Time Series Clustering
Created by Chia, Jonathan on Apr 09, 2022
Introduction
Below are some of my notes on how to use time series clustering vs. market basket analysis to find complement products.
In the end, I think state-of-the-art recommender systems are the better solution; however, I think time series clustering is an interesting topic!
Table of Contents
Case Study
See this paper for full information:
Output:
After time series clustering, D12 and D14 were found to be in the same cluster
How the data would look before clustering:
Strengths:
Bypasses the need for a large matrix for association rule mining, which usually ends up being sparse and leading to redundant and not very useful rules
Instead of having a row for each basket, you have a row for each product, so data is way smaller
"Time series clustering can be used to identify products that are commonly purchased across a certain time period. Such patterns are otherwise hard to discover using association rule mining, which analyses transactions without temporal consideration."
Potential Problems:
Product quantity sold is quite low in general since we have so many products
Products may be correlated by random chance - especially if they are not bought very often
We have thousands and thousands of different products
It might be better to group products by product attributes instead of doing individual products
Why Euclidean distance is Wrong
https://towardsdatascience.com/how-to-apply-k-means-clustering-to-time-series-data-28d04a8f7da3
Quantity Sold vs. Available
Check to see if the reason the product drops right after peaking is because of lack of supply
Complements or Substitutes?
Cross join all the products to get a table like in
one problem: we cannot find substitutes because jewelry that are substitutes will never be bought together so the substitute formula proposed in the article would not work
we could check if they are complements by referencing the market basket table (maybe we can lower the parameters so we have more data)
Document generated by Confluence on Apr 09, 2022 16:54
Last updated
