Better UI for models, now you can train multiple models inside for-loop and pipeline and assess their performance via quality metrics and diagrams.
We would like to note that for the models trained inside the for-loop brick, you may not only see the general information about the model performance but dive into the detailed performance for each value of the stratifying variable. This allows to highlight the weak and strong points of the solution and use this information for further improvement.
Moreover, the model output dashboard allows manually verifying the model response for the different input states by using so-called What-If analysis and provides a convenient interface for the model save and download. Also, you may get the complete information about data processing flow from Data Import to Model Training that is required for the results reproducing.
Datrics models can be used in your custom python code and within your own infrastructure. Now you can download the model trained in Datrics in JSON format and use our open-sourced library to use it in your python code or run integrate it to your website backend.
See documentation: https://wiki.datrics.ai/Datrics-Model-Deserialization-from-JSON-dbc736dc8eea4d9e94e33ba12a82e5b8
Now you can export your pipeline graph in HTML format to see all bricks, their description, statistics, and metrics. You can do the same thing for each model trained in Datrics. You can generate a model description that will include a graph of computations used to generated features and all model metrics. This file can be passed to your models` verification department and provides great transparency.
Now you can get predictions from your data using just 1 brick. Drag "Predictive model" brick to your pipeline, connect data to it, and select the target variable. We automatically perform the detecting of the supervised learning problem's type, based on the selected target variable, as well as the selection of the input features that are appropriate for the modeling, so you will receive a trained model and predictions just in a couple of seconds.
Charts brick has a lot of improvements:
Charts brick has a lot of improvements:
The errors section was redesigned. Now you can see not only errors but also notifications about the not obvious behavior of bricks which can give better transparency about the pipeline. All notifications were moved to the top of the right panel and shown on bricks when you hover to the icon in the top right corner of the brick.
Now handling outliers become much easier. The new brick supports three types of outputs - select, indicate, and remove and three algorithms - IQR, IForest, and One-Class SVM.
For your convenience, we have the simple and advanced mode in the Outliers Treatment functionality. In the simple mode, the user may change the outcome type only - outliers treatment is performed based on IQR strategy. In the advanced mode, you may not only choose the algorithm but configure its sensitivity.
Added a possibility to combine conditions for rows filtering with OR/AND logical operator.
Update Flatten JSON brick with new functionality to filter out JSON tags, ability to specify the maximum nested level to parse, added options to omit complex names for the newly generated columns.
Now it's become possible to compare dates in the Compare Brick.
API Input now properly converts strings to date-time and won't cause errors.
Now you can set custom random seed for the pipeline.
