Model with target as feature for benchmark I was reading this article (https://towardsdatascience.com/predicting-the-popularity-of-instagram-posts-deeb7dc27a8f) about predicting the popularity of Instagram post using different techniques and one of the steps the author takes is to use a XGBoost model with the target variable as a feature to get a benchmark score.
Anyone can explain me what is the point of this approach? Is it to have a guess about the maximum score a model without the target variable could get?
Thank you for your answers!
 A: This approach is not standard and not explained adequately in the post. In anything, if we used the target variable as a predictor as well as response in a adequately regularised linear model we should get: $R^2 = 1$. We do not get that (for reasons not explained in the post). 
Response variables might be used as explanatory variables within the context of list experiments analysis, a particular questionnaire analysis methodology, but this is not the case here. (See for example, Imai et al. 2015 Using the Predicted Responses from List Experiments as Explanatory Variables in Regression Models)
In general, it is a good practice to have a basic model to serve as a benchmark. This approach is particular popular in forecasting (e.g. when using metrics like MASE when the benchmark model is the one-step ahead forecast) and can be reasonably used in standard ML applications too. For example, when publishing a paper on a well-research area it is expected to benchmark a new model against known popular/good solutions. For personal use, it makes sense to use a simple model (e.g. a standard regularised regression approach using LASSO) to have an idea if we are actually doing better or just treading water. Notice that this is not a maximum score but rather a benchmark to surpass. 
