R&D implied a massive historical data analysis which later directed us to the hypothesis, trial & error circle until a fair solution was found. At last, we ended up building an algorithm which uses neural networks and users' navigation data in order to facilitate the handmade categorization. We also built a set of algorithms using deep learning and NLP tools in order to automatically suggest missing categories. Some of the tools used where Gensim, Keras and NLTK.
Added to this, as human validation is part of the solution, we built an ultra intuitive and intelligent frontend using React + Redux, ES6 and Webpack to communicate with the built API.
As it is an ongoing development, we are achieving results on a sprint basis. Up to now, we can argue that the categorization has become much more efficient and we are currently working in the complete automatization of category creation and product classification. Stay tuned.