Description of classification problem:
Assume a regular dataset X with n samples and d features.
This classification problem is somewhat hard (many features, few samples, low overall AUC ~70%).
It might be useful to mention that feature selection/extraction, dimension reduction, kernels, many classifiers have been applied. So I am not interested in trying these.
I am not looking forward to see an improvement in overall AUC. The goal is to find relevant features in haystack of features.
Description of my approach:
I select all pairwise combination of d features and create many two dimensional sub-datasets x with n samples.
On each sub-dataset x, I perform a 10-fold cross-validation (using all samples of the main dataset X). A very long process, assume weeks of computation.
I select top k pairs (according to highest AUC for example) and label them as +. All other pairs are labeled as -.
For each pair, I can compute several properties (e.g. relations between each pair using Expert's knowledge). These properties can be calculated without using the labels in main dataset X.
Now I have pairs which are labeled as + or -. In addition, each pair has many properties calculated based on Expert's knowledge (i.e. features). Hence, I have a new classification problem. Lets call this newly generated dataset Y.
I train a classifier on Y while following cross-validation rules. Surprisingly, I can predict the + and - labels with 90% AUC.
As far as I can see, it means that I am able to select relevant features. However, seeing a 90% AUC makes me worried about information leakage somewhere in this long process. Specially in step 3.
I was wondering if anyone can see any leakage in this approach.
Information Leakage:
Incorporation of target labels in the actual features. Your classifier will produce good prediction while did not learn anything.
Showing your test set to you classifier during the training phase. Your classifier will "memorize" the test set and its corresponding labels without "learning" anything.
Update 1:
I want to stress that indeed I am using all data points of X in step 1. However, I am not using them ever again (even for testing). The final 90% AUC is obtained from predicting labels of dataset Y.
On the other hand, it would be useful to note that, even if I randomize the values of my main dataset X, the computed features for dataset Y is going to be the same. However, the sample labels in Y would change because the previous + pairs might not be a good one anymore. Therefore they will be labeled as -.
Update 2:
Although I haven't got any opinion, I am going to state what I have got during 4 days of talking with pattern recognition researchers. Briefly I became confident that there is no information leakage (as long as I wont go back to the first dataset X and using its labels). Later on, in case I wanted to check to see if I could have better performance in X (i.e. predicting sample labels), I need to use only a part of dataset X for pairwise comparison (as training set). Then I can use the rest of samples in X as test set while using positively predicted pairs of Y as features.
I will set this as an answer in case no one could reject this method.
If your processes in step 1 uses all data. then the features you are learning have information from the whole data set. Since you selected based on the whole dataset and THEN validation, you are leaking serious information.
You should probably stick with tools that are well known / already done for you before running out and trying weird strategies like this. Try using a model with L1 regularization to do feature selection for your, or start with some of the simpler searches like Sequential Backward Selection.
If you do cross validation correctly in the end, each training will perform its own independent feature selection. If you do one global feature selection and then do CV, you are going to be doing it wrong and probably leaking information.
Related
I'm fairly new to data analysis and machine learning. I've been carrying out some KNN classification analysis on a breast cancer dataset in python's sklearn module. I have the following code which attemps to find the optimal k for classification of a target variable.
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from sklearn.neighbors import KNeighborsClassifier
import matplotlib.pyplot as plt
breast_cancer_data = load_breast_cancer()
training_data, validation_data, training_labels, validation_labels = train_test_split(breast_cancer_data.data, breast_cancer_data.target, test_size = 0.2, random_state = 40)
results = []
for k in range(1,101):
classifier = KNeighborsClassifier(n_neighbors = k)
classifier.fit(training_data, training_labels)
results.append(classifier.score(validation_data, validation_labels))
k_list = range(1,101)
plt.plot(k_list, results)
plt.ylim(0.85,0.99)
plt.xlabel("k")
plt.ylabel("Accuracy")
plt.title("Breast Cancer Classifier Accuracy")
plt.show()
The code loops through 1 to 100 and generates 100 KNN models with 'k' set to incremental values in the range 1 to 100. The performance of each of those models is saved to a list and a plot is generated showing 'k' on the x-axis and model performance on the y-axis.
The problem I have is that when I change the random_state parameter when spliting the data into training and testing partitions this results in completely different plots indicating varying model performance for different 'k'values for different dataset partitions.
For me this makes it difficult to decide which 'k' is optimal as the algorithm performs differently for different 'k's using different random states. Surely this doesn't mean that, for this particular dataset, 'k' is arbitrary? Can anyone help shed some light on this?
Thanks in anticipation
This is completely expected. When you do the train-test-split, you are effectively sampling from your original population. This means that when you fit a model, any statistic (such as a model parameter estimate, or a model score) will it self be a sample estimate taken from some distribution. What you really want is a confidence interval around this score and the easiest way to get that is to repeat the sampling and remeasure the score.
But you have to be very careful how you do this. Here are some robust options:
1. Cross Validation
The most common solution to this problem is to use k-fold cross-validation. In order not to confuse this k with the k from knn I'm going to use a capital for cross-validation (but bear in mind this is not normal nomenclature) This is a scheme to do the suggestion above but without a target leak. Instead of creating many splits at random, you split the data into K parts (called folds). You then train K models each time on K-1 folds of the data leaving aside a different fold as your test set each time. Now each model is independent and without a target leak. It turns out that the mean of whatever success score you use from these K models on their K separate test sets is a good estimate for the performance of training a model with those hyperparameters on the whole set. So now you should get a more stable score for each of your different values of k (small k for knn) and you can choose a final k this way.
Some extra notes:
Accuracy is a bad measure for classification performance. Look at scores like precision vs recall or AUROC or f1.
Don't try program CV yourself, use sklearns GridSearchCV
If you are doing any preprocessing on your data that calculates some sort of state using the data, that needs to be done on only the training data in each fold. For example if you are scaling your data you can't include the test data when you do the scaling. You need to fit (and transform) the scaler on the training data and then use that same scaler to transform on your test data (don't fit again). To get this to work in CV you need to use sklearn Pipelines. This is very important, make sure you understand it.
You might get more stability if you stratify your train-test-split based on the output class. See the stratify argument on train_test_split.
Note the CV is the industry standard and that's what you should do, but there are other options:
2. Bootstrapping
You can read about this in detail in introduction to statistical learning section 5.2 (pg 187) with examples in section 5.3.4.
The idea is to take you training set and draw a random sample from it with replacement. This means you end up with some repeated records. You take this new training set, train and model and then score it on the records that didn't make it into the bootstrapped sample (often called out-of-bag samples). You repeat this process multiple times. You can now get a distribution of your score (e.g. accuracy) which you can use to choose your hyper-parameter rather than just the point estimate you were using before.
3. Making sure you test set is representative of your validation set
Jeremy Howard has a very interesting suggestion on how to calibrate your validation set to be a good representation of your test set. You only need to watch about 5 minutes from where that link starts. The idea is to split into three sets (which you should be doing anyway to choose a hyper parameter like k), train a bunch of very different but simple quick models on your train set and then score them on both your validation and test set. It is OK to use the test set here because these aren't real models that will influence your final model. Then plot the validation scores vs the test scores. They should fall roughly on a straight line (the y=x line). If they do, this means the validation set and test set are both either good or bad, i.e. performance in the validation set is representative of performance in the test set. If they don't fall on this straight line, it means the model scores you get from you validation set are not indicative of the score you'll get on unseen data and thus you can't use that split to train a sensible model.
4. Get a larger data set
This is obviously not very practical for your situation but I thought I'd mention it for completeness. As your sample size increases, your standard error drops (i.e. you can get tighter bounds on your confidence intervals). But you'll need more training and more test data. While you might not have access to that here, it's worth keeping in mind for real world situations where you can assess the trade-off of the cost of gathering new data vs the desired accuracy in assessing your model performance (and probably the performance itself too).
This "behavior" is to be expected. Of course you get different results, when training and test is split differently.
You can approach the problem statistically, by repeating each 'k' several times with new train-validation-splits. Then take the median performance for each k. Or even better: look at the performance distribution and the median. A narrow performance distribution for a given 'k' is also a good sign that the 'k' is chosen well.
Afterwards you can use the test set to test your model
while reading the paper :" Tactile-based active object discrimination and target object search in an unknown workspace", there is something that I just can not understand:
The paper is about finding object's position and other properties using only tactile information. In the section 4.1.2, the author says that he uses GPR to guide the exploratory process and in section 4.1.4 he describes how he trained his GPR:
Using the example from the section 4.1.2, the input is (x,z) and the ouput y.
Whenever there is a contact, the coresponding y-value is stored.
This procedure is repeated several times.
This trained GPR is used to estimate the next exploring point, which is the point where the variance is maximum at.
In the following link, you also can see the demonstration: https://www.youtube.com/watch?v=ZiLq3i-BJcA&t=177s . In the first part of video (0:24-0:29), the first initalization takes place where the robot samples 4 times. Then in the next 25 seconds, the robot explores explores from the corresponding direction. I do not understand how this tiny initialization of GPR can guide the exploratory process. Could someone please explain how the input points (x,z) from the first exploring part could be estimated?
Any regression algorithm simply maps the input (x,z) to an output y in some way unique to the specific algorithm. For a new input (x0,z0) the algorithm will likely predict something very close to the true output y0 if many data points similar to this was included in the training. If only training data was available in a vastly different region, the predictions will likely be very bad.
GPR includes a measure of confidence of the predictions, namely the variance. The variance will naturally be very high in regions where no training data has been seen before and low very close to already seen data points. If the 'experiment' takes much longer than evaluating the Gaussian Process, you can use the Gaussian Process fit to make sure you sample regions where you are very uncertain of your answer.
If the goal is to fully explore the entire input space, you could draw a lot of random values of (x,z) and evaluate the variance at these values. Then you could perform the costly experiment at the input point where you are most uncertain in y. Then you can retrain the GPR with all the explored data so far and repeat the process.
For optimization problems (Not the OP's question)
If you wish to find the lowest value of y across the input space, you are not interested in doing the experiment in regions that you know give high values of y, but you are just uncertain of how high these values will be. So instead of choosing the (x,z) points with the highest variance, you might choose the predicted value of y plus one standard deviation. Minimizing values this way is named Bayesian Optimization and this specific scheme is named Upper Confidence Bound (UCB). Expected Improvement (EI) - the probability of improving the previously best score - is also commonly used.
I used the caret package to train a random forest, including repeated cross-validation. I’d like to know whether the OOB, as in the original RF by Breiman, is used or whether this is replaced by the cross-validation. If it is replaced, do I have the same advantages as described in Breiman 2001, like increased accuracy by reducing the correlation between input data? As OOB is drawn with replacement and CV is drawn without replacement, are both procedures comparable? What is the OOB estimate of error rate (based on CV)?
How are the trees grown? Is CART used?
As this is my first thread, please let me know if you need more details. Many thanks in advance.
There are a lot of basic questions here and you would be better served by reading a book on machine learning or predictive modeling. Thats probably why you haven't gotten much of a response.
For caret you should also consult the package website where some of these questions are answered.
Here are some notes:
CV and OOB estimation for RF are somewhat different. This post might help explain how. For this application, the OOB rate from random forest is computed while the model is being build whereas CV uses holdout samples that are predicted after the random forest model is computed.
The original random forest model (used here) uses unpruned CART trees. Again, this is in many text books and papers.
Max
I recently got a little confused with this too, but reading chapter 4 in Applied Predictive Modeling by Max Kuhn helped me to understand the difference.
If you use randomForest in R, you grow a number of decision trees by sampling N cases with replacement (N is the number of cases in the training set). You then sample m variables at each node where m is less than the number of predictors. Each tree is then grown fully and terminal nodes are assigned to a class based on the mode of cases in that node. New cases are classified by sending them down all the trees and then taking a vote; the majority vote wins.
The key points to note here are:
how the trees are grown - sampling WITH replacement (a bootstrap). This means that some cases will be represented many times in your bootstrap sample and others may not be represented at all. The bootstrap sample will be the same size as your training dataset.
The cases that are not selected for building trees are referred to as the OOB samples- an OOB error estimate is calculated by classifying the cases that aren't selected when building a tree. About 63% of the data points in the bootstrap sample are represented at least once.
If you use caret in R, you will normally use caret::train(....) and specify the method as "rf" and trControl="repeatedcv". You can change trControl to "oob" if you want out of the bag. The way this works is as follows (I'm going to use a simple example of a 10 fold cv repeated 5 times): the training dataset is split into 10 folds of roughly equal size, a number of trees will be built using only 9 samples - so omitting the 1st fold (which is held out). The held out sample is predicted by running the cases through the trees and used to estimate performance measures. The first subset is returned to the training set and the procedure repeats with the 2nd subset held out, and so on. The process is repeated 10 times. This whole procedure can be repeated multiple times (in my example, I do this 5 times); for each of the 5 runs, the training dataset with be split into 10 slightly different folds. It should be noted that 50 different held out samples are used to calculate model efficacy.
The key points to note are:
this involves sampling WITHOUT replacement - you split the training data and build a model on 9 samples and predict the held out sample (the remaining 1 sample of the 10) and repeat this process as above
the model is built using a dataset that is smaller than the training dataset; this is different to the bootstrap method discussed above
You are using 2 different resampling techniques which will yield different results therefore they are not comparable. The k fold repeated cv tends to have low bias (for k large); where k is 2 or 3, bias is high and comparable to the bootstrap method. K fold cv tends to have high variance though...
I have been working through the concepts of principal component analysis in R.
I am comfortable with applying PCA to a (say, labeled) dataset and ultimately extracting out the most interesting first few principal components as numeric variables from my matrix.
The ultimate question is, in a sense, now what? Most of the reading I've come across on PCA immediately halts after the computations are done, especially with regards to machine learning. Pardon my hyperbole, but I feel as if everyone agrees that the technique is useful, but nobody wants to actually use it after they do it.
More specifically, here's my real question:
I respect that principle components are linear combinations of the variables you started with. So, how does this transformed data play a role in supervised machine learning? How could someone ever use PCA as a way to reduce dimensionality of a dataset, and THEN, use these components with a supervised learner, say, SVM?
I'm absolutely confused about what happens to our labels. Once we are in eigenspace, great. But I don't see any way to continue to move forward with machine learning if this transformation blows apart our concept of classification (unless there's some linear combination of "Yes" or "No" I haven't come across!)
Please step in and set me straight if you have the time and wherewithal. Thanks in advance.
Old question, but I don't think it's been satisfactorily answered (and I just landed here myself through Google). I found myself in your same shoes and had to hunt down the answer myself.
The goal of PCA is to represent your data X in an orthonormal basis W; the coordinates of your data in this new basis is Z, as expressed below:
Because of orthonormality, we can invert W simply by transposing it and write:
Now to reduce dimensionality, let's pick some number of components k < p. Assuming our basis vectors in W are ordered from largest to smallest (i.e., eigenvector corresponding to the largest eigenvalue is first, etc.), this amounts to simply keeping the first k columns of W.
Now we have a k dimensional representation of our training data X. Now you run some supervised classifier using the new features in Z.
The key is to realize that W is in some sense a canonical transformation from our space of p features down to a space of k features (or at least the best transformation we could find using our training data). Thus, we can hit our test data with the same W transformation, resulting in a k-dimensional set of test features:
We can now use the same classifier trained on the k-dimensional representation of our training data to make predictions on the k-dimensional representation of our test data:
The point of going through this whole procedure is because you may have thousands of features, but (1) not all of them are going to have a meaningful signal and (2) your supervised learning method may be far too complex to train on the full feature set (either it would take too long or your computer wouldn't have a enough memory to process the calculations). PCA allows you to dramatically reduce the number of features it takes to represent your data without eliminating features of your data that truly add value.
After you have used PCA on a portion of your data to compute the transformation matrix, you apply that matrix to each of your data points before submitting them to your classifier.
This is useful when the intrinsic dimensionality of your data is much smaller than the number of components and the gain in performance you get during classification is worth the loss in accuracy and the cost of PCA. Also, keep in mind the limitations of PCA:
In performing a linear transformation, you implicitly assume that all components are expressed in equivalent units.
Beyond variance, PCA is blind to the structure of your data. It may very well happen that the data splits along low-variance dimensions. In that case, the classifier won't learn from transformed data.
I'm starting to use LIBSVM for regression analysis. My world has about 20 features and thousands to millions of training samples.
I'm curious about two things:
Is there a metric that indicates the accuracy or confidence of the model, perhaps in the .model file or elsewhere?
How can I determine whether or not a feature is significant? E.g., if I'm trying to predict body weight as a function of height, shoulder width, gender and hair color, I might discover that hair color is not a significant feature in predicting weight. Is that reflected in the .model file, or is there some way to find out?
libSVM calculates p-values for test points based upon the certainty of the classifier (i.e., how far is the test point from the decision boundary and how wide are the margins).
I think you should consider the determination of feature importance a separate problem from training your SVMs. There are tons of approaches for "feature selection" (just open any text book) but one easy to understand, straightforward approach would be a simple cross-validation as follows:
Divide your dataset into k folds (e.g., k = 10 is common)
For each of the k folds:
Separate your data into train/test sets (the current fold is the test set, the rest are the training set)
Train your SVM classifier using only n-1 of your n features
Measure the prediction performance
Average the performance of your n-1 feature classifier for all k test folds
Repeat 1-3 for all remaining features
You could also do the reverse where you test each of the n features separately but you will likely miss out on important second and higher order interactions between the features.
In general, however, SVMs are good at ignoring irrelevant features.
You may also want to try and visualize your data using Principal Components Analysis to get a feel for how the data is distributed.
The F-score is a metric commonly used for features selection in Machine Learning.
Since version 3.0, LIBSVM library includes a directory called tools. In that directory is a python script called fselect.py, which calculates F-score. To use it, just execute from the command line and pass in the file comprised of training data (and optionally a testing data file).
python fselect.py data_training data_testing
The output is comprised of an fscore for each of the features in your data set which corresponds to the importance of that feature to the model result (regression score).