6Tribes was a new aged social media app, that put user interests at the heart of connectivity; rather than who you know, it’s about what you know and what you want to know.
Lead Data Scientist + User Researcher | 6Tribes | 2016 | Acquired to become DriveTribe
To handle the cold-start problem, we decided to lean on the giant of social media – Facebook. By using Facebook, we were able to allow users to provide us with permission for photos, locations, current groups and recent posts. After some research and user testing, we discovered Facebook profile data was aged and not very relevant.
The data is seemingly updated as a one time process, whilst the user creates an account. Although ironic, to use Facebook data as we tried to build a social media platform, there is a large degree of enrichment and unlocked potential in the data.
Use of Natural language techniques alongside data enrichment using API services and data collection. This was an iterative process developing across all sources to optimize our recommendations.
NLP techniques for longer content included bag of words, n-grams and TF-IDF. For shorter content we used the Rapid Automatic Keyword Extraction (RAKE) along with cleaning processes.
Evolved into Topic Extraction, using LDA, based from the theory each document has a multinomial distribution of topics, matched to a multinomial distribution of words. The documents can assumed to be a mix, assuming the nature in content variety that is posted in Facebook, to therefore generate content.
- A main limitation of topical modelling is the topics are extracted words. We needed to manually create a mapping ontop of both topics and the RAKE extractor, to interests themselves.For example. If someone expresses interest in football, cricket, tennis then this implies a hobby of sport, as well as specific hobbies.
Geo Data from photos alongside geographical information and landmarks can indicate interests.
Posts highlight key interests through keywords, links, events and locations through additional analysis.
Groups provide key themes for extraction from their subject and posts.
Across apps that present recommendations, users are often intrigued as to why an item has been recommended, even furious if a recommendation is incorrect. With Social DNA, we wanted to present users back their transformed data in a meaningful way hence did this via profile tags of their inferred interests. Users could then additionally delete and add tags, which allowed a feedback loop for algorithmic improvements.
Most apps make it easy to share assets and invite users via a array of platforms. However, with the use of Facebook, we were able to make contact invitation more intelligent. Through using Facebook we can see the connected friends whom has liked a post. If this is a repeated occurrence, this creates a weak signal they may also have this as a common interest. Hence the 6Tribes user, upon wanting to share can have more intelligent invitee recommendations. The invitee then becomes more willing to join, as the content itself appears relevant to them.
From collated interests we recommended Tribes. These were formed based on interest weighting, alongside business and marketing objectives to appease to our growing user base. We matched using ElasticSearch matching and creating custom scoring based on both extracted interests. Tribes also contain keywords evolved in a similar process we’ve discussed above, allowing us to reinforce and evolve based on member postings.
To ensure we had a good onboarding process for users, we additionally created Trending Tribes, by analyzing frequency of posts and replies, per Tribe.
Just like Facebook and all those other apps but awesomer –
I love this app so much I’ve gotten to meet some really cool people from here and their all amazing :D It’s really cool how you can findsomeone who has similar interest in things as you by making a tribe or joiningone.