Prediction Block [Machine Learning]

This airtable block allows you to train a machine learning model on your base table data, evaluate the performance of the model and predict target variables given some features. You data is completely private as model training and evaluation all happens locally in your browser.

User Story

Jack is a real estate agent who has data (~5000 records) on housing prices across various cities in California. However, there is a new house coming up on the market which is unlisted (and hence no price) at the moment. Jack is a champion for his clients and would like to start discussing this new house. To be effective in this conversation Jack would like to have some estimate of what the price of the house might be given its features - location, number of bedrooms, baths, proximity to parks etc. Jack also wants to do this in a manner that is fast and private (sharing his data to external model building servers is a no no).

Jack uses the ML prediction block

  • Jack selects the target feature of interest -housing price. Since this column is a number, a regression model is suggested.
  • Jack opts to train a regression model using all other columns as features. The prediction block automatically infers the data type for each column and applies appropriate transformations. e.g categorical columns are one hot encoded
  • Jack views some metric on training progress and some numbers that indicate if the model "good or not".
  • Jack is satisfied with the model's performance and saves it. It is saved as a new base with each saved model a records and the model file as an attachment.
  • Following this, Jack (and all his collaborators) can now enter features of the any new house and get a prediction on its price using a saved model. As more data arrives, Jack can train new models. Jack now has more information about the likely price of the house to advise his clients during their meeting.

The best part is that Jack accomplishes all of this without exposure to privacy leakages (e.g. an external server). The model is trained in Tensorflow.js right in Jack's browser and his data never leaves his airtable base or local machine!

How does all this work? Hint: Tensorflow.js

Machine learning has demonstrated potential in unlocking value for businesses across multiple domains. Until recently, they required expensive server environments to run. With Tensorflow.js, things are different. Tensorflow.js, a framework for machine learning in javascript, is used to build a neural network model that is then applied to the task of prediction or classification.

When to Use Prediction Block

Training a model in the browser can be resource intensive and not designed from very large datasets. For data running into 100s of thousands of records, server based neural network implementations are strongly recommended.


Housing Price Prediction

  • Predict the price of a house given its features.

Built With

Share this project: