Functions, Users, and Comparative Analysis
We decided that Docs should have prime location.
Build AI products you can trust.
We’re super excited to share that Aporia is now the first ML observability offering integration to the Databricks Lakehouse Platform. This partnership means that you can now effortlessly automate your data pipelines, monitor, visualize, and explain your ML models in production. Aporia and Databricks: A Match Made in Data Heaven One key benefit of this […]
Fundamentals of ML observability
Metrics, feature importance and more
We’re excited 😁 to share that Forbes has named Aporia a Next Billion-Dollar Company. This recognition comes on the heels of our recent $25 million Series A funding and is a huge testament that Aporia’s mission and the need for trust in AI are more relevant than ever. We are very proud to be listed […]
Monitoring system can support all of the existing ML platforms and frameworks used by the team.
The solution is able to integrate with existing databases and datalakes that store production data (e.g. S3, ADLS, etc..).
The solution can integrate with python-based serving infrastructure.
There’s a centralized place where users can see all production models with their health status and recent activity.
View the performance over time of 2 different model versions in comparison mode to quickly identify the best performing one.
When ground truth is available, view performance metric (accuracy, f1, etc.) and how they change over time.
In cases where ground truth is not available, view average prediction and other aggregations like mean/sum/std dev on predictions over time to evaluate model performance.
Live distribution analysis of production data & predictions.
For investigation purposes, allow distribution comparison of:
* Different model versions
* Different time frames
* Different data segments
The platform provides tooling to visualize various metrics and the way the change over time to identify correlations.
View and compare live data & prediction statistics including the following info for each feature: Numeric – Feature name, Mean, Std Dev, Zeros, Min, Median, Max Categorical – Missing, Unique, Top, Freq. Top
Define segments of interest (e.g. state = “CA” and age >30) and provide tools to analyze prediction distribution and performance across different segments.
Slice the data by segment groups i.e. segment by Age groups will result in: age<10, 10<age<20, 20<age<30, etc.. For each, a segment analysis will be available with a view of group behavior for identifying misbehavior in specific segments.
Get drift score for each feature and prediction for quickly identifying drifting features.
The solution supports creating customized monitors to detect Data Integrity issues:
* Missing Values
* Model Activity (inference count)
* New Values
* Out of range
The solution supports creating customized monitors to detect data drift.
The solution supports creating customized monitors to detect prediction drift.
As different models have different data and performance metrics, the solution will allow an easy way to customize the thresholds and monitoring logic of each monitor.
The solution supports creating customized monitors to detect anomalies and sudden changes in metrics such as:
* Avg
* Min
* Max
* Variance
* Standard Deviation
The solution supports customized monitors for performance degradation and comes out of the box with the standard performance metrics:
* Accuracy
* Precision
* Recall
* F1 Score
* AUCROC
* MSE
* RMSE
* MAE
* Logloss
* WAPE
* MAPE
Users are able to define their own custom metrics within the platform and monitor them for anomalies and degradation.
The solution allows setting the training set as a baseline for a monitor (e.g. data drift compared to training).
The solution allows monitoring data anomalies over time (e.g. unexpected seasonal changes in missing values)
The solution allows monitoring specific populations (data slice) for anomalies and unexpected behavior.
Users can easily analyze specific prediction and see what was the contribution of each input to the final prediction.
Users are able to explore what-if scenarios by changing some input features, and watching the effect on model’s prediction.
The system is able to generate a non-technical explanation sentence for each prediction.
Alerts can be received via e-mail.
Alerts can be received via slack.
System supports generic integration to 3rd party solution by triggering a webhook.