Your first ML project

Machine learning is when a computer learns patterns from examples (data) instead of being given exact step-by-step rules for every case. After learning, it can make predictions or decisions on new data it has not seen before. This is different from traditional programming, where a human writes all the rules by hand.

The examples used to teach the model are called the training data. Each example usually has input features (what we know) and often a label (the correct answer we want to predict). Good training data is very important because the model learns its patterns from it.

A feature is one piece of input information that describes an example, like the height of a plant or the number of rooms in a house. The model uses features to find patterns and make a prediction. A dataset usually has several features for each example.

A label is the correct answer attached to a training example, such as 'cat' or 'dog' for a photo. In supervised learning, the model studies inputs together with their labels so it can learn to predict the label for new inputs. Without labels, the learning would be unsupervised instead.

In supervised learning, the training data includes labels (correct answers), so the model learns to predict those answers. In unsupervised learning, there are no labels, so the model instead finds hidden patterns or groups in the data, like clustering similar customers together. The key difference is whether labeled answers are provided.

We train the model on the training set and then check its accuracy on a separate test set the model has never seen. This shows how well the model will work on new, real-world data instead of just memorizing the examples. If we tested on the same data we trained on, the score would be unfairly high.

Overfitting means the model has memorized the training data too closely, including its noise and quirks, so it performs well on training data but poorly on new data. It is like a student who memorizes exact answers but cannot solve a slightly different question. We want a model that generalizes, not one that just memorizes.

Underfitting happens when a model is too simple to capture the real patterns in the data, so it performs poorly even on the training data. It is like trying to draw a straight line through data that actually curves. The fix is usually a more powerful model or better features.

Accuracy is the fraction of predictions the model got correct out of all predictions made. For example, if a model labels 90 out of 100 images correctly, its accuracy is 90%. It is a simple, popular measure, but it can be misleading when the classes are very unbalanced.

The usual first step is to collect and prepare the data, which includes cleaning it, handling missing values, and organizing features and labels. Clean, well-prepared data leads to a much better model. People often say 'garbage in, garbage out' to stress how important data quality is.

scikit-learn (imported as sklearn) is a popular and beginner-friendly Python library for classic machine learning. It provides ready-made tools for splitting data, training models, and measuring accuracy with just a few lines of code. It is a great choice for a first ML project.

from sklearn.tree import DecisionTreeClassifier
model = DecisionTreeClassifier()

The .fit() method trains the model by feeding it the training inputs and their labels so it can learn the patterns. After fitting, the model is ready to make predictions. You typically call .fit(X_train, y_train) where X is the features and y is the labels.

model.fit(X_train, y_train)

The .predict() method uses a trained model to produce outputs (predictions) for new input data. For example, model.predict(X_test) returns the model's guessed labels for the test examples. You call it after the model has been trained with .fit().

predictions = model.predict(X_test)

The Iris dataset contains measurements (like petal and sepal length and width) of three species of iris flowers. It is small and clean, so beginners use it to practice classification, predicting the species from the measurements. It is built into scikit-learn, making it easy to load.

Classification predicts a category or class, such as 'spam' or 'not spam'. Regression predicts a continuous number, such as the price of a house or tomorrow's temperature. The type of answer you need decides which kind of task you have.

Predicting the exact price of a used bicycle is a regression problem because the answer is a continuous number. Tasks that predict a category, like whether an email is spam, are classification problems. The clue is whether the output is a number on a scale or a label.

A model is the result of training: it is the learned set of patterns or rules that can turn inputs into predictions. You create an empty model, train it with data, and then use it to predict on new inputs. Different algorithms (like decision trees) produce different kinds of models.

Models learn patterns from data, so too little or poor-quality data leads to weak, unreliable predictions. More good examples help the model learn real patterns instead of random noise. This is why people say 'garbage in, garbage out'.

A decision tree is a model that makes predictions by asking a series of yes/no questions about the features, like a flowchart. Each answer sends you down a branch until you reach a final decision at a leaf. It is popular for beginners because it is easy to understand and visualize.

Generalization is a model's ability to perform well on new, unseen data, not just the data it was trained on. Good generalization is the real goal of machine learning. A model that only does well on training data but fails on new data has poor generalization.

Data cleaning means fixing problems in the data, such as removing duplicates, correcting errors, and handling missing values. Clean data helps the model learn accurate patterns and avoids misleading results. It is often one of the most time-consuming but important steps.

A prediction is the output the trained model gives for a new input, like guessing that an email is spam or that a house costs a certain price. Predictions can be categories or numbers depending on the task. They are the main reason we build models.

The model learned patterns specific to the first city's data, which may not match the new city's conditions, prices, or habits. This is a form of bias from unrepresentative training data. To work everywhere, the training data should reflect the places where the model will be used.

Bias means the data unfairly favors or ignores certain groups or situations, often because it was not collected fairly. A biased dataset can make the model give unfair or wrong predictions for some people. Checking data for bias is an important part of responsible ML.

ML projects often use personal data, so collecting or sharing it without permission can harm people's privacy. Responsible practice means getting consent, using only the data you need, and keeping it secure. Respecting privacy builds trust and follows the law in many places.

The target, often written as y, is the value we want the model to predict, such as the flower species or house price. The features (X) are the inputs, and the target is the desired output. During training, the model learns to map X to y.

Scaling means adjusting features so they are on a similar range, for example converting all values to be between 0 and 1. This stops features with large numbers from unfairly dominating the model and can help many algorithms learn better. It is a common data-preparation step.

This is a false positive: the model predicted 'spam' (positive) when the email was actually not spam. False positives and false negatives are different kinds of errors that matter in different ways. For a spam filter, false positives can be costly because real emails get hidden.

If you test on the training data, the model may simply repeat answers it memorized, giving a falsely high score. This hides overfitting and does not show real-world performance. A separate test set gives an honest measure of how the model handles new data.

During training, the model makes guesses, compares them to the correct answers, and gradually adjusts itself to reduce its mistakes. Over many examples, it gets better at matching inputs to the right outputs. This repeated adjust-and-improve process is the core of learning from data.

pandas is a library for loading and organizing data in table form called a DataFrame. It makes it easy to read files like CSVs, clean data, and select features and labels. It is a very common first step before training a model.

import pandas as pd
data = pd.read_csv('flowers.csv')

A validation set is used during development to tune the model's settings and compare choices without touching the final test set. This keeps the test set truly unseen until the very end, giving an honest final score. It helps prevent accidentally overfitting to the test data.

A good first project is classifying iris flowers or handwritten digits using a small, clean, ready-made dataset. These projects are simple enough to finish quickly while teaching the full workflow: load data, train, test, and measure accuracy. Starting small builds confidence before tackling harder problems.

Looking at wrong predictions, called error analysis, helps you understand where and why the model struggles. You might find messy data, a confusing category, or a missing feature you can improve. This insight guides how to make the next version of the model better.