Machine Learning in Python with Scikit-learn

Hello students, welcome to Kwickprep. Your first model scores ninety percent. Is that good or bad? By the end of this lesson, you will know how to answer. We will see what scikit learn gives us, split data for training and testing, and build a model with fit, predict and score. Then we will read the result properly.

First, a quick reminder of machine learning. The computer learns patterns from many examples. We do not write the rules by hand. For example, after seeing many fruits, it can guess a new fruit's type from its size and colour. What it learns is called a model.

Scikit learn is a free Python library for machine learning. It gives ready datasets, so you can practise without collecting data. It gives ready models for classification, regression and clustering. Classification predicts a category, regression predicts a number, and clustering makes groups. It has a function to split data, and metrics, which are ways to measure a model. You install it as scikit dash learn, but import it as S K learn.

We will use the famous iris flower dataset that comes with scikit learn. It has one hundred fifty flowers of three species. Petal length is a feature, which means an input the model looks at. Petal width is another feature, and there are four features in all. The species is the label, which means the answer we want the model to predict.

Load iris gives us the data. With return ex why equals True, it returns two parts. By custom, capital ex holds the features, and small why holds the labels. ex dot shape is one hundred fifty comma four, so there are one hundred fifty rows and four feature columns. Why dot shape is one hundred fifty, one label for each flower. The species are stored as numbers zero, one and two.

Why do we split the data? A model can simply memorise the examples it has already seen. So we keep two sets. The training set is what the model learns from. The test set is kept hidden until the end. It is like a board exam. If the questions were the same as the practice sheet, full marks would prove nothing.

Train test split divides the data for us. Test size equals zero point two keeps twenty percent for testing. So one hundred twenty flowers are for training, and thirty are for testing. The rows are shuffled first, and random state fixes that shuffle, so you get the same split every time. It returns four parts, in this order: ex train, ex test, why train and why test.

Every scikit learn project follows the same steps. First, load the features and labels. Second, split them into training and test sets. Third, fit the model on the training data, which means let it learn. Fourth, predict the answers for the test data. Fifth, score the model, and read what that score really means.

We will use a simple model called kay nearest neighbours. To guess a new flower, it finds the three most similar flowers it has seen, and takes a majority vote. It is like asking your three nearest neighbours for directions. The three inside the brackets is the number of neighbours. Model dot fit, with ex train and why train, is the learning step.

Model dot predict takes features only, with no answers, and returns its guesses. Here we ask about the first five test flowers. The model guesses two, one, zero, one, one. The real labels are two, one, zero, one, two. Pause and predict. How many did it get right? Four of five, because the last flower was wrong.

Model dot score compares the predictions with the real test labels. For a classifier, it returns the accuracy, which means the fraction of correct answers. Here the score is zero point nine. So twenty seven of the thirty test flowers were right, and three were wrong. Every scikit learn model uses the same three words: fit, predict and score.

Now let us read the number, not just say it. Zero point nine means twenty seven of thirty test flowers were right. Compare it with random guessing, which gets about thirty three percent with three species, so the model has clearly learned. But the test set is small, so a single mistake changes the score by three point three percent. And we must also ask which flowers went wrong.

In one typical run, here is how the thirty test flowers did, species by species. All seven setosa flowers were right. Eleven of twelve versicolor flowers were right, and one was called virginica. Nine of eleven virginica flowers were right, and two were called versicolor. So the model only mixes up the two species that look alike. That is far more useful than the single number.

A few more checks. On the training data, the score is zero point nine seven five, and on the test data it is zero point nine. A small gap is normal, but a very large gap means the model memorised instead of learning. Change the random state, and the score may become one point zero or drop lower, so one split is not the full story. And in serious uses like health, missing a sick patient matters more than overall accuracy.

Let us revise what we learned today. Scikit learn gives datasets, models, splitting and metrics. Capital ex holds the features, and small why holds the labels. Train test split keeps a hidden test set. Fit learns, predict answers, and score measures. And always read the result, by comparing with guessing, checking the mistakes, and minding the test size. So is ninety percent good? Now you know it depends.