Does clustering make sense at structured data? I have a sales data set from a retail company. The data is structured into Product Lines, Product Groups and Colors and Articles.

I wanted to make a cluster analysis to find similar articles or colors, on which I can make a time series analysis and build my predictive model on.

Clustering the data with k-means or two-step algorithm will bring probably each product line as a cluster. Right?

  • $\begingroup$ It ALL depends on the distance you will consider for clustering. One could probably achieve all possible partitions of your data provided a particular choice of clustering algorithm/distance/representation/normalization/projections of your data. $\endgroup$ – mic Sep 5 '15 at 22:37
  • $\begingroup$ If you have sales data, do you also have a customer id? Because then clustering might show which customers bought the same kind products / productgroups. Also first check your time series to see how everything looks before you start modeling. $\endgroup$ – phiver Sep 6 '15 at 7:11
  • $\begingroup$ @phiver Nope. I dont have any personal data like Customer ID or user. Just Store, Article, its Product Group, its Product line, Price, and Amount. $\endgroup$ – Messy Sep 6 '15 at 11:09
  • $\begingroup$ @phiver btw to "check time series" I need first to choose a product group or article. Without doing it, all data is mixed up and its not a time series. For example I have in one product group many articles which were sold more than one time in a month. So this month comes up already 3-4 times in the data set. $\endgroup$ – Messy Sep 6 '15 at 11:24
  • $\begingroup$ At least you can have a look if there are clusters with regards to the stores. Clustering on the products might give you some insight in which products / product colors are likely sold together in certain months. You can also check if there are any seasonal effects in the monthly sales data. $\endgroup$ – phiver Sep 6 '15 at 12:17

Cluster analysis cannot do magic.

It's not as if these tools will just work and yield the desired result.

On contrary, doing cluster analysis right will require many iterations. It is an exploratory technique with many parameters. You need to carefully study the results, then change parameters as to make them more understandable, repeat.

A clustering algorithm may (unless you control parameterd well) divide the products into two clusters:

  • products with an even-length name
  • products with an odd-length name

That is why it's part of explorative data analysis: it's a method for exploring data. Exploration will have may dead ends, and then you need to go back and try something different.

  • $\begingroup$ Yes man, I agree. but think of it in the way of a thesis. You need to explore something interesting. Not only play around with different parameters. My question was, if there can be other clusters. So far the analysis shows me a cluster = a product line. $\endgroup$ – Messy Sep 6 '15 at 19:01
  • $\begingroup$ You can only find other clusterd by experimenting. Usually, cluster analysis orefers the obvious. $\endgroup$ – Anony-Mousse Sep 6 '15 at 20:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.