2
Bratabase recommendations are weird
It recommends bras too small or too big for some reason. I remeasured after I outgrew my CK Florence in 32D (its band was altered), and BTB recommended Florence in 30D for me.
My boobs got bigger so I remeasured today and BTB recommended Florence in 30F and 30DD to me. And it's the only CK model it recommends to me. There are other brands, ofc, I'm just familiar with the fit of this particular model and it makes me confused. How do BTB algorithms work?
Filed under Bratabase
4 comments
I also don't understand recommendations. I wear 28 bands, and even though some are a bit too tight, most of them are fine. But bratabase recommends 30 and 32 bands :\
Edit: Well, to be fair, I went and checked the ribcage measurements of people who wore 28 bands, and they are all smaller than me. So, apparently I'm the outlier.
Welp, I think I'm the adult in charge of answering this. This may get a bit long.
To be honest, it is very hard even for me to predict what the algorithm will throw out. One critical difference is that unlike everything else online, this algorithm doesn't obtain the results from an arithmetic formula. This will match ranges of your measurements to the ranges of all the other measurements of what's on the database. So it has the fatal flaw of learning from what people enter here.
The way that the final result is obtained is by aggregating together the results of a bunch of "channels", each channel will implement a different algorithm on different set of measurements and find its own results. Then we put all these results together each with a confidence value and sort the first batch.
There are also what I call "anti channels" which are results that should be excluded from the final recommendations, and here is where there may be need to improvement based on what you describe. The antichannels work similarly, there's a few of them, each with different algorithm and get aggregated with a confidence level.
Here's a quick overview of the channels:
https://www.bratabase.com/help/recommendations/
There are a couple of other corrective charts where the results get contrasted with, some "global measurements per size" matrices that get updated with all the measurements as well.
Something very backwards about this approach that's been bugging me the last couple of years is that, not all channels are always enabled. Each gets activated only if the user has provided enough data for this channel's algorithm to work (must have set fit on one bra, or must have indicated that at least one bra doesn't fit, must have entered such and such measurements on at least one bra and so on). So when more data is added, more channels are enabled, this means that the final confidence level is divided by a bigger denominator (number of active channels) making the cut thresholds much more sensible. As if there are less channels enabled, the thresholds are bigger and you may get even more results. So adding more data could pay off by giving less results (because it can be more accurate). It is a similar story to when less people vote and the bad candidates get stronger results.
Overall, I think the general approach is very worth it, but a bunch of things need to be tuned and there are a few dozen of knobs in there, and some more to be added, for example band comfort and what each may accept as tight or loose.
When you look at your results page, hovering on each result will show a small flap indicating the confidence level from which channel got it to be included, this is a hint if you think something is wrong, it is possible to use this to investigate a bit further, usually the biggest offender for false positives is "Based on profile measurements", this means that someone else with similar measurements indicates the bra fits. And when only one person said this, the signal gets confused with being stronger. Adding a minimum of input to normalize the signal ends up excluding too much of the database since we'd need to have at least that many minimun of users that reported that a particular bra model and size fit, and it is only a handful of such combination that have a high enough number of entries, which kills the discovery factor.
Speaking of discovery, sometimes I feel that people would be much more satisfied if the results included less information and just do like every other calculator out there and only tell you "30F" and don't try to indicate which specific models. I think that this is the biggest cause of this engine having such bad fame, it tries more and gets judged differently, saying "30F" it claimed so little that it could not be judged if any bra in 30F doesn't fit. But trying to be more accurate it is much more falsifiable/verifiable.
Another aspect of this more precise judgement ability, is that for many users with good data, the results will not be surprising, so when things go well, they are not worthy of raising any eyebrows, but when you suddenly get a 32A in the list, then the full results get judged because of the bad sheep. The expectations get much higher and errors are much less forgiven.
Thanks for coming to my Ted talk :)
Thanks for adding your data to the site :)
This thread has 4 comments. Log in to read them