in Figure 5: Unsubscription prediction accuracy
Introduction and Motivation each period
4. SVM as classifier, treat k -period features and predetermined
categories as the training data set.
Business model of online game companies usually depends on:
Complete Scheme
• Sale of virtual items
100
• Monthly subscriptions in which gamers must pay for credits Playtime features
Density features
to continue their adventures in the virtual world. Both features
The combination of our classification method and prediction
90
Being able to predict how long people will stay in the game will model:
Accuracy (%)
directly affects game companies’ revenue.
80
Fade-out
prediction model
This study provides a practical scheme for predicting player un-
70
Staying
subscription: Classification
model
Leaving?
• Input– a player’s game hours 60
Fade-out
Leaving
• Output– whether or not he will renew an expiring subscrip-
5 10 15 20 Gameplay
k value Trend?
tion. Sudden-out
• Our rationale– if we can predict the departure of a player
Figure 3: The classification accuracy of different values
before he actually quits a game, the game operator can
of k.
take remedial measures to prevent it from happening and
improve the game along the way based on the feedback Figure 6: The complete unsubscription prediction scheme
provided by such a player. To find the optimal k value, we experiment within the range of
[2, 20], ten-fold cross-validating for each value. From Fig.3, we
Our traces are from ShenZhou Online, a mid-scale commercial find that the optimal value is 10. Input – a player’s incomplete trace
MMORPG in Taiwan sustaining at any moment thousands of Only incomplete data is available in real-life prediction – Output – three way output:
players online. gamers’ traces, with last n days cut off (n in [3, 60]), are fed • Sudden-out pattern (or just unpredictable)
into the SVM model. Fig.4 shows the result.
• Staying for the time being
ShenZhou Online Traces Summary
2003-03-01 • Leaving within a specific number of days.
2007-02-15 The accuracy of our complete prediction scheme is shown in
Longer than 2 years
Fig.7, and Fig.8 shows the three types of errors.
100
1,447 days
Longer than 1 year
102,233,240 Longer than 0.5 year
Shorter than 0.5 year
90
Accuracy (%)
162,980
20,514
80
Longer than 2 years
100
Longer than 1 year
70
Longer than 0.5 year
Shorter than 0.5 year
Figure 1: Summary of ShenZhou traces 90
Accuracy (%)
60
0 10 20 30 40 50 60
80
Before unsubscription(days)
70
Figure 4: Predictivity of our classification method.
Classification of Online Gamers
60
0 10 20 30 40 50 60
d value (days)
From observations on gamers’ playing history (as in Fig.2), an
intuitive categorization of unsubscribing players: Model for Predicting Unsubscription Figure 7: Accuracy of our complete prediction scheme
• Fade-out, with ever-decreasing daily playtime and login
frequency
How to predict whether a gamer is leaving in d days? (d in [3,
• Sudden-out, no noticeable tendency in daily playtime or False leaving False staying False sudden−out
60])
50
50
50
login frequency Longer than 2 years
Longer than 1 year
Longer than 2 years
Longer than 1 year
Longer than 2 years
Longer than 1 year
Similar to our classification method, for each gamer: Longer than 0.5 year
Shorter than 0.5 year
Longer than 0.5 year
Shorter than 0.5 year
Longer than 0.5 year
Shorter than 0.5 year
40
40
40
1. Assign prediction point at d days before his quitting
30
30
30
Error rate (%)
Error rate (%)
Error rate (%)
20
20
20
20 20 20 2. Derive two random observation windows, counting from the
Daily playtime (hrs)
Daily playtime (hrs)
Daily playtime (hrs)
10
10
10
15 15 15
gamer’s first login day
10 10 10
0
0
0
5 5 5 3. Leaving window – contains the prediction point, unsubscribe 0 10 20 30
d value (days)
40 50 60 0 10 20 30
d value (days)
40 50 60 0 10 20 30
d value (days)
40 50 60
0 0 0 within d days after window; staying window – not contain
0 200 400 600 800 1000 0 200 400 600 800 0 200 400 600 800
Day Day Day the prediction point, still stay at least d days after window
20 20 20
4. Extract 10-period features from each window Figure 8: False positives and false negatives of our pre-
Daily playtime (hrs)
Daily playtime (hrs)
Daily playtime (hrs)
15 15 15
5. Fed to the SVM along with corresponding window type diction scheme
10 10 10
5 5 5 The prediction accuracy for each d value is shown in Fig.5
0 0 0
0 200 400 600 800 0 200 400 600 800 1000 0 200 400 600 800 1000
Day Day Day
Fade−out Sudden−out
Longer than 2 years Longer than 2 years
Conclusion
100
100
Longer than 1 year Longer than 1 year
Longer than 0.5 year Longer than 0.5 year
Shorter than 0.5 year Shorter than 0.5 year
Figure 2: The playing history of six sample gamers
90
90
Accuracy (%)
Accuracy (%)
The ability to predict a gamer’s departure is coveted by the
80
80
MMORPG industry as it allows the game operators to target
How to perform automated classification?
70
70
their resources on keeping subscribers motivated and to bene-
1. Randomly choose 2,000 gamers, classify them with the human fit from these loyal customers. To this end, we hope that our
60
60
0 10 20 30 40 50 60 0 10 20 30 40 50 60
eye d value (days) d value (days) scheme will prove helpful to operators, as well as gamers who
2. Divide each gamer’s history into k periods of equal length may enjoy a better gaming environment because of it.