## Where communities thrive

• Join over 1.5M+ people
• Join over 100K+ communities
• Free without limits
##### Activity
Paul Mineiro
@pmineiro

@pmineiro In this case, should we ignore reward signals after a window has expired, or should we still process them trusting that central limit theorem will help us achieve accuracy over time as we observe more events?

I'm not trying to be salty, but there's no CLT issue here. When you update VW, you are saying "for this context i observed this reward". If you do it again, you are saying "i happened to observe the exact same context again but this time i got this other reward". So the best estimate after that is the average of the first and second reward, which is probably not what you want. With respect to time limit, if you define reward as, e.g., "1 if a click within 30 minutes of presentation else 0" then what happens after 30 minutes is irrelevant.

Wes
@wmelton
@pmineiro i dont mind saltiness - just here to learn. I referenced CLT to highlight the same situation as found in the example you gave, at least in my mind. At a sufficiently large sample size, errors in reporting reward with perfect accuracy should regress to the mean over time, correct? I may have wrongly assumed this conclusion given my current understanding that features are “shared” across many users in a given model, so i assumed attribution errors would ultimately more or less tend to the mean given a large enough corpus of events. If Im totally wrong, no sweat haha. Like i said, just here to learn - trying to make the mental leap from a more traditional non-contextual approach to this approach.
Paul Mineiro
@pmineiro
@wmelton using the bandit analogy, if you do 2 vw updates you'll get the equivalent of (n += 2) in the bandit setting. with vw, every time you send in a reward ("c") you get the equivalent of an increment in the number of trials ("n"). so it'll cause you problems.
Max Pagels

@wmelton yeah, just to be clear:

If you have a bernoulli bandit, what some people do is that when an arm is pulled, they record +1 trials and update the posterior, and only when they get a reward for that pull do they update +1 successes. In a context-free setting this is sort of OK and will be kind of eventually consistent. I've done this before, primarily because it saves me from keeping track of pulls that get zero rewards and assigning those explicitly. It isn't "correct", however. In bandit settings you should learn when the reward is available, not do such a half-step. But it's a practical compromise.

In contextual bandits, and in VW, doing this will fail because of the issue @pmineiro mentioned. The way to overcome this is to keep track of all predictions and their context in some DB or memory store and learn only when a reward arrives for a particular prediction/context, or a suitable amount of time has passed such that you can assume zero reward and learn on that.

@wmelton regarding the second question, I've found no flags directly in VW for this. I've made my own system with bootstrapping
Max Pagels

If anyone has any comments on this message I posted I'd be very grateful:

@pmineiro thanks for your patience answering all my questions. I did a quick sanity check: I'd expect explore_eval with 100% exploration against a "world" that never changes and where exactly half of actions are positive (-1 cost) and half negative (+1) would get an estimated average loss of 0, but that's not the case. I'm not sure if this is due to some systemic bias, because in this particular case --cb_explore_adf reports the loss I'd expect. I made an issue but I'm not sure if it's a bug or intended behaviour: VowpalWabbit/vowpal_wabbit#2621

olgavrou
@olgavrou

@maxpagels_twitter : you definitely do not ever run --cb_explore (or --cb_explore_adf) on an offline CB dataset without --explore_eval. you only run --cb_explore either 1) online, i.e., acting in the real-world, 2) offline with a supervised dataset and --cbify (to simulate #1) or 3) offline with --explore_eval and an offline CB dataset (to simulate #1). nothing else is coherent.

Max Pagels
@olgavrou yeah, I already read that and tested explore_eval as suggested, but it gives a loss i wouldn't expect against a uniform random dataset with exactly as much positive and negative feedback. The reported loss is systematically wrong. Which is why I'm wondering if it's a feature of explore_eval or if there is a bug
Paul Mineiro
@pmineiro

In contextual bandits, and in VW, doing this will fail because of the issue @pmineiro mentioned. The way to overcome this is to keep track of all predictions and their context in some DB or memory store and learn only when a reward arrives for a particular prediction/context, or a suitable amount of time has passed such that you can assume zero reward and learn on that.

This join operation is done for you by Azure Personalizer (https://azure.microsoft.com/en-us/services/cognitive-services/personalizer/). We done presentations and workshops at AI NextConn conferences where we show the detailed dataflow diagram, maybe you can find one of those ... or you could just use APS.

Max Pagels

More questions: why, in cb_explore_adf with epsilon set to 0.0, do I se probability distributions with values other than 0.0 or 1.0? This only happens in the start of a dataset:

maxpagels@MacBook-Pro:~$vw --cb_explore_adf test --epsilon 0.0 Num weight bits = 18 learning rate = 0.5 initial_t = 0 power_t = 0.5 using no cache Reading datafile = test num sources = 1 average since example example current current current loss last counter weight label predict features 0.666667 0.666667 1 1.0 known 0:0.333333... 6 0.833333 1.000000 2 2.0 known 1:0.5... 6 0.416667 0.000000 4 4.0 known 2:1... 6 0.208333 0.000000 8 8.0 known 2:1... 6 0.104167 0.000000 16 16.0 known 2:1... 6 0.052083 0.000000 32 32.0 known 2:1... 6 0.026042 0.000000 64 64.0 known 2:1... 6 0.013021 0.000000 128 128.0 known 2:1... 6 0.006510 0.000000 256 256.0 known 2:1... 6 finished run number of examples = 486 weighted example sum = 486.000000 weighted label sum = 0.000000 average loss = 0.003429 total feature number = 4374 maxpagels@MacBook-Pro:~$

All examples have the same number of arms (3), and on different datasets, I see the same thing at the start of a dataset. One large dataset I have takes some 20,000 examples before giving correct probabilities

--first works as expected, but not --epsilon, which at 0.0 exploration should be greedy, ie. the probability vector should have one value of 1.0 and the reset of 0.0.
Max Pagels
Update on the above: apparently if the raw predicted cost for 2 or more arms is exactly the same, tie breaks are done at random, resulting in a probability other than one even though --epsilon 0.0
olgavrou
@olgavrou
Hi @omelyanchikd thanks for bringing this up, this looks like a bug in cover. The distribution (and the resulting prediction) should not be affected by the number of predictions in the test dataset. Will ping you again once some progress is made here.
Diana Omelianchyk
@omelyanchikd
Thank you @olgavrou. I will be looking forward to it. We have decided to go with bagging approach for now.
Wes
@wmelton
@olgavrou can you help me understand how vw treats vectors passed as features? E.g. feature=[0.3,-1.3,...,n] - when using this, vw does not throw an error, yet it is not apparent how vw interprets this. Does it understand it as a vector or does it treat it more like a string and one-hot encode it, or something different entirely?
Allegra Latimer
@alatimer
Hi all--I'm wondering if anyone can point me to papers that demonstrate real-world/industry examples of building a model to filter a large action space down to a reasonable number of actions before using CB to select a recommended action from that subset. I have seen cases where the action space was naturally limited by context, eg by business rules (https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/thompson.pdf) or curation by domain experts (https://arxiv.org/abs/1003.0146), but I haven't been able to find a good example of a hierarchical modeling approach where eg a high-recall recommender system is used to first filter down actions to a manageable subset before applying CB. Any ideas?
Amil Khare
@amil.khare_gitlab
Hey all!
I was wondering if VW can handle cases with one class datasets. Basically finding similar data points based on initial dataset. Is there some way VW can help in such cases?
Allegra Latimer
@alatimer
Hi all, I see CATS is a relatively new algorithm for bandits in continuous action space that uses a tree policy class. Does that mean we can use a tree policy class generally in VW? EG if I am using vw --cb_explore_adf, is there a command line argument to make the policy class be decision trees?
olgavrou
@olgavrou
Hi @alatimer right now that will not be possible. CATS uses the cats_tree reduction under the hood which is a single line reduction. cb_explore_adf is a multi line reduction. Offset_tree was also added though that will use cb_explore (again it is single line so no adf involved here).
2 replies
George Fei
@georgefei
Hi all, I’m wondering if the python wrapper or the cli tool provides a way to output the reward estimation of each arm in the cb/cb_explore mode (I found this post asking the exact same question but it was unanswered: https://stackoverflow.com/questions/60678450/how-to-get-cost-reward-current-estimation-for-all-arms-using-vowpal-wabbit-with)?
Wenjuan Dou
@darlwen
Hi all , I am reading VW source code recently, I am confused about how reduction stack work, to be more detail, how does this setup_base function initialize all enabled reductions?
My train setting is: vw -d debug.txt --foreground --ccb_explore_adf --cb_type mtr --epsilon 0.01 --ftrl -f debug.model, and based on the debug info:
how does all these reductions enabled?
Wenjuan Dou
@darlwen
@olgavrou could you pls help explain it?
Alexey Taymanov
@ataymano

hi @darlwen ,
Stack of reductions for every vw run is defined by 2 things:
1) DAG of dependencies that are defined in setup function for every reduction.
i.e. here:
2) topoligical order here: https://github.com/VowpalWabbit/vowpal_wabbit/blob/b8732ffec3f8c7150dace1c41434bf3cdb4d8436/vowpalwabbit/parse_args.cc#L1246

So, final stack of reduction for each vw run is actually sub-stack from 2) that contains:
1) reductions that you explicitly provided in your command line
2) reductions that defined in input model file (if any)
3) reductions populated as dependencies.

In your case you have ccb_explore_adf, ftrl provided explicitly by you, others are populated as dependencies:

Wenjuan Dou
@darlwen

thanks @ataymano much more clear now. In VW::LEARNER::base_learner* setup_base(options_i& options, vw& all)
when enter the following logic,

 else
{
all.enabled_reductions.push_back(std::get<0>(setup_func));
return base;
}

my understanding is that it won't do auto setup_func = all.reduction_stack.top(); anymore, for example, when we get "ftrl_setup" then it enters the else logic, then how it makes the rest reductions(scorer, ccb_explore_adf etc.) enabled?

2 replies
Kev. Noel
@arita37
thanks for you presntation at Neurips !
Sam Lendle
@lendle
howdy! I'm trying to get a pyvw.vw object to process a data file when I instantiate it with a --data argument. Based on this fairly recent s.o. answer https://stackoverflow.com/a/62876763, my understanding is that it should do just that, but I am not having any luck. I'm using vw version 8.9.0, did something change in a recent release? I have confirmed that using the same options from the command line works so I don't think I'm doing something obviously wrong like using a wrong file name
6 replies
Andrew Clegg
@andrewclegg
what does ring_size do? will increasing this help I/O performance, or is it not something to worry about?
Jack Gerrits
@jackgerrits
Ring_size refers to the initial size of the example pool. It will resize if it needs more room to store parsed examples waiting to be processed by the learner. I wouldn't worry too much about changing it.
Andrew Clegg
@andrewclegg
thanks!
buildvoc
@buildvoc
Was wondering if you could help with a very simple question please note that I am a beginner in VW, someone explain multi class or multi label to me and how it works in VW for example can answer questions like “What’s the predicted price of this house?” or “Should I buy this house today?" but it's more difficult to apply it to a problem like "What type of house is this?" (multiclass) or "What are the adjectives that best describe this house?" (multilabel)
George Fei
@georgefei
hey @jackgerrits could you help answer my question from 2 weeks ago: I’m wondering if the python wrapper or the cli tool provides a way to output the reward estimation of each arm in the cb/cb_explore mode (I found this post asking the exact same question but it was unanswered: https://stackoverflow.com/questions/60678450/how-to-get-cost-reward-current-estimation-for-all-arms-using-vowpal-wabbit-with)?
9 replies
Sam Lendle
@lendle
I have a bunch of questions, mostly about bandits & policy optimization options, not so much about bandits w/ exploration. Questions below in separate messages so responses/discussions can be threaded. In exchange for everyone's patience, I'll update the wiki where I can.
Is my understanding correct that dm, ips, and dr all run the cost sensitive classification reduction with estimated cost for each action as the cost in the csc? The difference between dm, ips, dr is in the estimated costs:
• dm: naively estimates cost w/ regression. Subject to bias due to confounding
• ips: estimate cost as reported cost/probability, or 0 if cost is not reported. (c(a) = cost/probability * I(observed action = a)). Unbiased if probabilities are correct, usually high variance
• dr: doubly robust, uses regression and ips, often lower variance
Is mtr the same as “importance weighted regression” in https://arxiv.org/pdf/1802.04064.pdf?
If so, is the method:
1. Estimate costs w/ regression and 1/probability importance weights. (Essentially the same regression for cost as in the dm method, but with importance weights?)
2. Policy: predict cost for each action, take action w/ lowest predict cost
What does --baseline do with respect to both simple regression and with contextual bandits? What is an example's ‘enabled flag’ referred to in the help string for --check_enabled?
4 replies
Is there a reason MTR isn’t/can’t be implemented for --cb rather than --cb_adf? Related: it’s trivial to manually convert a simple cb example to an adf cb example. I would think that --cb gets internally converted to an adf type problem, but since mtr is not available for --cb, it suggests that is not the case. What else is different between --cb and --cb_adf, when there are not actually any dependent features other than an action indicator?
4 replies

This is the only thing I’ve found that describes the implementation for csoaa: http://users.umiacs.umd.edu/~hal/tmp/multiclassVW.html. As I read it, that means csc based bandit methods:

1. Fit a regression for each action where the target of the regression is the estimated cost of that action
2. Policy: predict cost for each action from the regression models, take action w/ lowest predicted cost
Is that right?

If so, is it reasonable to think of ips and mtr as essentially the same except:

1. IPS uses cost * I(action = observed action)/probability as target and 1 as weight
2. Mtr uses cost as target and I(action = observed action)/probability as weight
How is the progressive validation loss reported by the driver for bandits when not exploring? Is it just the IPS: mean(cost * I(observed action = predicted action) / probability)) or something more sophisticated, like https://arxiv.org/abs/1210?
3 replies
Is there an --eval type option for --cb_adf? How is the data specified?
2 replies
Are there plans to have OPE estimators ala https://github.com/VowpalWabbit/estimators available in mainline vw so they can be called from outside python? (context: I'm a binding author and would love access to this)
2 replies
Hi all. I wrote up a short little post about vw 8.9.0 mostly for folks newer to vw but thought I'd share here. I'd love any feedback or corrections, I'm sure I've missed plenty. https://travisbrady.github.io/posts/vowpal-8.9.0/
Jack Gerrits
@jackgerrits
Thanks for sharing Travis! Love the post! (And thanks for the call out :) ) By the way, I haven’t forgotten about supporting bindings, it’s been a bit lower on the todo list for a while though. We’ve started work on a new C binding interface that provides a more complete view at VW functionality and actually reports errors how a C api should.
3 replies
Sam Lendle
@lendle
For posterity, here's a recording of the NeurIPS 2020 presentation: https://slideslive.com/38942331/vowpal-wabbit, which I discovered in Travis's post above but haven't been able to find elsewehre. Great presentation, it helped clear up many of the questions I had last week. Thanks!
Jack Gerrits
@jackgerrits
We also just put up a blog post that contains timestamps to each section in case anyone is looking for anything in particular: https://vowpalwabbit.org/blog/neurips2020.html
Lena Gangan