Increasing user retention with Firebase Predictions (Firebase Developer Story)
Articles Blog

Increasing user retention with Firebase Predictions (Firebase Developer Story)

August 27, 2019


FRANCISCO GONZALEZ: Halfbrick
is a game studio based out of Brisbane,
Australian, with teams in Sydney, Adelaide, Spain,
Bulgaria, and Los Angeles. We are probably best known
for the game “Fruit Ninja,” though we also have several
other big titles, including “Jetpack Joyride.” In mobile games,
user retention is one of the toughest
problems, so being able to use machine
learning to proactively target users before they
churn sounded almost too good to be true. When we heard about
Firebase Predictions, we were excited to try it
out in one of our games, “Dan the Man.” MIGUEL PASTOR: We
had experimenting with building churn
prediction models in house, but they struggled to find the
time or resources to properly devote to the problem. Even once we had a model,
the problem wasn’t solved. It was still difficult
to implement changes based on targeting that
the model provided, so we decided to put prediction
to the test with a [INAUDIBLE] experiment to see
if we could boost user retention by offering
a gift [INAUDIBLE] currency. We divided our
experiment population into three groups, Group
0, Group 1, and Group 2. Group 0 was the control,
and received no promotion. Users in Group 1
received the in-game gift when they completed Halfbrick’s
[INAUDIBLE] heuristic, beating level three. Finally, users in
Group 2 received the gift if they were identified
by prediction as, will churn. The results speak
for themselves. By sending the game
promotion to users who were predicted to churn,
we saw a 5% point increase in retention, which
equated to a 20% gain. FRANCISCO GONZALEZ: We were
so excited at the success from “Dan the Man,”
that we decided to try this feature in other titles. “Fruit Ninja” and
“Jetpack Joyride” are already preparing
experiments using predictions.

Only registered users can comment.

  1. This is a great video. I think the value of predictions is underestimated by most teams. It is good to see a real life use case. Thanks!

Leave a Reply

Your email address will not be published. Required fields are marked *