Gavin Potter at IntroAnalytics has written a white paper titled Making Behavioral Recommendations Pay, An Analysis of Returns Achieved. I am pleased to see some of the early results discussed publicly. While not exactly earth-shattering in terms of depth, I applaud Gavin for making the results available and urge him to continue to update. In fact, IntroAnalytics should follow OKCupid and blog when they have enough data to work with. That would be outstanding. A few tidbits of information below.
- There has been a substantial increase in activity on all sites which have implemented the IntroAnalytics technology. Detailed analysis of site traffic indicates that the introduction of the IntroAnalytics product leads to an average 20% increase in site traffic.
- Detailed analysis of site activity shows that most users prefer to view profiles based on the selections provided by the IntroAnalytics technology rather than the traditional search built into the sites.
- One surprising result is that the technology has a greater impact on female activity – especially messaging activity than on male activity.
Those wishing to read the full paper should us the contact form or email info@introanalytics.com.
No related posts.
{ 6 comments… read them below or add one }
If anyone is interested in the full whitepaper … feel free to contact us on the site.
At the moment, we are running a free trial for any dating site or social network.
Nick Tsinonis
CEO, introAnalytics.com
Recommendation Systems Ltd
Which is exactly the range of that recommendation engine invented by IntroAnalytics?
e.g: In a big database as the one at Match, it recommends you X persons per 1,000 persons screened.
How IntroAnalytics can demonstrate its recommendation engine can achieve that range?
Fernando, one of IntroAnalytics’ founder is Gavin Potter.
He did pioneering work on movie recommendations using psychology, maths, machine learning for the Netflix Prize. See Wired article: http://www.wired.com/techbiz/media/magazine/16-03/mf_netflix
The algorithms have evolved significantly and are constantly changing and improving as science and machine learning advances. So usage “behavioural analytics” can range from the a simple user analysis to cutting edge artificial intelligence.
I wouldn’t be surprised if Match wasn’t using some kind of behavioral algorithm but not sure on the details. Perhaps Dave can do some investigative journalism and let us know.
Our major addition to the dating industry is that we provide this technology as a SaaS (Software as a service). So any dating site or social network can easily plug this service into their site without hiring a team of Phds.
Hi Nick!
By my own experience I know that the proprietary Bidirectional Recommendation Engine (Matching based on Self-Reported Data by personal preferences & likes and dislikes) actually in use at Match (“the Daily5″at USA site) is in the range of 3 to 4 persons recommended per 1,000 persons screened, in other words any member receives on average 3 or 4 prospective mates as recommended for dating purposes per 1,000 (one thousand) members screened in Match’s big database.
Can you contact Dr. Potter and ask him:
Which is exactly the range of the Bidirectional Recommendation Engine offered by IntroAnalytics as a SaaS?
To avoid misunderstanding, there are 3 possible answers:
1) Far better than the one used at Match because the one offered by IntroAnalytics can reduce false positive recommended prospects; e.g. in the range of 1.5 to 2 persons recommended per 1,000 persons screened ( 15 to 20 recommended per 10,000 screened )
2) No improvement, in exactly the same range as the one used at Match ( 30 to 40 recommended per 10,000 screened )
3) Much worse than the one used at Match because the one offered by IntroAnalytics really increases false positive recommended prospects; e.g. in the range of 6 to 8 persons recommended per 1,000 persons screened ( 60 to 80 recommended per 10,000 screened )
Please, do not think I am rude or not polite, but: Do you really know the range of the Bidirectional Recommendation Engine you offer or not?
Thanks.
Fernando
Nick has asked me to respond to your query. It is little difficult to make a comparison with Match based on the data that you have provided. This is because we have built a behaviorally adaptive system and the number of recommendations made will depend on how successful previous recommendations have been.
In our system, every person is ranked in terms of how successful we believer the recommendation is likely to be. The number of recommendations made, will however, depend on the success of the recommendations. If a particular person selects and communicates to many of the recommendations then we will present more to that individual and vice versa.
The key difference between our system and others that we are aware of, is that we are trying to drive the system from user behaviour. This means that the presentation of the recommendations will depend on how receptive that person is to receiving the recommendations. Other systems tend to impose arbitrary limits on the number of recommendations received (either by providing a fixed number of recommendations or only providing recommendations that exceed a set threshold etc). By creating an adaptive system we believe that we can better match the individual users requirements and therefore provide a better overall user experience of a dating site.
Therefore we can’t say how many recommendations are made per person because it will vary depending on the success of the recommendations.
From a technical perspective we are adjusting not only the ROC curve but the threshold used – so single point comparisons with Match or any other site are not possible. Proper comparisons would require full ROC/AUC data from each site so that meaningful statistical comparisons can be made.
I hope that helps.
Hello Gavin:
Many online dating sites had been using Behavioural Bidirectional Recommendation Engines for years, like PlentyOfFish, and they could not outperform compatibility Matching Methods based on personality profiling.
It seems there is a range convergence phenomenon between the 3 mains tools online dating sites can offer: searching by your own, Bidirectional Recommendation Engines and Compatibility Matching Methods. Any member receives on average 3 or 4 prospective mates as selected / recommended / compatible for dating purposes per 1,000 (one thousand) members screened in the database.
They all 3 are performing the same for serious daters, with a high percentage of false positives, like gun machines shooting flowers.
I would respectfully suggest you to obtain/construct an average range indicator for the online dating sites using your Behavioural Bidirectional Recommendation Engine (like the ones powered by White Label Dating or yesnomayB, smooch, Singles365, etc ), or roughly estimate that range to see if your engine can outperform the 3 to 4 persons recommended -on average- per 1,000 persons screened.
Regards.