To Regression or To Classify

Hi Team
Im investigating an opportunity to value used vehicle. I have data about the vehicles make, model, mileage, features, condition and i have the “Y” the price these cars sold for.

To me this is an obvious regression problem which suggests i use a neural network hidden layers using RELU activation and final layer with linear activation, using a loss function like squared error

However would there be benefit instead converting this problem to a classification problem?
that is, creating buckets to possible selling prices and classifying the car into the correct selling bucket?
Then using softmax, and SparseCategoricalCrossentropy

Another angle is that the training data i have has been collected over a 3 year period during which time the market has gone through various dips and peaks. So would it be better to look at this as a time series model?

I’m NOT looking for a “correct” answer here. Rather id appreciate any discussion this topic generated from the community…

Hi @darthShana

great question and great application!

So, one guiding point to start with according to CRISP-DM… is business understanding: You can ask yourself: E.g. if you provide a prediction for a range of fair prices (using the available prices as labels) this could be interesting for the end user who either wants to buy or wants to sell a vehicle and your price prediction (preferably with confidence intervals) would help the user to reduce uncertainty on the end user‘s side. This speaks in favour of modelling it as regression problem, (e.g. with a probabilistic model if you want to provide uncertainty or confidence estimates.)

In conclusion: I do not see how a classification problem could contribute to solve a users problem or contribute to business understanding in this example.

There are other examples where things might be not so clear: let’s assume you want to predict how a certain device or machine is doing. Here you would have the opportunity (assuming sufficient data and labels are available):

  • to model a remaining useful life (regression problem)
  • to model a multi-class problem (e.g. with classes: normal, failure1, failure 2, …) if you have also failure labels which are often hard to get (classification)
  • to model an anomaly detection in an unsupervised way, e.g. with an autoencoder for example, if you have tons of normal data only

I would suggest to let your business understanding considering also technical boundaries guide you in your problem definition.

Note also dependent on your designed system, regression, classification and unsupervised methods can be combined technically, see also this application here.

Hope that helps! Please let me know what you think.

Best regards

1 Like

Concur with @Christian_Simonis above on 1) problem drives solution and 2) it doesn’t have to be a single model or output layer.

I can imagine a regression to produce a prediction for current market price, but another model that classifies it as below/at/above-market or looks at recent trends for that make/model/year/condition. These could either be entirely separate use cases or treated as an ensemble to produce a recommended bid price. Ie you might want to help a user bid less than true current market price if recent prices are trending downward.

1 Like