Albert now provides property predictions for tasks! Predictions are generated by Albert’s machine learning models that are continuously trained on your historical data for the given property or data column. Predictions will be shown in cases where models have high confidence and proven accuracy against historical data.
Albert's models are still under beta testing, as we continue improvements to our models. The following article outlines the how Albert is leveraging machine learning to show predictions and how to interpret the outputs.
How do Albert’s Property Models work?
A property model is a tool that generates predictions based on formulation and process information as shown in the below diagram.
These models are built on historical data using machine learning. Models use data to understand relationships between inputs and the output. In Albert's property models, inputs come in the form of formulation data (broken down to the CAS level) and process parameters which are then mapped to property measurement outputs.
Predictions in Tasks
When a task is created, Albert uses stored property models to make a prediction for that task. For predictions with high confidence, you will see a prediction and standard deviation appear in the task's data template grid in the ‘Albert AI Prediction’ row. This prediction can be used as an estimate for the end result of a task.
Standard Deviation
Standard Deviation, shown as ±16, is a measure of the amount of variation (or deviation) that might be expected between the actual value and the predicted value. The prediction’s standard deviation is a byproduct of the closeness between the prediction’s inputs and those of the data that the model is trained on. If a prediction is made on a task whose inputs are close to another task that the model has seen, will have higher confidence in the prediction and the resultant standard deviation will be lower. Conversely, when a prediction falls far from the model’s training data, the model’s prediction will have higher uncertainty.
What determines what predictions I see in the platform?
Predictions are shown based on a few key factors: historical model performance, model confidence (or standard deviation), and percent of known CAS in the formula. These are continuously updated and monitored by Albert to ensure the most proven model predictions are shown.
Additional References
Below are links to resources that dive into more detail on Gaussian Process models, standard deviation, Bayesian optimization, and more on the topic of AI/ML.
https://towardsdatascience.com/gaussian-process-models-7ebce1feb83d
https://medium.com/analytics-vidhya/bayesian-optimization-9ddb3aff0eb4
Comments
Please sign in to leave a comment.