3 Advantages of goedle.io over Google Firebase Predictions
Google Firebase released Predictions at the end of October 2017. When one reads the summary of the Firebase Dev Summit on Google’s blog, one almost has trouble ignoring the similarities between Firebase Predictions and our platform at goedle.io. Below you can see two images. First is our Techstars Demo Day presentation in September 2016 (see the full video). Second is a screenshot from Firebase Prediction’s introductory video on YouTube.
Naturally, as CEO and co-founder of goedle.io, I prefer to see customers using our platform. However, I’m not saying that there are no scenarios in which Firebase Predictions is not a good solution. In support of my argument, let me briefly describe my background. For years, I have been working on the prediction of user behavior, in particular churn and purchase prediction. I have published research papers in this field (see here and here) and have also applied this knowledge in the industrial context. In 2014, we started working on goedle.io with the goal of democratizing machine learning and making it accessible to a wider range of areas and businesses, including mobile apps and websites. In the process, we accumulated significant experience and knowledge, including the understanding that customization and high-quality support are very important, because AI and machine learning nearly always require a profound introduction.
Our platform provides similar capabilities as those of Firebase Predictions, including churn prediction, purchase prediction, and the setup of your own predictions. We make these predictions available through our dashboard and API. We have similar use cases such as sending smarter push notification with a 65% increase in retention or ramping up your CRM and speeding up your sales process. Compared to Firebase Predictions, we provide more options to individualize the offering for our customers and we are very flexible when it comes to making the predictions accessible. Following are three reasons why I encourage you to reach out to us if you are considering Firebase Predictions or if you have it in place already.
#1 — Custom Features
Machine learning is still heavily dependent on feature engineering and many features rely on a detailed data tracking plan. As we built our platform over the past few years, we accumulated a large set of features. Still, every app or website is different, with its own properties and dynamics. It is often helpful to add a small number of custom features to each customer’s feature set. For example, in many cases the timing between two events is very important. One instance involves the time between starting and finishing a level in a mobile game. However, when considering all timings between all possible events in a game, this quickly becomes a mess and is often rather obstructive. For that reason, we either selectively add such features or use smart feature selection algorithms to find the most important timing events. We have applied this approach successfully in our churn prediction for NCSOFT’s Blade & Soul and in the student at-risk prediction within the ENVISAGE project.
#2 — Flexible Prediction Windows and User Segments
Firebase’s churn prediction forecasts whether users disengage over the next week. In many cases, a fixed window of seven days is not appropriate to really define churn or other behavior. Think of apps or services you do not use on a weekly basis. If you run a software-as-a-service business, it’s like that not all users will log in within the next seven days. However, many of them might still might be loyal customers. Here are two images showing the maximum absence time for two apps from different areas:
As you can see, many users — even for a high-frequency app — still return after much longer than seven days and, hence, should not be treated as potential churners. For that reason, we allow the adaptation of the prediction window to match our customer’s needs, while Firebase Prediction fixes it at seven days.
Secondly, how long do you observe players? Do you want to observe all players for an equal amount of time (1), or do you want to consider everybody who has been active in the past n days (2)? We have seen that it is beneficial to learn different models for different use cases. If you have drip campaigns in place, you may reach out to all users after one week but want to personalize the content depending on the churn likelihood (use case 1). In other cases, you care only about whether users have been active during the past 14 days, and you wish to react according to their most recent behavior (use case 2). Here, goedle.io helps you customize your setup, and sets up different churn or conversion predictions for you.
Another element of customization is the number of groups in which our prediction segments your user base. Firebase uses three levels of risk tolerance to avoid approaching the wrong users. We go in a different direction. We also begin with three groups (see dashboard screenshot above), but with our platform you can have an arbitrary number of groups. In the past, we have seen that groups with low algorithmic certainty, i.e., a high risk level, are particularly interesting. For those users, the leverage of marketing campaigns can be the largest because these users are on the fence. Why message people who are leaving with very high certainty? One push message won’t change all of their behavior, and spamming churners is also bad practice. However, convincing users on the fence helps you grow your user base or increase revenue.
#3 — Different Sources and More Integrations
While Firebase is app-centric, we not only support mobile apps but also work with websites, for example, e-commerce shops or SaaS platforms. We also support Google Tag Manager as a data source and, akin to Firebase, a Segment-integration is possible with only a few mouse clicks. On top of that, we provide a raw HTTPS-API and we can track data from other third-party tools such as Adjust. Lastly, we support bulk data upload for historic data import or pre-integration test runs.
When it comes to accessing the predictions, Firebase offers Remote Config and push notifications via FCM. We are more flexible, as we offer an HTTPS-API to access the predictions on a user level. Of course, we also support other integrations such as sending data back to Segment, emailing customers via third-party services (e.g., with our Mailgun-integration), and push notifications. All of these integrations can be managed easily with the help of our dashboard.
We also provide a service similar to Remote Config, in which different strategies and content can be tested and distributed to users. For example, our dynamic difficulty adjustment helped a quiz game increase its ad revenue by 74%.
Let me also give you a reason to work with Google Firebase Prediction. Certainly, Google Firebase is a comprehensive self-service with many dashboards and extensive documentation. If you are already well-informed about the requirements of machine learning and if you feel comfortable with the topic, Firebase Predictions is certainly an interesting product through which to get started right away, in particular when your entire setup is already tied to Google Firebase. While our platform does not support all of Firebase’s features, e.g., crash reporting, we focus on the prediction of user behavior and we are more than happy to tune our algorithm to your setting.
A decision in favor of one of the two platforms is also a decision between two different paradigms. Either you prefer Google’s all-in-one platform, or you go with a best-of-breed solution and choose the leading service for each task. When it comes to predictions, we encourage you to reach out to us for our solution. If you are just getting started with predictions and machine-learning-enhanced marketing, we also offer workshops explaining how the technology works, brainstorm ideas with you, and prepare you for the next level of marketing. Please feel free to reach out to us, we would love to show you a demo. And don’t forget to subscribe to our newsletter and follow us on Twitter.
Originally published at blog.goedle.io.