I believe that as your user base and daily message posting rate in your forum climbs higher than the average user will read in a day, the noise to signal ratio of the forum starts to increase quickly with the result that more active forums see a continual drop-off of engagement and return users (as measured in Daily Active Users / Total Registered Users. Facebook had this very quickly as their message rate climbed, and I think I’m seeing it too.
One solution I think, would be to address it as facebook did with their Edgerank algorithm - to start customizing the discussion feed for each user, so that the topics and people that are of most interest to a user are shown prioritized above those message topics of less interest.
Would you like to see this feature in your forum?
Yes - More relevant discussion topics for users is very valuable
What you’re not factoring in here is the cost of development. This is a very, very difficult feature to get right – if you had to estimate the number of man-years and money Facebook had spent on it, I’d say it would be in the tens of millions of dollars if not hundreds of millions of dollars.
This is “let’s build successful AI” level difficult. Not quite put a man on the moon, but not terribly far off, either.
(It is much “easier” for Google since people literally type in what they want into search, e.g “cute cats”, but even then look at the size of Google and the number of years they’ve spent on the problem, even starting with an explicit statement from the user where the user typed in exactly what they wanted.)
I think what I forgot was to ask people how many messages a day / and users they have in their forum before they answer this question. This feature is really only an issue for higher volume forums - where the number of messages significantly exceed the average daily messages read by the user base.
The current customization is really very minimal in my mind.
And Jeff,
The good news is that while Facebook first implemented their feed algorithm back in 2009 when there was little published in this area - there has been a lot of research, and a lot of people have built up expertise in this area during the past 7 years - so getting a minimally effective algorithm seems like it could be much easier.
and (in case you missed my other posting)
And, I’m sure you’re an Eric Reis fan - it seems you could start with some sort of minimum viable algorithm. I suspect (correct me if I’m wrong) that any simple algorithm that even just takes into consideration 1 factor - the titles of the topics that the user has “liked” - would be better than just a chronological firehose feed of all the data / messages / topics.
What I’m saying is the following:
Investment could start low with a minimum viable algorithm using a contract data scientist (Stanford and Berkeley have lots of these people, and if you combine the data from all your hosted forums, you have some good data to evaluate). Perhaps get some PHD students for cheap at UC Berkeley to do this for free or near free as a PHD thesis.
Constrain the number of variables initially - just try to get a prioritization of the Topics (not the individual messages - since really each topic is one conversation).
I may be wrong on this - but it seems that virtually any effort at prioritization is better than no prioritization (other than time).
I think you’re quite severely wrong. What if my “simple predictor of what BCHK will like” algorithm happens to pick 5 things you dislike, and puts them at the top of your list?
That would be considerably worse than a simple chronological list of topics, or the “top” topics in a given timeframe.
Yes, as an analogy, if I usually visit the Baking forum’s Bread category, how would the forum know when I had company coming and was instead interested in Cake that particular day?
I much prefer Settings than choice hi-jacking or attempting to read my oft fickle mind.
Jeff - I agree, you could be completely right on this.
the issue of course, is how much work/time/effort it would take to develop a minimally viable algorithm that is better than current time based feed. You are in the best position to decide this - I would just encourage you to grab lunch with a data scientist friend or aquaintance down the street at Berkeley sometime soon and get his or her opinion.
Stumbled here from a Google search. My community is growing quite fast and I can see the “news feed” problem arising from a distance.
What if there was a way for the platform to use relationships instead of content to decide what to show first?
For example, if John reads all posts from Mary, has chatted with Mary and has liked several of her posts, then clearly John tend to prefer content from Mary.
What I’m proposing here is to build a “friendship score” for each user.
If the friendship score is above a certain treshold, the post shows on top. If not, then it show the post chronologically.
The Facebook algorithm has 4 big factors: the person who post and the relationship with that person, the popularity/engagement of the content, the type of content and the users preference toward content type, time decay.
I’m sure this is insanely more complex than what I’ve just written out, but a really dumbed down version using only two of these factors (relationship + time) could be a huge step forward for Discourse.
I know it’s been a year, but I’m curious to see if anyone has any other/simpler ideas.
One other aspect would be to just offer an additional view, next to latest, categories, unread, etc… which is like a facebook news feed but completely raw, i.e. a backwards chronological microblog feed of all activity on the site, kind of like the activity stream buddypress offers
The reason I would find this something worth exploring is twofold:
a) lots of average users would any platfom outside social networks to catch up and offers similar behavior
b) it might trigger more activity by displaying more activity (i.e. you see users joined, moderators doing things, people creating posts, folks earning badges etc)
