I am quite new to Data Analytics. I was just wondering whether we can use cluster analysis in Multiple Regression. Let me give you a scenario so that it becomes easier to visualize.

I have a dataset of Property Transactions in the year 2013. The dataset has Property price, Region, Property Area in sq.m, Properties' locational attributes like is it close to bus stop, super market and so on.

Now if I use multiple regression model over here, I can use Price as my dependent variable and other variables as my independent variables and figure out what independent variables cause major influence on Price.

Instead of this approach if I use cluster analysis and figure out which regions have maximum prices and what are the locational attributes that are causing this increase, divide the dataset based on these clusters and then do multiple regression on these datasets to see what regression analysis results I get, will it make sense?

  • 2
    $\begingroup$ It's difficult to know what you're imagining here: "if I use cluster analysis and figure out which regions have maximum prices and what are the locational attributes that are causing this increase". You don't need cluster analysis to see where maximum prices are; in fact it is not clear that cluster analysis would help in that at all. How would the cluster analysis tell you anything about influences that the multiple regression won't? (Very generally, not much in statistics, including multiple regression, tells us about causes; your wording of influences is nearer the mark.) $\endgroup$ – Nick Cox Oct 26 '14 at 10:17
  • 2
    $\begingroup$ I think you misunderstand the purpose of cluster analysis. It will look for properties that are similar across all the variables in your data set. This may tell you something useful about the properties, but (as @NickCox said) you already know which properties have the highest price. Doing multiple regression within each cluster might be useful but it's hard to say. It would be more typical to look at interactions or possibly a regression tree. $\endgroup$ – Peter Flom Oct 26 '14 at 10:45
  • $\begingroup$ It's not easy to just "figure out...what are the locational attributes that are causing this increase." What your technique currently does is reduce price from a continuous variable to a categorical variable (ie a cluster label). you are proposing to first build a classifier for price cluster and then use the features in that classifier to build a continuous prediction model for price. That might not be a bad route but it's probably not great and it's probably not what you really want to do $\endgroup$ – shadowtalker Oct 26 '14 at 13:22

Since you have labeled data, a supervised approach will usually outdo any unsupervised approach.

I agree with Peter Flom, who in the comments noted that you "misunderstand the purpose of cluster analysis". Clusters are not meant to find regions with maximum prices.

Chances are that by partitioning your data into clusters without paying attention to price, your multiple regeression approach will be worse, because it only sees part of the data; and it may have discontinuities at the borders. But in other cases, exactly this can help.

Why not just give it a try and see for yourself? But beware of overfitting, don't increase the number of variables too much.

  • $\begingroup$ Thank you all for your inputs. I am surely going to try it out now. $\endgroup$ – maverick Oct 27 '14 at 15:58

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.