Classification overview and evaluation techniques

AI provides us with various classification techniques, but machine learning classification would be the best to start with as it is the most common and easiest classification to understand for the beginner. In our daily life, our eyes captures millions of pictures: be they in a book, on a particular screen, or maybe something that you caught in your surroundings. These images captured by our eyes help us to recognize and classify objects. Our application is based on the same logic.

Here, we are creating an application that will identify images using machine learning algorithms. Imagine that we have images of both apples and oranges, looking at which our application would help identify whether the image is of an apple or an orange. This type of classification can be termed as binary classification, which means classifying the objects of a given set into two groups, but techniques do exist for multiclass classification as well. We would require a large number of images of apples and oranges, and a machine learning algorithm that would be set in such a way that the application would be able to classify both image types. In other words, we make these algorithms learn the difference between the two objects to help classify all the examples correctly. This is known as supervised learning.

Now let's compare supervised learning with unsupervised learning. Let's assume that we are not aware of the actual data labels (which means we do not know whether the images are examples of apples or oranges). In such cases, classification won't be of much help. The clustering method can always ease such scenarios. The result would be a model that can be deployed in an application, and it would function as seen in the following diagram. The application would memorize facts about the distinction between apples and oranges and recognize actual images using a machine learning algorithm. If we took a new input, the model would tell us about its decision as to whether the input is an apple or orange. In this example, the application that we created is able to identify an image of an apple with a 75% degree of confidence:

Sometimes, we want to know the level of confidence, and other times we just want the final answer, that is, the choice in which the model has the most confidence.

Evaluation

We can evaluate how well the model is working by measuring its accuracy. Accuracy would be defined as the percentage of cases that are classified correctly. We can analyze the mistakes made by the model, or its level of confusion, using a confusion matrix. The confusion matrix refers to the confusion in the model, but these confusion matrices can become a little difficult to understand when they become very large. Let's take a look at the following binary classification example, which shows the number of times that the model has made the correct predictions of the object:

In the preceding table, the rows of True apple and True orange refers to cases where the object was actually an apple or actually an orange. The columns refer to the prediction made by the model. We see that in our example, there are 20 apples that were predicted correctly, while there were 5 apples that were wrongly identified as oranges.

Ideally, a confusion matrix should have all zeros, except for the diagonal. Here we can calculate the accuracy by adding the figures diagonally, so that these are all the correctly classified examples, and dividing that sum by the sum of all the numbers in the matrix:

Here we got the accuracy as 84%. To know more about confusion matrices, let's go through another example, which involves three classes, as seen in the following diagram:

Source: scikit-learn docs

There are three different species of iris flowers. The matrix gives raw accounts of correct and incorrect predictions. So, setosa was correctly predicted 13 times out of all the examples of setosa images from the dataset. On the other hand, versicolor was predicted correctly on 10 occasions, and there were 6 occasions where versicolor was predicted as virginica. Now let's normalize our confusion matrix and show the percentage of the cases that predicted image corrected or incorrectly. In our example we saw that the setosa species was predicted correctly throughout:

Source: scikit-learn docs

During evaluation of the confusion matrix, we also saw that the system got confused between two species: versicolor and virginica. This also gives us the conclusion that the system is not able to identify species of virginica all the time.

For further instances, we need to be more aware that we cannot have really high accuracy since the system will be trained and tested on the same data. This will lead to memorizing the training set and overfitting of the model. Therefore, we should try to split the data into training and testing sets, first in either 90/10% or 80/20%. Then we should use the training set for developing the model and the test set for performing and calculating the accuracy of the confusion matrix.

We need to be careful not to choose a really good testing set or a really bad testing set to get the accuracy. Hence to be sure we use a validation known as K-fold cross validation. To understand it a bit better, imagine 5-fold cross validation, where we move the testing set by 20 since there are 5 rows. Then we move the remaining set with the dataset and find the average of all the folds:

Quite confusing, right? But scikit-learn has built-in support for cross validation. This feature will be a good way to make sure that we are not overfitting our model and we are not running our model on a bad testing set.