14 Takeaways

This document covered anomalies in detail. We conclude it with a list of key points:

• Anomaly dectection is an unsupervised learning method used to detect instances that do not follow a regular pattern.

• BigML anomaly use an optimized implementation of the Isolation Forest algorithm, a highly scalable and efficient method that usually yields the best results compared to other anomaly detection techniques.

• BigML computes an anomaly score for each instance and a measure to indicate the relative contribution of each input field to the anomaly score.

• BigML anomalies support categorical and numeric fields as inputs, text and items fields will not be taken into account to compute the anomaly score.

• BigML anomalies also supports missig data.

• To create anomalies you just need an existing dataset. Then anomalies can be used to make a single score prediction or a batch score prediction. Additionally, you can create a dataset from anomalies. (See Figure 14.1 .)

• You can use the 1-click option to create your anomaly or you can configure the several parameters provided by BigML before.

• When the anomaly has been created, you get a list of your TOP ANOMALIES ranked by score.

• You can inspect your anomalous instances values in the DATA INSPECTOR.

• You can create a new dataset removing your anomalous instances or including them.

• You can use your anomaly to score single or multiple instances in batch not seen before by the model.

• You can create, configure, update, and use your anomalies programmatically via the BigML API and bindings.

• You can download your anomalies to locally score your new instances.

• You can add descriptive information to your anomalies.

• You can move your anomalies between projects.

• You can share your anomalies with other people using the secret link or embedding them into your own applications.

• You can stop your anomalies creation by deleting them.

• You can permanently delete your existing anomalies.