Content-Based Recommendation With Textual Data

ideas:
now the idea is to take the interest points / descriptors and exchange them by the features of items a customer has purchased or that the scraper has found in combination (recommendation ares) and

use the GMM and Fisher Vector to create a “taste” or “combination” fingerprint. than we can use it to check how good an item would fit to a group of items.

local refs:
https://git.picalike.corpex-kunden.de/picalike/personalization
https://git.picalike.corpex-kunden.de/tschulz/complementary-products

refs:
https://www.slideshare.net/ren4yu/image-retrieval-with-fisher-vectors-of-binary-features-miru14
https://arxiv.org/abs/1609.08291
https://lear.inrialpes.fr/people/verbeek/mlcr.slides.14.15/4.FisherVectors.pdf

use cases:
(1) complementary recommendations for all type of products

(2) personalization based on click and/or shop cart. based on click should optimize the recommendations based on what the customer has already looked at during a session (2-12 hours).

if we know the customer (user-session with infinite session lifetime) and we got some purchased item information (shopping cart data), than we should recommend items that could be of interest to the customer based on his preferences in the past.