Figure 1 shows the general outline of the ML-backend. Except the primary goal of creating an efficient backend one design guideline has been keeping training data and the ML-artifact separate for each user, this to be able to target privacy and security issues.
Our context, commuting, gives us some metrics that we need to target. One central metric is accuracy; others are response time and training time. Aspects of the accuracy is a function of many factors and will be central in further discussion during this project. To target the situation when a user starts the app we want to deliver journey predictions within a few seconds. Regarding training time demands depend on the ML approach, in an interactive training setting, training can be allowed to take hours since the users don't expect new correct predictions until they are in the same situation next time. In a machine teaching session users can expect retraining to take place in less than 10-20 seconds so the teacher can evaluate the training during the session.
Figure 1: Backend that saves labelled data from the app and delivers predictions to the app. When new labelled data is entered by the user (this could be done explicitly or when the user selects a new route) the labelled data is saved and can be used for training. The coordination for this is done via the real-time database and cloud functions. Retraining is currently done one hour after the last use of the app. Predictions are initiated when the user/app uploads features (weekday, time, location, activity) to the real-time database, the Node-server then performs the predictions and returns the result via the real-time database.
Initially we made a technical verification that uses the backend and delivers predictions in the app. We reached our main metric for the backend and can deliver predictions in most cases in a few seconds. Retraing in this initial setting took around 10 minutes due to selected hardware (Non GPU server) and ML Framework (TensorFlow). These test were mostly to verify the functionality. An animation that visualize a prediction can be seen on this site in the animated gif to the left.