Training a multilayer perceptron in Weka - machine-learning

I am using Weka to implement classification algorithms. I was dealing with Multilayer Perceptron. I have some doubts while training the model. I used toy datasets that are already available in Weka. The name of the datasets are contact-lenses.arff and weather.nominal.arff.
I am attaching some screenshots.
]
I was using 5fold cross validation method.
As per the definition available in weka for hidden layer, There are also wildcard values: 'a' = (attribs + classes) / 2, 'i' = attribs, 'o' = classes , 't' = attribs + classes.
For the 1st screenshot a should be (No. of attribute + classes)/2 = (4+3)/2 = 7/2 = 3.5 = 4
So we can see the 4 nodes in the hidden layer.
Now for screenshot 3, (4+2)/2 = 3. But we can see 5 nodes in the hidden layer. Why there is a mismatch between actual nodes and calculated nodes?
Secondly if we consider 1st two screenshots we can see age uses three different values in the model namely, young, pre-prebyopic, prebyopic. However, attribute spectacle-prescrip has two different values namely, myope and hypermetrope but only one value hypermetrope used to train the model. What about the other values myope? Same doubts raised for other attributes as well.
A brief explanation will be helpful.

By default, MultilayerPerceptron applies the unsupervised NominalToBinary filter to the input data, which will increase the number of attributes for non-binary nominal attributes. That explains the different number of nodes in the hidden layer.
Also, a binary nominal attribute can be modeled with just a single node (below threshold is one label, above threshold the other one) - and NominalToBinary, with its default settings, does not change these types of attributes.

Related

Specifying class or sample weights in Keras for one-hot encoded labels in a TF Dataset

I am trying to train an image classifier on an unbalanced training set. In order to cope with the class imbalance, I want either to weight the classes or the individual samples. Weighting the classes does not seem to work. And somehow for my setup I was not able to find a way to specify the samples weights. Below you can read how I load and encode the training data and the two approaches that I tried.
Training data loading and encoding
My training data is stored in a directory structure where each image is place in the subfolder corresponding to its class (I have 32 classes in total). Since the training data is too big too all load at once into memory I make use of image_dataset_from_directory and by that describe the data in a TF Dataset:
train_ds = keras.preprocessing.image_dataset_from_directory (training_data_dir,
batch_size=batch_size,
image_size=img_size,
label_mode='categorical')
I use label_mode 'categorical', so that the labels are described as a one-hot encoded vector.
I then prefetch the data:
train_ds = train_ds.prefetch(buffer_size=buffer_size)
Approach 1: specifying class weights
In this approach I try to specify the class weights of the classes via the class_weight argument of fit:
model.fit(
train_ds, epochs=epochs, callbacks=callbacks, validation_data=val_ds,
class_weight=class_weights
)
For each class we compute weight which are inversely proportional to the number of training samples for that class. This is done as follows (this is done before the train_ds.prefetch() call described above):
class_num_training_samples = {}
for f in train_ds.file_paths:
class_name = f.split('/')[-2]
if class_name in class_num_training_samples:
class_num_training_samples[class_name] += 1
else:
class_num_training_samples[class_name] = 1
max_class_samples = max(class_num_training_samples.values())
class_weights = {}
for i in range(0, len(train_ds.class_names)):
class_weights[i] = max_class_samples/class_num_training_samples[train_ds.class_names[i]]
What I am not sure about is whether this solution works, because the keras documentation does not specify the keys for the class_weights dictionary in case the labels are one-hot encoded.
I tried training the network this way but found out that the weights did not have a real influence on the resulting network: when I looked at the distribution of predicted classes for each individual class then I could recognize the distribution of the overall training set, where for each class the prediction of the dominant classes is most likely.
Running the same training without any class weight specified led to similar results.
So I suspect that the weights don't seem to have an influence in my case.
Is this because specifying class weights does not work for one-hot encoded labels, or is this because I am probably doing something else wrong (in the code I did not show here)?
Approach 2: specifying sample weight
As an attempt to come up with a different (in my opinion less elegant) solution I wanted to specify the individual sample weights via the sample_weight argument of the fit method. However from the documentation I find:
[...] This argument is not supported when x is a dataset, generator, or keras.utils.Sequence instance, instead provide the sample_weights as the third element of x.
Which is indeed the case in my setup where train_ds is a dataset. Now I really having trouble finding documentation from which I can derive how I can modify train_ds, such that it has a third element with the weight. I thought using the map method of a dataset can be useful, but the solution I came up with is apparently not valid:
train_ds = train_ds.map(lambda img, label: (img, label, class_weights[np.argmax(label)]))
Does anyone have a solution that may work in combination with a dataset loaded by image_dataset_from_directory?

How to get jlibsvm prediction probability in multi-class classification

I am new to SVM. I am using jlibsvm for a multi-class classification problem. Basically, I am doing a sentence classification problem. There are 3 Classes. What I understood is I am doing One-against-all classification. I have a comparatively small train set. A total of 75 sentences, In which 25 sentences belongs to each class.
I am making 3 SVMs (so 3 different models), where, while training, in SVM_A, sentences belong to CLASS A will have a true label, i.e., 1 and other sentences will have a -1 label. Correspondingly done for SVM_B, and SVM_C.
While testing, to get the true label of a sentence, I am giving the sentence to 3 models and I am taking the prediction probability returned by these 3 models. Which one returns the highest will be the class the sentence belong to.
This is how I am doing. But I am getting the same prediction probability for every sentence in the test set for all models.
A predicted:0.012820514
B predicted:0.012820514
C predicted:0.012820514
These values repeat for all sentences in the training set.
The following is how I set parameters for training:
C_SVC svm = new C_SVC();
MutableBinaryClassificationProblemImpl problem;
ImmutableSvmParameterGrid.Builder builder = ImmutableSvmParameterGrid.builder();
// create training parameters ------------
HashSet<Float> cSet;
HashSet<LinearKernel> kernelSet;
cSet = new HashSet<Float>();
cSet.add(1.0f);
kernelSet = new HashSet<LinearKernel>();
kernelSet.add(new LinearKernel());
// configure finetuning parameters
builder.eps = 0.001f; // epsilon
builder.Cset = cSet; // C values used
builder.kernelSet = kernelSet; //Kernel used
builder.probability=true; // To get the prediction probability
ImmutableSvmParameter params = builder.build();
What am I doing wrong?
Is there any other better way to do multi-class classification other than this?
You are getting the same output, because you generate the same model three times.
The reason for this is, that jlibsvm is able to perform multiclass classification out of the box based on the provided data (LIBSVM itself supports this too). If it detects, that more than two class labes are provided in the given data, it automatically performs multiclass classification. So there is no need for a manually 1vsN approach. Just supply the data with class-labels for each category.
However, jlibsvm is still in beta and relies on a rather old version of LIBSVM (2.88). A lot has changed. For a more intiuitive Java binding (in comparison to the default LIBSVM version), you can take a look at zlibsvm, which is available via Maven Central and based on the latest LIBSVM version.

Parameter selection and k-fold cross-validation

I have one dataset, and need to do cross-validation, for example, a 10-fold cross-validation, on the entire dataset. I would like to use radial basis function (RBF) kernel with parameter selection (there are two parameters for an RBF kernel: C and gamma). Usually, people select the hyperparameters of SVM using a dev set, and then use the best hyperparameters based on the dev set and apply it to the test set for evaluations. However, in my case, the original dataset is partitioned into 10 subsets. Sequentially one subset is tested using the classifier trained on the remaining 9 subsets. It is obviously that we do not have fixed training and test data. How should I do hyper-parameter selection in this case?
Is your data partitioned into exactly those 10 partitions for a specific reason? If not you could concatenate/shuffle them together again, then do regular (repeated) cross validation to perform a parameter grid search. For example, with using 10 partitions and 10 repeats gives a total of 100 training and evaluation sets. Those are now used to train and evaluate all parameter sets, hence you will get 100 results per parameter set you tried. The average performance per parameter set can be computed from those 100 results per set then.
This process is built-in in most ML tools already, like with this short example in R, using the caret library:
library(caret)
library(lattice)
library(doMC)
registerDoMC(3)
model <- train(x = iris[,1:4],
y = iris[,5],
method = 'svmRadial',
preProcess = c('center', 'scale'),
tuneGrid = expand.grid(C=3**(-3:3), sigma=3**(-3:3)), # all permutations of these parameters get evaluated
trControl = trainControl(method = 'repeatedcv',
number = 10,
repeats = 10,
returnResamp = 'all', # store results of all parameter sets on all partitions and repeats
allowParallel = T))
# performance of different parameter set (e.g. average and standard deviation of performance)
print(model$results)
# visualization of the above
levelplot(x = Accuracy~C*sigma, data = model$results, col.regions=gray(100:0/100), scales=list(log=3))
# results of all parameter sets over all partitions and repeats. From this the metrics above get calculated
str(model$resample)
Once you have evaluated a grid of hyperparameters you can chose a reasonable parameter set ("model selection", e.g. by choosing a well performing while still reasonable incomplex model).
BTW: I would recommend repeated cross validation over cross validation if possible (eventually using more than 10 repeats, but details depend on your problem); and as #christian-cerri already recommended, having an additional, unseen test set that is used to estimate the performance of your final model on new data is a good idea.

Encog query classification

I'm trying to process this dataset using Encog. In order to do so, I combined the outputs into one (can't seem to figure out how to use multiple expected outputs, even tho I unsuccessfully tried to manually train a NN with 4 output neurons) with the values: "disease1", "disease2", "none" and "both".
Starting from there, used the analyst wizard in the CSV, and the automatic process trained a NN with the expected outputs. A peak from the file:
"field:1","field:2","field:3","field:4","field:5","field:6","field:7","Output:field:7"
40.5,yes,yes,yes,yes,no,both,both
41.2,no,yes,yes,no,yes,second,second
Now my problem is: how do I query it? I tried with classification, but as far as I've understood, the result only gives me the values {0,1,2}, so there are two classes which I can't differentiate (both are 0).
This same problem applies to the Iris example presented in the wiki. Also, how does Encog extrapolate from the output neuron values to the 0/1/2 results?
Edit: the solution I have found was to use a separate network for disease 1 and disease 2, but I really would like to know if it was possible to combine those into one.
You are correct, that you will need to combine the output column to a single value. Encog analyst will only classify to a single output column. That output column can have many different values. So normalizing the two output columns to none,first,second,both will work. If you use the underlying neural networks directly, you could actually train for two outputs each doing an independent classification. But for this discussion I will assume we are dealing with the analyst.
Are you querying the network using the workbench, or in code? By default Encog analyst encodes to the neural network using equilateral encoding. This results in a number of output neurons equal to n-1, where n is the number of classes. If you choose one-of-n encoding in the analyst wizard, then the regular classify method on the BasicNetwork will work, as it is only designed for one-of-n.
If you would like to query (in code) using equilateral, then you can use a method similar to the following. I am adding this to the next version of Encog.
/**
* Used to classify a neural network that has been encoded using equilateral encoding.
* This is the default for the Encog analyst. Equilateral encoding uses an output count
* equal to the number of classes minus one.
* #param input The input to the neural network.
* #param high The high value of the activation range, usually 1.
* #param low The low end of the normalization range, usually -1 or 0.
* #return The class that the input belongs to.
*/
public int classifyEquilateral(final MLData input,double high, double low) {
MLData result = this.compute(input);
Equilateral eq = new Equilateral(getOutputCount()+1,high,low);
return eq.decode(result.getData());
}

How are binary classifiers generalised to classify data into arbitrarily large sets?

How can algorithms which partition a space in to halves, such as Suport Vector Machines, be generalised to label data with labels from sets such as the integers?
For example, a support vector machine operates by constructing a hyperplane and then things 'above' the hyperplane take one label, and things below it take the other label.
How does this get generalised so that the labels are, for example, integers, or some other arbitrarily large set?
One option is the 'one-vs-all' approach, in which you create one classifier for each set you want to partition into, and select the set with the highest probability.
For example, say you want to classify objects with a label from {1,2,3}. Then you can create three binary classifiers:
C1 = 1 or (not 1)
C2 = 2 or (not 2)
C3 = 3 or (not 3)
If you run these classifiers on a new piece of data X, then they might return:
C1(X) = 31.6% chance of being in 1
C2(X) = 63.3% chance of being in 2
C3(X) = 89.3% chance of being in 3
Based on these outputs, you could classify X as most likely being from class 3. (The probabilities don't add up to 1 - that's because the classifiers don't know about each other).
If your output labels are ordered (with some kind of meaningful, rather than arbitrary ordering). For example, in finance you want to classify stocks into {BUY, SELL, HOLD}. Although you can't legitimately perform a regression on these (the data is ordinal rather than ratio data) you can assign the values of -1, 0 and 1 to SELL, HOLD and BUY and then pretend that you have ratio data. Sometimes this can give good results even though it's not theoretically justified.
Another approach is the Cramer-Singer method ("On the algorithmic implementation of multiclass kernel-based vector machines").
Svmlight implements it here: http://svmlight.joachims.org/svm_multiclass.html.
Classification into an infinite set (such as the set of integers) is called ordinal regression. Usually this is done by mapping a range of continuous values onto an element of the set. (see http://mlg.eng.cam.ac.uk/zoubin/papers/chu05a.pdf, Figure 1a)

Resources