Thanks for your feedback. Don't hesitate to check our series on Differentiable Programming (see tv.lokad.com/journal/2019/4/10/solving-unsolvable-problems/ ). Since we have released this video, I now consider this approach as the revised state of the art as far as forecasting new products is concerned. Hope it helps! Joannes Vermorel
That was a very insightful video on forecasting new products. For the deep learning model you recently introduced, which one is more important/relevant and giving more accurate results in terms of cross entropy - to provide proper hierarchies with categories, or to provide plain text (label) as attributes, or maybe all three at the same time?
Deep learning - our 5th generation forecasting engine - did indeed provides a nice boost to the extra accuracy obtained from plain text labels, and the best results are obtained by leveraging the two when present. If you had to choose (we did quite a few benchmarks), hierarchies and categories are frequently winning against plain text labels. Indeed, hierarchies and categories - when present - tend to be of high quality for many companies, especially eCommerce, because those are used to let customers efficiently find the products they seek (aka facets). However, for some eCommerce that heavily rely on search to let their clients find the products they seek, the situation is reversed. Hope it helps!