ARMOR Overview

The overview of ARMOR is illustrated above. In the training phase, ARM-Net is trained to model feature interactions in a selective and dynamic manner. In the inference phase, given the input tuples, ARMOR supports key functionalities such as prediction, global interpretability and local interpretability for various structured data analytics.

Let us consider a use case when a company would like to make predictions about monthly sales, and a data table containing attribute fields of (month, regionID, storeID, productID) and some predictive targets total sales are available. In such an application, ARMOR can learn to predict the monthly sales target and disclose the cross features that have been used to make the prediction. In this example, a particular store may perform better at selling a particular product locally, and all the stores may sell way more of a particular product in certain months/regions globally. ARMOR is able to dynamically identify the interactions of these features and highlight such cross features in human-understandable terms, on which the predictive analytics are based.

View our ARM-Net publication in the ACM Special Interest Group on Management of Data (SIGMOD) 2021

ARM-Net: Adaptive Relation Modeling Network for Structured Data [ paper link]

(A patent on ARMOR is being filed.)