Data Scientist

Tip 1: one can try ALS based methods which probably can give embeddings for both in the same space.
Tip 2: the problem can be framed as graph, but given that the user / product attributes were not provided if you just use the products and users as nodes then its difficult to scale.
Tip 3: something like prod2vec can be leveraged.

Tip 1: NLP based embedding can be a way of solving this problem. 
Tip 2: Hybrid recommendation approaches are a plus point as they help to solve the cold start problem and also supports good amount of user personalisation 

Tip 1: re-adjust product embeddings using the recency of click. 
Tip 2: hybrid recommenders which supports attribute level embeddings are a plus point

Tip 1: try to use labelled data as less as possible, self-supervised approaches can be leveraged.
Tip 2: try to explain how the embeddings can be used in realtime search (scalability)

Tip 1: emphasise on how to understand the region of interest in the image (use less labelled data)
Tip 2: think about using the image embedding along with the existing product embedding. Do not just concentrate on the image embedding. Because lot of important information is present in the product attributes.