3 ms·
Fun fact: We did this exact analysis at reddit many years ago, and used it to figure out which subreddits were related to each other. We never got around to p
by jedberg 12y ago
Fun fact: We did this exact analysis at reddit many years ago, and used it to figure out which subreddits were related to each other. We never got around to productizing it, unfortunately, but the idea was to use it to suggest new reddits to you.
- sinemetu11 12y agoI guess this might get into some special sauce territory, but was there a specific reason why this type of recommendation system was deprecated?
- Houshalter 12y agoAll the reddit subreddit recommenders I've seen produce garbage recommendations. Outside of a handful of popular, general subreddits which everyone already knows about, everything is niche special interest stuff that you need to find on your own.
- bduerst 12y agoSubreddit specificity is so messily complex that it would be very difficult to do any recommendations based on your own subscriptions. Without reddit's cooperation in categorization (unlikely) it's probably not going to happen.
- jedberg 12y agoAs others already said below, the feedback loop was dangerous. Also it took a lot of resources to calculate and we just didn't have the time to build efficient map/reduce jobs to do it regularly. It was done by hand in Mathmatica.
- swalsh 12y agoOne potential downside of using an algorithm like that is the possibility of a feedback loop.
- eevilspock 12y agoThe very difficult trick is to find a virtuous feedback loop. Positive feedback loops might appear to be virtuous, but without balancing mechanisms they encourage viral dynamics (when did the virus become out role model for success?), group think and gaming. Google's PageRank is an example of an incestuous positive feedback loop. Web pages with many inbound links get ranked higher, and pages that rank higher get more inbound links (because that's how many people find things they link to).