I am training an object detection model for multi-class objects in the image. The dataset is custom collected and labelled data with bounding boxes and class labels in the ground truth data.
I trained the MobileNet+SSD , SqueezeDet and YoloV3 networks with this custom data but get poor results. The rationale of choosing these models is their fast performance and light weight (low memory foot print). Their single shot detector approach is shown to perform well in literature as well.
The class instance distribution in the dataset is as below
Class 1 -- 2469
Class 2 -- 5660
Class 3 -- 7614
Class 4 -- 13253
Class 5 -- 35262
Each image can have objects from any of the five classes. Class 4 and 5 have very high incidence.
The performance is very skewed with high recall scores and Average Precision for the class 4 and 5 , and an order of magnitude difference (lower) for the other 3 classes.
I have tried fine tuning on different filtering parameters , NMS threshold, model training parameters to no avail.
Question,
How to tackle such class imbalance to boost the detection Average precision and object detection accuracy for all classes in object detection models. ?
Low precision means your model is suffering from false positives. So you can try hard negative mining. Run your model. Find False positives. Include them in your training data. You can even try using only false negatives as false examples.
As you expect another way can be collecting more data if possible.
If it is not possible you may consider adding synthetic data. (i.e. change brightness of image, or view point(multiply with a matrix so it looks stretched))
One last thing may be having data for each class i.e. 5k for each.
PS: Keep in mind that flexibility of your model has a great impact. So be aware of over fitting under fitting.
In generating your synthetic data as mentioned by previous author, do not apply illumination or viewpoint variations..etc to all your dataset but rather, randomly. The number of classes is also way off, and will be best to either limit the numbers or gather more datasets for those classes. You could also try applying class weights to penalize the over representing classes more. You are making alot of assumptions that simple experimentation will yield results that could surprise you. Remember deep learning is part science and alot of art.
Related
I would like to build a GBM model with H2O. My data set is imbalanced, so I am using the balance_classes parameter. For grid search (parameter tuning) I would like to use 5-fold cross validation. I am wondering how H2O deals with class balancing in that case. Will only the training folds be rebalanced? I want to be sure the test-fold is not rebalanced.
In class imbalance settings, artificially balancing the test/validation set does not make any sense: these sets must remain realistic, i.e. you want to test your classifier performance in the real world setting, where, say, the negative class will include the 99% of the samples, in order to see how well your model will do in predicting the 1% positive class of interest without too many false positives. Artificially inflating the minority class or reducing the majority one will lead to performance metrics that are unrealistic, bearing no real relation to the real world problem you are trying to solve.
For corroboration, here is Max Kuhn, creator of the caret R package and co-author of the (highly recommended) Applied Predictive Modelling textbook, in Chapter 11: Subsampling For Class Imbalances of the caret ebook:
You would never want to artificially balance the test set; its class frequencies should be in-line with what one would see “in the wild”.
Re-balancing makes sense only in the training set, so as to prevent the classifier from simply and naively classifying all instances as negative for a perceived accuracy of 99%.
Hence, you can rest assured that in the setting you describe the rebalancing takes action only for the training set/folds.
A way to force balancing is using a weight columns to use different weights for different classes, in H2O weights_column
I am analyzing the Secom dataset from the UCI Machine Learning repository by lasso-regularized logistic regression, but the results are bad.
https://archive.ics.uci.edu/ml/datasets/SECOM
Characteristics:
1546 data samples with 590 numeric attributes
106 positive samples (production failures)
The goal is to predict the positive class accurately, and also to perform feature selection.
I optimize lambda by 10 fold cross validation with the glmnet package in R. But the results are terrible, as the model tends to assign all test samples to only a single class.
Is it simply the wrong kind of model for this dataset?
Predicting with imbalanced classes can be a very difficult problem to solve. Thankfully, there is a huge bibliography on how to solve such problems. There is a particular one that worked really well for me. It involves using up-sampling and/or down-sampling techniques:
down-sampling: randomly subset all the classes in the training set so that their class frequencies match the least prevalent class. For example, suppose that 80% of the training set samples are the first class and the remaining 20% are in the second class. Down-sampling would randomly sample the first class to be the same size as the second class (so that only 40% of the total training set is used to fit the model). caret contains a function (downSample) to do this.
up-sampling: randomly sample (with replacement) the minority class to be the same size as the majority class. caret contains a function (upSample) to do this.
hybrid methods: techniques such as SMOTE and ROSE down-sample the majority class and synthesize new data points in the minority class. There are two packages (DMwR and ROSE) that implement these procedures.
I took the above bullet points from this caret's documentation. The post contains examples about each one of the above bullet points and R code. You should be able to use a Lasso logistic regression and have better results after you have transformed your data based on the above techniques.
I am building a sentiment analysis program using some tweets i have gathered. The labeled data which i have gathered would go through a neural network which classifies them into two classes, positive and negative.
The data is still being labeled. So far i have observed that the positive category has very small number of observations.
The records for positive category in my training set could be about 5% of the training data set (the same ratio could reflect in the population as well).
Would this create problems in the final "program" ?
Size of the data set is about 5000 records.
Yes, yes it can. There are two things to consider:
5% of 5000 is 250. Thus you will try to model data distribution of your class based on just 250 samples. This might be orders of magnitude to small for neural network. Consequently you might need 40x more data to have a representative sample of your data. While you can easily reduce the majority class through subsampling, without the big risk of destroing the structure - there is no way to get "more structure" from less points (you can replicate points, add noise etc. but this does not add structure, this just adds assumptions).
Class imbalance can also lead to convergence to naive solutions, like "always False" which has 95% accuracy. Here you can simply play around with the cost function to make it more robust to imbalance (in particular - train splits suggested by #PureW is nothing else like "black box" method of trying to change the loss function so it has bigger weight on minority class. When you have access to your classifier loss, like in NN you should not due this - but instead change the cost function and still keep all the data).
Without even splits of the different classes, you may want to introduce weights in your loss function such that errors in the smaller class are deemed more important.
Another solution, since 5000 samples may or may not be much data depending on your problem, could be to sample more datasets. You basically take this set of 5000 samples, and sample datapoints from it such that you have a new dataset with even split of the classes. This means the new dataset is only 10% of your original dataset. But it is evenly split between the classes. You can redo this sampling many times and end up with several datasets, useful in bootstrap aggregating.
I have very small data that belongs to positive class and a large set of data from negative class. According to prof. Andrew Ng (anomaly detection vs supervised learning), I should use Anomaly detection instead of Supervised learning because of highly skewed data.
Please correct me if I am wrong but both techniques look same to me i.e. in both (supervised) Anomaly detection, and standard Supervised learning, we train data with both normal and anomalous samples and test on unknown data. Is there any difference?
Should I just perform under-sampling of negative class or over-sampling of positive class to get both type data of same size? Does it affect the overall accuracy?
Actually in supervised learning, you have the data set labelled (e.g good, bad) and you pass the labelled values as you train the model so that it learns parameters that will separate the 'good' from 'bad' results.
In anomaly detection, it is unsupervised as you do not pass any labelled values.. What you do is you train using only the 'non-anomalous' data. You then select epsilon values and evaluate with a numerical value (such as F1 score) so that your model will get a good balance of true positives.
Regarding trying to over/under sample so your data is not skewed, there are 2 things.
Prof Ng mentioned something like if your positive class is only 10 out of 10k or 100k then you need to use anomaly detection since your data is highly skewed.
Supervised learning makes sense if you know typically what 'bad' values are. If you only know what is 'normal'/'good' but your 'bad' value can really be very different every time then this is a good case for anomaly detection.
In anomaly detection you would determine model parameters from the portion of the data which is well supported (As Andrew explains). Since your negative class has many instances you would use these data for 'learning'. Kernel density estimation or GMMs are examples of approaches that are typically used. A model of 'normalcy' may thus be learnt and thresholding may be used to detect instances which are considered anomalous with respect to your derived model. The difference between this approach and conventional supervised learning lies in the fact that you are using only a portion of the data (the negative class in your case) for training. You would expect your positive instances to be identified as anomalous after training.
As for your second question, under-sampling the negative class will result in a loss of information whilst over-sampling the positive class doesn't add information. I don't think that following that route is desirable.
How should I set my gamma and Cost parameters in libSVM when I am using an imbalanced dataset that consists of 75% 'true' labels and 25% 'false' labels? I'm getting a constant error of having all the predicted labels set on 'True' due to the data imbalance.
If the issue isn't with libSVM, but with my dataset, how should I handle this imbalance from a Theoretical Machine Learning standpoint? *The number of features I'm using is between 4-10 and I have a small set of 250 data points.
Classes imbalance has nothing to do with selection of C and gamma, to deal with this issue you should use the class weighting scheme which is avaliable in for example scikit-learn package (built on libsvm)
Selection of best C and gamma is performed using grid search with cross validation. You should try vast range of values here, for C it is reasonable to choose values between 1 and 10^15 while a simple and good heuristic of gamma range values is to compute pairwise distances between all your data points and select gamma according to the percentiles of this distribution - think about putting in each point a gaussian distribution with variance equal to 1/gamma - if you select such gamma that this distribution overlaps will many points you will get very "smooth" model, while using small variance leads to the overfitting.
Imbalanced data sets can be tackled in various ways. Class balance has no effect on kernel parameters such as gamma for the RBF kernel.
The two most popular approaches are:
Use different misclassification penalties per class, this basically means changing C. Typically the smallest class gets weighed higher, a common approach is npos * wpos = nneg * wneg. LIBSVM allows you to do this using its -wX flags.
Subsample the overrepresented class to obtain an equal amount of positives and negatives and proceed with training as you traditionally would for a balanced set. Take note that you basically ignore a large chunk of data this way, which is intuitively a bad idea.
I know this has been asked some time ago, but I would like to answer it since you might find my answer useful.
As others have mentioned, you might want to consider using different weights for the minority classes or using different misclassification penalties. However, there is a more clever way of dealing with the imbalanced datasets.
You can use the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to generate synthesized data for the minority class. It is a simple algorithm that can deal with some imbalance datasets pretty well.
In each iteration of the algorithm, SMOTE considers two random instances of the minority class and add an artificial example of the same class somewhere in between. The algorithm keeps injecting the dataset with the samples until the two classes become balanced or some other criteria(e.g. add certain number of examples). Below you can find a picture describing what the algorithm does for a simple dataset in 2D feature space.
Associating weight with the minority class is a special case of this algorithm. When you associate weight $w_i$ with instance i, you are basically adding the extra $w_i - 1$ instances on top of the instance i!
What you need to do is to augment your initial dataset with the samples created by this algorithm, and train the SVM with this new dataset. You can also find many implementation online in different languages like Python and Matlab.
There have been other extensions of this algorithm, I can point you to more materials if you want.
To test the classifier you need to split the dataset into test and train, add synthetic instances to the train set (DO NOT ADD ANY TO THE TEST SET), train the model on the train set, and finally test it on the test set. If you consider the generated instances when you are testing you will end up with a biased(and ridiculously higher) accuracy and recall.