I am thinking more of auto categorization. "programming languages," "art/design", "lulz," "gaming," etc. But certainly with "interesting to me" as a component of each category.
No, I wouldn't. And I wouldn't use it, even if it were free.
The current system of labelling via filters gives me fine-grained control from the very first e-mail. Machine learning will require a (possibly long) training period, and may still then be prone to false positives and negatives.
But what I would use (might pay for) is a service that makes my attachments easily findable. An attachment browser, basically. The UI would have to be within Gmail, no external page. Don't put the emails central, put the attachments central. Order by date received. Group by received from same people/same filename. Optional grouping by type (ie: All | Images | Office docs | ...) Add a left-hand link "Attachments".
How is this different creating labels and filtering your emails into various labels you created?
That is what I have set up. If it pertains to bill, money or bank stuff the sender i.e. support@bankofamerica.com goes into bank & money label. That is just an example as I have 15 or more labels where email goes accordingly.
Beware of building a business around an obvious missing feature to an existing product. Especially if said product has an active team that's constantly churning out new features, and even goes as far as to have a "labs" section.
So yeah, if you did this for Outlook instead, you might have a viable business. For Gmail, you run a very real risk that your feature may simply appear for free inside of Gmail itself one day, leaving you out in the cold.
Bayesian filtering is based on word frequency, or in sophisticated forms, Markov analysis.
I'm not confident it could reliably discriminate between useful categories. It'd be more likely to pick up on individuals' different vocabularies than useful categories such as client companies, contracts or projects.
To be useful, It'd have to have > 95% accuracy (post training). I'm just not sure a Bayesian model could do that reliably.
Unless, of course, it heavily weighted features such as "sender domain", but by that point you might as well be using explicit filters.
No, I already use automated labeling based on sender, and the amount of mail that doesn't fall under these filters is generally 1) small in volume, and 2) spam.
This sounds like a whole lot of work for something that already works well enough. In my experience, Bayesian filtering works well, until it doesn't. I then find myself frustrated because the filter is essentially a black box to me. I could keep trying to train it, but I already have two children and three pets. My patience for teaching things to a computer is usually nonexistent.
This idea has a larger than usual chance of being crushed by competition.
Normally I wouldn't be bothered about the risk of Google/Microsoft stepping into my market. (In fact, Microsoft are already in my market.)
But in this case:
- They have access to more data than you. A lot more.
- The product is bang in the middle of their core competence. They love automated data analysis. One of their engineers would probably hack up a competitor at the weekend for fun.
- They already have a low-friction way of trying out ideas with Gmail (labs) which would help counter the usual large-but-slow competitor disadvantage.
I've been considering building things on top of Gmail, but because of the above I have decided that if I do it'll be limited to products that Google doesn't seem to like. (And/or hobby projects for my own amusement.)
16 comments
[ 2.9 ms ] story [ 51.1 ms ] threadThe current system of labelling via filters gives me fine-grained control from the very first e-mail. Machine learning will require a (possibly long) training period, and may still then be prone to false positives and negatives.
That's how it always seems to go with these things.
Something like this: http://poorbuthappy.com/ease/archives/2010/05/14/4728/can-so...
That is what I have set up. If it pertains to bill, money or bank stuff the sender i.e. support@bankofamerica.com goes into bank & money label. That is just an example as I have 15 or more labels where email goes accordingly.
So yeah, if you did this for Outlook instead, you might have a viable business. For Gmail, you run a very real risk that your feature may simply appear for free inside of Gmail itself one day, leaving you out in the cold.
I'm not confident it could reliably discriminate between useful categories. It'd be more likely to pick up on individuals' different vocabularies than useful categories such as client companies, contracts or projects.
To be useful, It'd have to have > 95% accuracy (post training). I'm just not sure a Bayesian model could do that reliably.
Unless, of course, it heavily weighted features such as "sender domain", but by that point you might as well be using explicit filters.
This sounds like a whole lot of work for something that already works well enough. In my experience, Bayesian filtering works well, until it doesn't. I then find myself frustrated because the filter is essentially a black box to me. I could keep trying to train it, but I already have two children and three pets. My patience for teaching things to a computer is usually nonexistent.
Normally I wouldn't be bothered about the risk of Google/Microsoft stepping into my market. (In fact, Microsoft are already in my market.)
But in this case:
- They have access to more data than you. A lot more.
- The product is bang in the middle of their core competence. They love automated data analysis. One of their engineers would probably hack up a competitor at the weekend for fun.
- They already have a low-friction way of trying out ideas with Gmail (labs) which would help counter the usual large-but-slow competitor disadvantage.
I've been considering building things on top of Gmail, but because of the above I have decided that if I do it'll be limited to products that Google doesn't seem to like. (And/or hobby projects for my own amusement.)