Predictions
aito.ai extension is a great companion when adding and working with data, as it can auto-fill column values based on predictions.
Last updated
aito.ai extension is a great companion when adding and working with data, as it can auto-fill column values based on predictions.
Last updated
Once the training data has been uploaded, you can immediately start using the predictions through the extension.
Predictions work simply by choosing a cell or cells in your table. The extension will display the top 5 predicted results for this particular cell. aito.ai works best for predicting categories, such as short strings, single and multiselects, collaborators and so forth. We may at a later stage add support for other types of inference, such as regression ("forecast a numeric value").
The elements of the extension's predictions are explained below the screenshot.
The switch at the top allows the extension to automatically fill empty cells with the predicted value if the confidence is higher than the set percentage (90% in the picture). The threshold value is editable from the pen icon on the right.
The predicted values are on the left, with the confidence values on the right, with the question mark that will display explanations. Use button allows applying the value to the cell manually. If the selected cell in Airtable is not empty, a confirmation dialogue is shown. For multiple select fields, the Use buttons are replaced with + and - buttons to add and remove values.
You may also choose multiple cells either vertically or horizontally to predict multiple cells at the same time. Maximum of 10 cell values are predicted at once.
Hovering over the question mark next to confidence opens a tooltip that shows the top contributors for each prediction. This helps you to understand what datapoints aito.ai used to make each prediction.
In this example, the words dayton and store in the column lineText
were the top negative contributors, and having productCategory
CLASS-1274 together with words rent, management ... real in lineText
were the top positive contributors.