MACHINE LEARNING INCREASE OUTBOUND MARKETING CAMPAIGN CONVERSION BY 20%
Outbound marketing campaigns are common practice in games where players received messages via different channels from the publisher. These messages could be delivered in-game; or via external channels such as e-mail or push notifications.
The most basic campaign would simply blast the same message to all the players in what we call spray-and-pray approach, hoping for any conversion you can get. This method is easy to execute but can be an annoyance to players who may not be the correct audience of such messages. For example, messaging a player who hasn’t been back in the game for the past month about a new weapon he could purchase for the game would likely go straight to trash; whereas a message inviting the same player to come back to the game with a token bonus would be more relevant and effective.
More advanced campaigns would segment their players based on certain attributes and send a targeted message to each segment. These segments are simply parameters they use in database queries to group them into buckets. For instance, we could group players into monetized vs. non-monetized groups and send them a different message. The monetized group are players who have made a purchase in the game with payment information saved whereas the others are players who have not. We could send the monetized group a message about getting 20% more tokens if they purchase within the next week, and the non-monetized group would get a message about getting free tokens when they enter their payment information but not get charged for it. This type of segmentation would allow you to optimize the success rate for each of these two groups of players.
With the machine-learning algorithm, we could get an even better conversion rate by predicting which players would positively respond to our message. One of our clients with an FPS (First Person Shooter) game wanted to promote a new expansion pack they had just launched to the existing players. Using logistic regression machine-learning method, we created a model to predict which players are most likely to purchase this new pack. The result in a 20% increase in conversion compared to a previous campaign we ran using the segmentation method.
Here are a few more details for those interested in the math. The probability of players would like to buy (Purchase_%) could be expressed as:
Where b0 is the interception, bi are coefficients, xi are variables. (xi includes players’ demographics parameters such as age, gender, country, historical game behaviors and purchase history such as purchase frequency, average purchase amount, recent purchase date, recent purchase amount etc.)
With over 3 million+ players, 70% were assigned as the training set, 15% as the validation set, 15% as the test set.
We used the training data set to find the best fit coefficients bi and derive Purchase_% using the formula above. By targeting the players with at least 50% score in Purchase_%, we were able to increase the conversion by almost 20%.