I was tinkering a bit recently in an effort to build a simple system that finds 'related' podcasts and see if I can see the network effect play out over time.

I was tinkering a bit recently in an effort to build a simple system that finds 'related' podcasts and see if I can see the network effect play out over time. I did this by building a graph of people (hosts/guests) and episodes and started folding in tags/topics. None of this is in my wheelhouse, and I found: - It takes a lot of work to curate a substantial collection of podcasts. There are lists all over the place but it's hard to know what's really in there. - I attmpmted to use SpaCy and/or NLTK to do some 'Named Entity Recognition' in order to extract topics/people/orgnaziations from episode titles and descriptions. This was surprisingly brittle. The string 'Sean Carroll', for example, wasn't detected as a person by either framework (IIRC). It also seems quite brittle to punctuation and other context (e.g. beginning or end of a sentence). This was using the default models shipped with both. I started off with just the english models but expanded as there were lots of names being skipped silently. That helped less than I had hoped. - I have yet to find a good UI for exploring a graph. I used Neo4j and the built in 'browser' is not intended for that purpose. Gephi has good capability for filtering and analytics, but it takes quite a bit of getting used to and the graph itself isn't amenable to dynamic navigation. That's all. Bookmarking this as it would really help.