32
$\begingroup$

My set-up is the following:

I am following the guide lines in "Applied Predictive Modeling". Thus I have filtered correlated features and end up with the following:

  • 4900 data points in the training set and 1600 data points in the test set.
  • I have 26 features and the target is a continuous variable.

I apply 5-fold cross-validation to train models using the caret package. When I apply a MARS model then I get a mean absolute error (MAE) of approximately 4 on the training set as well as on the test set.

However applying XGBgboost (either the tree algorithm or the linear one) I get something like 0.32 (!) on the training set and 2.4 on the test set.

Thus if the test error is 8 times the training error then I would say: I have overfit the training data. Still I get a smaller error on test anyways.

I use the following parameters on xgboost:

  • nrounds = 1000 and eta = 0.01 (increasing nrounds and decreasing eta could help but I run out of memory and run time is too long)
  • max_depth = 16: if I compare other posts and the default of 6 then this looks large but the problem is pretty complex - maybe 16 is not too large in this case.
  • colsample_bytree = 0.7,subsample = 0.8 and min_child_weight = 5: doing this I try to reduce overfit.

If I reduce max_depth then train and test-error get closer but still there is a large gap and the test-error is larger (a bit above 3).

Using the linear booster I get the roughly the same train and test error on optimal parameters:

  • lambda = 90 and `alpha = 0: found by cross-validation, lambda should prevent overfit.
  • colsample_bytree = 0.8,subsample = 0.8 and min_child_weight = 5: doing this I try to reduce overfit.

My feeling is that XGBoost still overfits - but the training error and as far as I can see in the real time test (I have used the XGBoost models and an ensemble of them in reality for 4 days) looks ok-ish (the error is larger than the test error but there are is more uncertainty in real life about the forecast of features and other variables).

What do you think: can I accept overfit if (if this is possible) real life performance is superior? Does XGBoost in my setting tend to overfit?

$\endgroup$

1 Answer 1

52
$\begingroup$

Is overfitting so bad that you should not pick a model that does overfit, even though its test error is smaller? No. But you should have a justification for choosing it.

This behavior is not restricted to XGBoost. It is a common thread among all machine learning techniques; finding the right tradeoff between underfitting and overfitting. The formal definition is the Bias-variance tradeoff (Wikipedia).

The bias-variance tradeoff

The following is a simplification of the Bias-variance tradeoff, to help justify the choice of your model.

  • We say that a model has a high bias if it is not able to fully use the information in the data. It is too reliant on general information, such as the most frequent case, the mean of the response, or few powerful features. Bias can come from wrong assumptions, for exemple assuming that the variables are Normally distributed or that the model is linear.

  • We say that a model has high variance if it is using too much information from the the data. It relies on information that is revelant only in the training set that has been presented to it, which does not generalize well enough. Typically, the model will change a lot if you change the training set, hence the "high variance" name.

Those definition are very similar to the definitions of underfitting and overfitting. However, those definition are often too simplified to be opposites, as in

  • The model is underfitting if both the training and test error are high. This means that the model is too simple.
  • The model is overfitting if the test error is higher than the training error. This means that the model is too complex.

Those simplifications are of course helpful, as they help choosing the right complexity of the model. But they overlook an important point, the fact that (almost) every model has both a bias and a variance component. The underfitting/overfitting description tell you that you have too much bias/too much variance, but you (almost) always have both.

If you want more information about the bias-variance tradeoff, they are a lot of helpful visualisation and good ressource available through google. Every machine learning textbook will have a section on the bias-variance tradeoff, here are a few

  • An introduction to statistical learning and Elements of statistical learning (available here).
  • Pattern Recognition and Machine Learning, by Christopher Bishop.
  • Machine Learning: A Probabilistic Perspective, by Kevin Murphy.

Also, a nice blog post that helped me grasp is Scott Fortmann-Roe's Understanding the Bias-Variance Tradeoff.

Application to your problem

So you have two models,

$$ \begin{array}{lrrl} & \text{Train MAE} & \text{Test MAE} &\\ \text{MARS} & \sim4.0 & \sim4.0 & \text{Low variance, higher bias},\\ \text{XGBoost} & \sim0.3 & \sim2.4 & \text{Higher variance, lower bias},\\ \end{array} $$

and you need to pick one. To do so, you need to define what is a better model. The parameters that should be included in your decisions are the complexity and the performance of the model.

  • How many "units" of complexity are you willing to exchange for one "unit" of performance?
    • More complexity is associated with higher variance. If you want your model to generalize well on a dataset that is a little bit different than the one you have trained on, you should aim for less complexity.
    • If you want a model that you can understand easily, you can do so at the cost of performance by reducing the complexity of the model.
    • If you are aiming for the best performance on a dataset that you know comes from the same generative process than your training set, you can manipulate complexity in order to optimize your test error and use this as a metric. This happens when your training set is randomly sampled from a larger set, and your model will be applied on this set. This is the case in most Kaggle competitions, for exemple.

The goal here is not to find a model that "does not overfit". It is to find the model that has the best bias-variance tradeoff. In this case, I would argue that the reduction in bias accomplished by the XGBoost model is good enough to justify the increase in variance.

What can you do

However, you can probably do better by tuning the hyperparameters.

  • Increasing the number of rounds and reducing the learning rate is a possibility. Something that is "weird" about gradient boosting is that running it well past the point where the training error has hit zero seems to still improve the test error (as discussed here: Is Deeper Better Only When Shallow Is Good?). You can try to train your model a little bit longer on your dataset once you have set the other parameters,

  • The depth of the trees you grow is a very good place to start. You have to note that for every one unit of depth, you double the number of leafs to be constructed. If you were to grow trees of size two instead of size 16, it would take $1/2^{14}$ of the time! You should try growing more smaller trees. The reason why is that the depth of the tree should represent the degree of feature interaction. This may be jargon, but if your features have a degree of interaction of 3 (Roughly: A combination of 4 features is not more powerful than a combination of 3 of those feature + the fourth), then growing trees of size larger than 3 is detrimental. Two trees of depth three will have more generalization power than one tree of depth four. This is a rather complicated concept and I will not go into it right now, but you can check this collection of papers for a start. Also, note that deep trees lead to high variance!

  • Using subsampling, known as bagging, is great to reduce variance. If your individual trees have a high variance, bagging will average the trees and the average has less variance than individual trees. If, after tuning the depth of your trees, you still encounter high variance, try to increase subsampling (that is, reduce the fraction of data used). Subsampling of the feature space also achieves this goal.

$\endgroup$
3
  • 2
    $\begingroup$ One should not comment saying "thanks" but for this long and interesting answer I would like to say "thank you". I aleady had some knowledge about some of the things you write but this was really nicely put together. I will go through some of your references and grow new trees and maybe come back with a comment. For now: thanks! Great! $\endgroup$
    – Richi W
    Commented Mar 30, 2016 at 10:37
  • $\begingroup$ The link to the interactions page stat.columbia.edu/~jakulin/Int is really great! $\endgroup$
    – Richi W
    Commented Apr 7, 2016 at 8:52
  • $\begingroup$ I took your advice and limited the depth of the trees to 3 but took nrounds 1200 and the resuls feels great: very quick computations, difference between train and test reduced and still on a good level. The rest of the story is here: stats.stackexchange.com/questions/205858/… $\endgroup$
    – Richi W
    Commented Apr 11, 2016 at 14:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.