Creating similarity metric with Doc2Vec and additional features

Question

I have a dataset which contains many features. Each record is company that has many features.

For example...

Company A:

• Keywords - data, big data, tableau, dashboards, etc.

• Industry - Information Technology

• Sub-Industry - Data Visualization

• Total Funding - $150,000,000

I want to create a similarity metric between multiple companies, incorporating both doc2vec embeddings trained on the keyword lists as well as the additional features listed. I had a hard time searching/finding papers that did something like this. Any ideas?

Answer 1

You could think of your similarity measure as a search problem if you consider one record a query, and the "near" records as search results.

I've had some good results following this paper: https://arxiv.org/pdf/1602.01137.pdf

As I understand it, the document vectors used in the paper were only good for improving the relevance of search results for results that were already decent (the top N results).

That to me suggests you might try developing a similarity score that works with your other attributes first, and then do something like a weighted average, where the significance of the doc2vec score decays quickly based on the first metric.