5 ms·

Ask HN: Bayesian FTW

9 points by Pamar 11y ago

Reading this HN entry (https://news.ycombinator.com/item?id=9780677) has rekindled my interest in Bayesian logic. My main introduction to the topic was through original Stanford AI MOOC. I managed to get decent scores on all the Bayesian-related stuff, but I cannot say I really internalize the ideas. What I am looking for is something that could help me making practical use of Bayes in my day-to-day life (professional and/or personal). So I am basically thinking of: - books or articles detailing practical examples of how to apply Bayesian models to day-to-day choices (if these covered debug and testing activities it would be a big plus) - an easy-to-use app (desktop or tablet/smartphone) to build and play with Bayesian networks models. Of course, if you have anything else to suggest along the same lines, please do.

jrgnsd 11y ago
I find that having a bayes classifier handy can help a lot with categorizing documents or text. I manually keep track of my bank accounts using a spreadsheet, and I started using a classifier to automatically categorize the transaction. It's quite a time saver.
- Pamar 11y ago
  Can you explain what you use as categorization elements?
  - jrgnsd 11y ago
    You train your own set by posting the document and the classification / category to the service. The training set is then used to calculate the prior probabilities needed by Bayes theorem. Once you have a proper training set, you can post a document to the service and it will return a list of probable categories from the training set. Your set might look something like this: Category - Document negative - I don't like ice cream positive - That's an awesome idea neutral - The wind is blowing today positive - We won! You'll get a categorization of either positive, negative or neutral.
    - Pamar 11y ago
      Yes, well - I understand the general theory, but I don't understand how it applies to managing your bank account. I do manage my bank account, but I don't see any need for an "classifier": when I input (or check) a supermarket bill I don't need much help in setting it to "GROCERIES". Basically anything that gets into my accounting files it's either something I know already what is about, or else something I need to investigate (e.g.: a speed ticket from a foreign country, routed to my credit card by the car rental company).
duncanawoods 11y ago
I honestly don't believe Bayes is that relevant to day to day decision making but I'm keen to see a counter-argument (I have some but I don't find them convincing yet). I see Bayes as a method to solve certain classes of problem but the key challenge in everyday decisions is more like a design problem i.e. to define what the problem is. Once you have defined "what would be a good outcome of this decision" sufficiently, the answer rarely requires statistical methods. The common sources of error in a decision are in its definition e.g. omitting a requirement that is later revealed as essential. For example, if you are analysing drug studies, Bayes is obviously relevant but for something more common place such as choosing a software tool, the main challenge is to understand what your goals are so that you can identify the criteria by which to judge the tools and the trade-offs you are willing to accept. The biggest steps to improving decision quality appear to be process related e.g. using prototypes to explore options before making a larger commitment. Such acts are so effective because they reveal information that lets you improve your goals rather than just clarifying the quality of an option.
- Pamar 11y ago
  Well, I wouldn't be surprised if the only practical benefit in building a Bayesian Network model ended up being that "that you had to think long and hard at what the various inputs are and how they should be linked". I am curious if there is any kind of resource that would help me to confidently add Bayes to my normal decision-making toolchest.
  - duncanawoods 11y ago
    Yep, I can see the value as a thinking tool. There are tools like http://www.bayesia.com/ http://www.bayesia.com/ and http://www.hugin.com/ http://www.hugin.com/ but they are enterprise pricey and I find them very clunky. I'd love to know how your ideal tool might work. Do you mind if I drop you an email (from your profile info)? I'm working on a decision making tool to let you quickly describe reasoning with simple bullet point lists. It aims to be as quick as a making a rough note so that with just a few different types of bullet, you can create belief networks and computable decision models. The tool is intended for a general audience so it currently uses simple weighted evidence. If you can tell me how you would like to work with probabilities then you might get the tool you want :)
roller 11y ago
I've had this one stacked up on my reading list for a while: http://camdavidsonpilon.github.io/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers/ http://camdavidsonpilon.github.io/Probabilistic-Programming-... It details how to use the PyMC library via iPython notebooks. I'm not sure if iPython qualifies as an easy to use app though.

9 points by Pamar 11y ago

jrgnsd 11y ago

I find that having a bayes classifier handy can help a lot with categorizing documents or text. I manually keep track of my bank accounts using a spreadsheet, and I started using a classifier to automatically categorize the transaction. It's quite a time saver.

Pamar 11y ago
Can you explain what you use as categorization elements?
- jrgnsd 11y ago
  You train your own set by posting the document and the classification / category to the service. The training set is then used to calculate the prior probabilities needed by Bayes theorem. Once you have a proper training set, you can post a document to the service and it will return a list of probable categories from the training set. Your set might look something like this: Category - Document negative - I don't like ice cream positive - That's an awesome idea neutral - The wind is blowing today positive - We won! You'll get a categorization of either positive, negative or neutral.
  - Pamar 11y ago
    Yes, well - I understand the general theory, but I don't understand how it applies to managing your bank account. I do manage my bank account, but I don't see any need for an "classifier": when I input (or check) a supermarket bill I don't need much help in setting it to "GROCERIES". Basically anything that gets into my accounting files it's either something I know already what is about, or else something I need to investigate (e.g.: a speed ticket from a foreign country, routed to my credit card by the car rental company).

duncanawoods 11y ago

I honestly don't believe Bayes is that relevant to day to day decision making but I'm keen to see a counter-argument (I have some but I don't find them convincing yet). I see Bayes as a method to solve certain classes of problem but the key challenge in everyday decisions is more like a design problem i.e. to define what the problem is. Once you have defined "what would be a good outcome of this decision" sufficiently, the answer rarely requires statistical methods. The common sources of error in a decision are in its definition e.g. omitting a requirement that is later revealed as essential. For example, if you are analysing drug studies, Bayes is obviously relevant but for something more common place such as choosing a software tool, the main challenge is to understand what your goals are so that you can identify the criteria by which to judge the tools and the trade-offs you are willing to accept. The biggest steps to improving decision quality appear to be process related e.g. using prototypes to explore options before making a larger commitment. Such acts are so effective because they reveal information that lets you improve your goals rather than just clarifying the quality of an option.

Pamar 11y ago
Well, I wouldn't be surprised if the only practical benefit in building a Bayesian Network model ended up being that "that you had to think long and hard at what the various inputs are and how they should be linked". I am curious if there is any kind of resource that would help me to confidently add Bayes to my normal decision-making toolchest.
- duncanawoods 11y ago
  Yep, I can see the value as a thinking tool. There are tools like http://www.bayesia.com/ http://www.bayesia.com/ and http://www.hugin.com/ http://www.hugin.com/ but they are enterprise pricey and I find them very clunky. I'd love to know how your ideal tool might work. Do you mind if I drop you an email (from your profile info)? I'm working on a decision making tool to let you quickly describe reasoning with simple bullet point lists. It aims to be as quick as a making a rough note so that with just a few different types of bullet, you can create belief networks and computable decision models. The tool is intended for a general audience so it currently uses simple weighted evidence. If you can tell me how you would like to work with probabilities then you might get the tool you want :)

roller 11y ago

I've had this one stacked up on my reading list for a while: http://camdavidsonpilon.github.io/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers/ http://camdavidsonpilon.github.io/Probabilistic-Programming-... It details how to use the PyMC library via iPython notebooks. I'm not sure if iPython qualifies as an easy to use app though.