I am trying to train a product search query (e-commerce) classifier for deducing probable product categories from search query with a dataset of 700k queries with probable categories labelled
I tried first with multi class but then needed to change to multi-label as some of search queries in dataset had multiple probable categories. now issue with multi label is I'm getting very low accuracy of 30%.
For training I also trained a custom gensim word embedding on a dataset of 8 million products names and 3 million random search queries but results are more or less same.
The training set is in form of 30 categories (columns) with 0 or 1 labelled with search query 1 being true.
What would be the better way to solve this issue?