Visualizing the Logged Data

Users can visualize the logged data in the Data Monitoring section of the dashboard.

This dashboard generates all the graphs corresponding to a model version which can be selected from a dropdown.

The Dashboard has three sections:

Model Summary

This section consists of time series graph of the prediction volumes, actual volume and custom-metrics graph. Here is how it looks in the dashboard.

Data Distribution

This section displays the distribution of features, predictions and actual for a given time range.

It plots the following graphs for features/predictions:

• Categorical: A histogram representing counts corresponding to each class.
• Numeric: A histogram with 10 buckets from minimum to maximum. Each bar representing the number of feature/prediction lying in that bucket.

Note: Each bar of histogram is color coded with a light color denoting a low custom_metric value which is the defined loss function. This can provide an insight that for what prediction value, feature value the model is not performing well.

You can also add comparison between two time windows:

Raw Data

This is the dataset browser, where user can browse through the dataset.

Applying Filters

Filters can be applied on features, predictions, actuals and custom metric values on all three pages.