Graphs as a Descriptive Analytics Tool
If you have had a chance to look at the Olist Case Study, you can see that even within one geographic region, IID (Independent Identically Distributed) is not a reasonable assumption to analyze purchasing patterns of customers. For some cities it is, but for others it is not. As discussed in the post, a reasonable approach in such scenarios is to use methods used to analyze IID data for the group of cities for which IID is a reasonable assumption. For the remainder, we can use Graph based analysis methods.
This pattern of finding different behaviors for different subsets of the data is a very frequent finding in real life datasets. Using a monolithic modeling approach can be sub-optimal. If we use an IID assumption for all the data, then we have a model that is not correctly specified for some parts of the data. We could use a graph based approach for the entire dataset, but we don’t need the neighborhood features for some parts of the data (the IID parts). So we have a lot of redundant computation. Doing a descriptive analytics exercise lets us pick the right modeling approach for each subset of the data.
If you look at all the examples, the raw dataset is tabular. This can be viewed as a mathematical relation. Conceptually, each row in the input dataset can be viewed as a set an interaction between one or more entities. In the Olist case, it is a customer’s interaction with the store’s inventory. For some subset of this input dataset, a graph representation is useful to analyze it. Since non-IID (dependent) data does occur frequently, learning a graph from a set of relations is a frequent problem.
I have a write up about this problem and a proposed solution sketch in the descriptive analytics repository: Learning Graphs from Relations.
Citation
@online{sambasivan2025,
author = {Sambasivan, Rajiv},
title = {Graph {From} {Relations}},
date = {2025-07-27},
url = {https://rajivsam.github.io/r2ds-blog/posts/graph_from_relations},
langid = {en}
}