Hey guys, I would like to know if anyone found an appropriate way of calculating feature importance after training a model?
I tried using sklearn/eli5 permutation methods but neither properly worked.
Then, I decided to code my own, where I first train a model, then do permutation importance in a held out set. I'm a bit concerned as to whether the results are significant, mainly because of all interactions created on the fly with VW. I should mention I am aware of the multilinearity/correlation problem, and this is not my biggest concern.
Does it even make sense to calculate feature importance in VW? (I assume so because this was one of the topics from the VW presentation at: https://slideslive.com/38942331/vowpal-wabbit
Thanks!
I tried using sklearn/eli5 permutation methods but neither properly worked.
Then, I decided to code my own, where I first train a model, then do permutation importance in a held out set. I'm a bit concerned as to whether the results are significant, mainly because of all interactions created on the fly with VW. I should mention I am aware of the multilinearity/correlation problem, and this is not my biggest concern.
Does it even make sense to calculate feature importance in VW? (I assume so because this was one of the topics from the VW presentation at: https://slideslive.com/38942331/vowpal-wabbit
Thanks!
@CP500 cats pdf should give you a pdf (probability density function) and not a pmf (probability mass function) as cats is predicting in a continuous action space. PDF is in the form of (left:right:pdf_value) triples so you could check that the pdf integrates to 1 by doing (left - right) pdf_value + (left - right) pdf_value for all the returned triples
7 replies
@Favoreto_B_twitter that repo is very much still a work in progress - so docs/examples are not there yet unfortunately. For CCB the approach that has been taken so far is to do CFE on the first slot only - so in this context I think it does make sense to use it. But it would need some adapting and I am not positive here.
9 replies
I'm fiddling around with CATs, and have a simple setup with a fixed context. Per round, I ask for an action (range 0-100) and calculate a cost that is zero at 50, otherwise quadratically the absolute distance from 50 in either direction. If tried gridsearching a whole mess of bandwith, epsilon and learning rate values, but the learning is just all over the place. I would have expected the system to converge to an optimal prediction of 50.0 per round pretty easily since the context is always fixed. Instead, it either bounces around or gets stuck on some non-optimal values around 40. Any tips?
Hi @maxpagels_twitter what is the parameter you pass to
--cats? have you experimented with that at all? For cats I would try different combinations of number of discrete actions used by the algorithm (passed in to the --cats arg) and bandwidths (bandwidth being a property of the continuous range). e.g. I would try a grid of num_actions [8, 16, 32, 64, 128, 256, 1024] and e.g. bandwidths [1, 2, 4, 6, 8, 10, 14, 20]. For different number of discrete actions you might need more data for CATS to converge to something sensible. CATS label support in pyvw should be available in the next release (coming soon-ish, we don't want to wait another year for the next vw release). Let me know if you get better results from CATS or not :)
I tried gridsearching a whole mess of options, including a bunch of action counts, and can get relatively close to an optimum, but the hyperparams seem to be super important to get just right or the learning is way off. But I'll experiment further and report back
3 replies
Hello guys!
Can someone help me understand why propensities scores are important when training a CB?
I've been thinking about this lately and just couldn't wrap my head around a good explanation...
Let's take the epsilon-greedy, for example. When we train a CB model with epsilon-greedy, the pmf output is always the same (just the indexes change). This makes me assume that propensities scores aren't supposed to teach the CB how to output probabilities. Moreover, I believe they are used for "importance weighting", i.e.,
Also, when we offline train a new model using logged CB data, how is the new CB able to achieve better performance than the logging policy? I mean, it's an excellent thing, but I would like to understand how that is possible.
Thanks!
Can someone help me understand why propensities scores are important when training a CB?
I've been thinking about this lately and just couldn't wrap my head around a good explanation...
Let's take the epsilon-greedy, for example. When we train a CB model with epsilon-greedy, the pmf output is always the same (just the indexes change). This makes me assume that propensities scores aren't supposed to teach the CB how to output probabilities. Moreover, I believe they are used for "importance weighting", i.e.,
prob(new_policy)/prob(logging_policy), but isn't this only for when we use IPS? I think I'm missing something quite obvious here...Also, when we offline train a new model using logged CB data, how is the new CB able to achieve better performance than the logging policy? I mean, it's an excellent thing, but I would like to understand how that is possible.
Thanks!
Hi all, I have a few questions related to contextual bandit evaluation:
1.How do I compare the performance of different policies’ decisions using --eval? Do I look at the average loss in the output? If the costs in the input data are all negative and a lower cost is better, does a lower average loss mean one policy is better? What does average loss represent?
1.How do I compare the performance of different policies’ decisions using --eval? Do I look at the average loss in the output? If the costs in the input data are all negative and a lower cost is better, does a lower average loss mean one policy is better? What does average loss represent?
2.How do I interpret the output of --explore_eval? More specifically update count, violation count, and final multiplier (what variables do they correspond to in the algorithm on slide 9 of https://pdfs.semanticscholar.org/presentation/f2c3/d41ef70df24b68884a5c826f0a4b48f17095.pdf). Do I also look at the average loss to compare different exploration algo + hyperparameter combinations?
3.In order to use -explore_eval I have to convert my data from cb format to cb_adf format since the cb format is not supported when using -explore_eval. For the example data with two arms below, are the two ways to represent the data equivalent?:
2:10.02:0.5 | x0:0.47 x1:0.84 x2:0.29
1:8.90:0.5 | x0:0.51 x1:0.65 x2:0.67shared | x0:0.47 x1:0.84 x2:0.29
| a1
0:10.02:0.5 | a2shared | x0:0.51 x1:0.65 x2:0.67
0:8.90:0.5 | a1
| a2
My question is this - within a given namespace, does the order of features or their names matter? Im assuming yes, but the VW documentation out there doesnt make it super clear how to handle a situation where two given documents have the same keywords in them, but after tokenization, the keywords are not in the same order due to variance in the number of keywords found in each document. Appreciate guidance there.
finally, i referenced Personalizer only because it sparked this train of thought largely because the documentation for it leverages only rhe JSON format of input data, but seems to neglect any instruction with regards to variation in keyword order if your features are keywords extracted from a document. Thanks!
18 replies
Particularly, I imagine lots of folks do evaluation by gridsearching learning with
cb_type ips/dm/dr and choosing the one with the best reported loss. Isn't that wrong, especially considering dm is biased? --eval throws an error if you use DM.
49 replies
OPE PR: VowpalWabbit/vowpalwabbit.github.io#193
@lalo @olgavrou et al, Note that I will need expert advice on this. There is a checklist that needs to be confirmed to absolute certainty or, if untrue, commented on to provide me with the correct interpretation.
Anyone here have experience using CBs in VW with a "no op" arm that by definition can't generate a reward?
For example, imagine a use case where we intend to potentially contact a user automatically. so arms are "email user", "send text message to user", "send push notification to user" or "do nothing". The first 3 options all have "click through" as the reward, but "do nothing" of course has no such obvious reward. Is there a standard way to handle this? All pointers much appreciated.
For example, imagine a use case where we intend to potentially contact a user automatically. so arms are "email user", "send text message to user", "send push notification to user" or "do nothing". The first 3 options all have "click through" as the reward, but "do nothing" of course has no such obvious reward. Is there a standard way to handle this? All pointers much appreciated.
Hi everyone! I’m currently using VWClassifier to predict binary labels (-1, 1) on some dummy dataset where y = sigmoid(X@w) for some random X and w. I am able to use pyvw to fit perfectly to the training dataset when I use VWClassifier without any regularization, but I’m noticing strange behavior once I add in regularization. For example, when I add in l1 regularization of 1e-3, all of my training and testing labels get pushed to 1, and the ROC_AUC score between the predicted and actual labels are 0.5 for both training and test. When compared to sklearn packages SGDClassifier and LogisticRegression, I get vastly different results—the labels do not get pushed to 1, and the ROC_AUC score are all >0.5 when I compare the predicted outcome to the actual outcome. Here is the code I'm running, and any help would be greatly appreciated! Thanks :)
Hi everyone! Does someone have a working python example of new VW's feature "CATS, CATS pdf for Continuous Actions"?
I'm really new to RL and this library. Can someone help me to understand how to process data with this new Bandit?
My task is the following. I have prices and I need to find optimal price (not big not low), my reward is a click rate.
Appreciate any help )
I'm really new to RL and this library. Can someone help me to understand how to process data with this new Bandit?
My task is the following. I have prices and I need to find optimal price (not big not low), my reward is a click rate.
Appreciate any help )
Olga created a tutorial Jupyter notebook which has not yet been merged but it is a great resource VowpalWabbit/jupyter-notebooks#6
4 replies
Hi all, I had a problem when using SquareCB algorithm to train contextual bandit model, especially when saving & loading it again.
I trained and saved SquareCB model in this way (using the simulation setting as in https://vowpalwabbit.org/tutorials/cb_simulation.html):
vw = pyvw.vw("--cb_explore_adf -q UA -f squarecb.model --save_resume --quiet --squarecb")
num_iterations = 5000
ctr = run_simulation(vw, num_iterations, users, times_of_day, actions, get_cost)
plot_ctr(num_iterations, ctr)
vw.finish()and then loaded the model :
vw_loaded=pyvw.vw('--cb_explore_adf -q UA -i squarecb.model')
num_iterations = 5000
ctr = run_simulation(vw_loaded, num_iterations, users, times_of_day, actions, get_cost, do_learn=False)
plot_ctr(num_iterations, ctr)
print(ctr[-1])and the loaded model seems like doing a random exploration.
Could anyone explain how to save and load this model correctly? Thanks in advance.
8 replies
Can anyone here help me understand how the bagging algorithm does counterfactual learning from logged bandit data? From the bagging algorithm in the bake-off paper, we can see it reduces to an oracle where the probability of choosing an action a is decided by the proportion of the current policies that evaluate action a as optimial. So how will the probability in the logged data be used? I am perplexed here.
Hey guys!
I was reading Microsoft's Personalizer docs and wondering why changing the Reward wait time triggers retraining? I understand that there might be delay-related bias. I don't understand how retraining the model is any better than simply using the existing model but with an updated reward wait time? Because in my mind, offline training doesn't suffer from delay-related bias (because all data is there, ready for use).
Does this make sense? What am I missing?
What happens if the user decides to add/modify attributes? Is there retraining or simply a matter of changing the Rank input?
Thanks!
I was reading Microsoft's Personalizer docs and wondering why changing the Reward wait time triggers retraining? I understand that there might be delay-related bias. I don't understand how retraining the model is any better than simply using the existing model but with an updated reward wait time? Because in my mind, offline training doesn't suffer from delay-related bias (because all data is there, ready for use).
Does this make sense? What am I missing?
What happens if the user decides to add/modify attributes? Is there retraining or simply a matter of changing the Rank input?
Thanks!
Now, if you change the wait time to 2 minutes, the training data must be recreated so as to include your click, which means the model too must be retrained otherwise you are training on different definitions of reward (old stuff has rewards calculated with 1 minute cutoff, newer with 2 minute cutoff). This leads to general weirdness
If you add features, at least in VW-land, you don't need to retrain but just continue training on new data. So in personaliser I think it's just a matter of changing the rank input
of course, if you add informative features later on, they will only be recorded for events after the change, but in an online system that doesn't really matter since it will correct itself over time
OPE tutorial is up for those interested: https://vowpalwabbit.org/tutorials/off_policy_evaluation.html