How can these models be validated? about mta HOT 1 CLOSED

Thank you @peterbayerle for interesting questions. Let's see..

• All attribution models are built using real collected data, i.e. you know exactly customer journeys and whether or not they converted (see the included sample data). When putting together a model you apply some reasoning to figure out each channel's contribution so that it agrees with the observed conversions and quits. Would this model be valid? Well, yeah, by design and for this data.
• What those authors are saying about classification, i.e. ".. high degree of accuracy in correctly classifying.." is about one of the models they describe in the paper, not about MTA models in general. Bad wording. What they mean though has to do with using logistic regression. The idea is that channels should be able to "predict" if a customers will convert or quit. Makes sense, you hope that channels do affect customer's decision and attribution does exist when you do attribution modelling. So you say aha, we can easily see a classification problem here, and fit a logistic regression, in this case, to the available data. Then you grab the fitted weights and use them to calculate each channel's contribution. Naturally, you'd like this logistic regression model to show a high accuracy, otherwise some would say that your channel contributions are garbage because they were derived without much agreement with the observed conversions/quits (prediction accuracy was bad!).

from mta.