Model Manager¶

When we apply machine learning models to time series data, we usually have a problem that data patterns change over time and the accuracy of the model gradually decreases. In this case, data scientists ask the data manager to get the new data and then build the model. Theu have to go through re-learning to get a certain level of accuracy and redeploy to the system. This can sometimes take up to several months.

Metatron Anomalysupports the Model Manager, which allows users who are not a data scientists nor a data manager to easily retrain the model.

The model manager consists of the following functions.

Model accuracy fluctuation

Model re-training and learning history

Comparison of models and application of new models

Click icon_model in the right menu of the created alarm rule detail page to enter the model manager.

Model accuracy fluctuation¶

The upper part shows how much the model accuracy has increased or decreased compared to the most recent learning, and the numerical value shows that the accuracy score changes over time when the mouse is over the graph. At the bottom, the information of the currently applied model and the timing of application are indicated. It’s possible.

Model re-training and learning history¶

If the accuracy is lower than the desired value, you can re-learn by clicking the Train button at the top right. Select the range of training data and algorithm type to be re-trained and press the Train button to start the job.

When re-learning starts, you can see the current status of learning in the menu recorded as the current time in Training History. You can also check the history of the past in the list.

Comparison of models and application of new models¶

Click the icon_comparison icon to the right of the new model to compare the previously applied model with the newly trained model. The previously applied model is marked with a blue line, and the newly selected model is marked with a pink line to show the values predicted by the two models. You can compare abnormal score values at the same time.

To apply the newly trained model to the rule, click Apply this training model from the icon_ham menu on the right. The applied model is tagged with icon_applied .