Last year at Firebase Summit, we introduced you to Predictions, a machine learning product that helps you smartly segment your users based on their predicted future behavior. Without requiring anyone on your app team to have ML expertise, Predictions gives you insight into which segments of users are likely to churn or spend (or complete another conversion event) so you can make informed product decisions and grow your app.
As of today, Predictions makes more than 6 billion predictions per day for our developers and allows them to take meaningful actions by making predictive segments available for targeting in Remote Config, Cloud Messaging, In-App Messaging, and A/B Testing.
This year at Firebase Summit, we announced that Predictions has graduated out of beta and into general availability with a host of new features that we added based on your feedback.
Since Predictions continuously update based on actual user behavior inside your app, we heard from many of you that you wanted to know how stable a prediction was before you integrate it into your app.
To help answer this question, we created a health indicator at the bottom of each predictive segment card that gives you a snapshot of how a certain prediction is performing:
Image 1: Green means it has been performing consistently well over the last two weeks
Image 2: Yellow means it is performing well today but did not meet the quality threshold some time in the past two weeks
Image 3: Red means it is not performing well today and had other performance issues over the last two weeks
It is worth mentioning that actions targeted with Predictions have a fail-safe mechanism, so if a predictive segment is performing poorly, it simply turns inactive. That means, if you are using Remote Config to deliver a set of values to users in that predicted group, Remote Config will gracefully fall back to your default values if the predictive segment decreases in reliability. Any notifications or in-app messages directed at that predictive segment will also not trigger until the predictive segment increases in accuracy.
To help you understand how we assess the quality of a prediction, we are now exposing our evaluation criteria. For every predictive segment, we use a portion of your historical data from the last 28 days that we hold out during the model training phase.
We then compare the results of the prediction to what actually happened. This gives us two ways to score the prediction: how many of the users in the predictive segment actually behaved in the predicted way (we call that true positive rate) and how many users in the predictive segment were incorrectly classified (or in more technical terms, the false positive rate).
You can access this data from the bottom of the prediction card
Tapping on the health indicator exposes these values.
By exposing these two scores to you, you can now make a better determination about which risk profile to choose for your action.
Another common question we received during our beta phase is what went into creating a predictive segment. We now offer a details page that gives you the ingredient list! You can click through and see what data our model makes use of. This includes event frequency, volume, and parameters as well as other data like device language, freshness of app install and more.
The last thing we are excited to announce is that now, you can export your raw predictions data into BigQuery. This will give you access to the raw prediction score, the thresholds we used for each risk profile, as well as the final result. You can use this data to create your own risk profiles or if you supply your own user_id property in analytics, to do sophisticated analysis with your analytics data. For example, you can find out which countries exhibit the highest potential to churn or spend!
We are humbled to have gained your trust over the past year and hope these improvements make it easier for you to make the most out of Predictions in your mobile apps and games. As always, if you have any questions, you can find us on Twitter (@firebase) and on Stack Overflow.
For more information on these updates, check out our docs below!
Predictions risk tolerance and performance
Predictions model inputs and details page
Predictions data export to BigQuery
If you're building or looking to build a visual app, you'll love ML Kit's new face contour detection. With ML Kit, you can take advantage of many common Machine Learning (ML) use-cases, such as detecting faces using computer vision. Need to know where to put a hat on a head in a photo? Want to place a pair of glasses over the eyes? Or maybe just a monocle over the left eye. It's all possible with ML Kit's face detection. In this post we'll cover the new face contour feature that allows you to build better visual apps on both Android or iOS.
With just a few configuration options you can now detect detailed contours of a face. Contours are a set of over 100 points that outline the face and common features such as the eyes, nose and mouth. You can see them in the image below. Note that as the subject raises his eyebrows, the contour dots move to match it. These points are how advanced camera apps set creative filters and artistic lenses over a user's face.
Setting up the face detector to detect these points only takes a few lines of code.
lazy var vision = Vision.vision() let options = VisionFaceDetectorOptions() options.contourMode = .all let faceDetector = vision.faceDetector(options: options)
The contour points can update in realtime as well. To achieve an ideal frame rate the face detector is configured with the fast mode by default.
fast
When you're ready to detect points in a face, send an image or a buffer to ML Kit for processing.
faceDetector.process(visionImage) { faces, error in guard error == nil, let faces = faces, !faces.isEmpty else { return } for face in faces { if let faceContour = face.contour(ofType: .face) { for point in faceContour.points { print(point.x) // the x coordinate print(point.y) // the y coordinate } } }
ML Kit will then give you an array of points that are the x and y coordinates of the contours in the same scale as the image.
The face detector can also detect landmarks within faces. A landmark is just an umbrella term for facial features like your nose, eyes, ears, and mouth. We've dramatically improved its performance since launching ML Kit at I/O!
To detect landmarks configure the face detector with the landmarkMode option:
landmarkMode
lazy var vision = Vision.vision() let options = VisionFaceDetectorOptions() options.landmarkMode = .all let faceDetector = vision.faceDetector(options: options)
Then pass an image into the detector to receive and process the coordinates of the detected landmarks.
faceDetector.process(visionImage) { faces, error in guard error == nil, let faces = faces, !faces.isEmpty else { return } for face in faces { // check for the presence of a left eye if let leftEye = face.landmark(ofType: .leftEye) { // TODO: put a monocle over the eye [monocle emoji] print(leftEye.position.x) // the x coordinate print(leftEye.position.y) // the y coordinate } } }
Hopefully these new features can empower you to easily build smarter features into your visual apps. Check out our docs for iOS or Android to learn all about face detection with ML Kit. Happy building!
We're delighted to announce the beta launch of Firebase Predictions. With this we are bringing the power of Google's machine learning systems to every developer that uses Firebase. Predictions is a product that can build dynamic user groups based on predicted behavior, determined using a machine learned model, and these user groups can then be targeted using Firebase Cloud Messaging, Remote Config and other technologies. User groups are updated daily to keep your predictions fresh.
Out of the box, there are four predicted groups:
You'll see these three groups right away when you select Predictions in the left nav bar of Firebase Console.
You'll notice that each of these cards has actions that you can take upon them.
Tolerance: The tolerance slider gives you the ability to tolerate low, medium or high risk of false positives. So, with a low tolerance, your population of users will be smaller, but so also will your risk of false positives. Similarly, with a high tolerance, you'll have a larger population of users, but at a risk of some of them being false positives. In the case of 'churn', a false positive would be a user who is predicted to churn, but in fact continues to use your app.
Target Users: This gives you a drop-down on which you can select Remote Config or Notifications for that user group. It also links to some handy guidance for offering in-app incentives.
Selecting Remote Config will take you to a new screen where you can specify the remote config parameter that you want to set up, and then the value for it for that population. So, for example if you've been building a game, and a lot of people have churned and you see from feedback that it's too difficult to play, you could set a remote config variable for the difficulty, so that likely churners could get a lower default value set for them, and thus would have an easier experience playing the game.
Selecting Notifications will take you to the familiar composer for messages to be sent using Firebase Cloud Messaging, but in addition to the usual options for picking target audience, you'll also get the predicted user group pre-populated as a user segment.
This allows you to target notifications at that user group. So, for example, for users at a risk of churning, you could send a notification with an enticement to continue using the app.
Creating your own predictions. You aren't limited to the built-in predictions cards, of course, and can create your own based on custom events that you set up in your app. In this case, you'll see a card that allows you to create a prediction.
And when you select it, you can then create a prediction for when your event will, or will not happen. This helps you identify users who are likely to engage in that conversion event:
So, for example, in the above case, whenever a user levels up in the game, the level_up conversion event is logged. Thus, you could create a prediction for players who may level up, and incentivize them to continue playing.
Then, once you've saved your prediction, over time a card will populate on the Firebase Console in the same way as the built-in ones.
And this card can be used in the same way as the others -- including targeting users with notifications and Remote Config.
Firebase Predictions is a Beta product, and we're continuing to work on it and improve it. If you have any questions or feedback, please reach out -- and for bugs and product suggestions, you can reach us at firebase.google.com/support.
Learn more about Firebase Predictions at firebase.google.com/products/predictions/ or dive straight into our docs right here.
