The Iris flower data set or Fisher's Iris data set is a multivariate data set used and made famous by the British statistician and biologist Ronald Fisher in his 1936 paper The use of multiple measurements in taxonomic problems as an example of linear discriminant analysis.
The data set consists of 50 samples from each of three species of Iris (Iris setosa, Iris virginica and Iris versicolor). Four features were measured from each sample: the length and the width of the sepals and petals, in centimeters.
├─ img/
├─ models/
├─ notebooks/
├─ src/
img: images used to document this project
models: serialized models
notebooks: notebooks used during analysis and model training
src: helper functions used during analysis
The objective of the project is to perform exploratory data analysis (EDA) and train a model capable of classifying a flower into one of three species. This project follows the CRISP-DM framework.
The dataset contains a set of 150 records under five attributes: sepal length, sepal width, petal length, petal width, and species. In this section, we explored the distribution of data across different species. The number of samples for each of the three species was the same, as we can see in the chart below, meaning we don't have any kind of imbalance between classes.
The next step was to understand the characteristics and statistics of each feature, including its distribution, mean values, standard deviation, and to check for outlier values. A summary of this can be seen through the boxplot below. As we can see, there are a few outliers in the distribution, but these are real values, which means we cannot discard them as they could also occur in future samples. Also through this chart, it is possible to identify that Iris Setosa has a very different distribution of values for petal length and petal width, with no overlap between its values and those of the other two classes.
After that, we looked at the relation between features. In this step, we first created three new features calculated from the existing ones.
| feature | observation |
|---|---|
| Petal_Prop | ratio between petal length and petal width |
| Sepal_Prop | ratio between sepal length and sepal width |
| SPL/SL | ratio between petal length and sepal length |
With these features, we explored the relation between each one, for each species. As we can observe, petal length and petal width seem to have a strong relation. Also from the plots, we can clearly see that Iris Setosa can be easily separated from the other two, as it was previously suggested when we analyzed the distribution of our data.
We can see how strong the correlation between features is through the correlation matrix. A correlation matrix is a table that displays the correlation coefficients between pairs of variables. This helps visualize the strength and direction (positive or negative) of relationships within a dataset.
To make a streamline process of all the transformations and features engineering we've created a pipeline where the new features as computed.
The modeling phase is where se select the models that will be used, our goal is to classify new samples of flower in one of the three species, so we will use classifiers and compare the results the models selected are
| Model | Library |
|---|---|
| Decision Tree Classifier | scikit-learn |
| Random Forest Classifier | scikit-learn |
| SVM Classifier | scikit-learn |
| KNN Classifier | scikit-learn |
| MLP Classifier | scikit-learn |
We first split the DataFrame into features (X) and our target variable (y). After that, it was necessary to encode our target variable. We used a LabelEncoder to assign a different numeric value to each of the species.
Next, we had to split our data into training and testing sets. For this, we used the train_test_split function from the scikit-learn library. The data doesn't have many observations, so we decided to use 20% of it for testing and 80% for training. random_state was set to 0 for reproducibility purposes
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)
For each model, we trained it using the training set without modifying any hyperparameters; we used the default values. After training, we predicted the class labels of observations in the test dataset and computed Accuracy, Balanced Accuracy, Precision, Recall, and F1-score for each model. The results are as follows:
| Model | Accuracy | Balanced Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|---|
| Desicion Tree | 96.67% | 97.44% | 96.67% | 96.67%% | 96.67% |
| Random Forest | 100% | 100% | 100% | 100%% | 100% |
| SVM | 100% | 100% | 100% | 100%% | 100% |
| KNN | 100% | 100% | 100% | 100%% | 100% |
| MLP | 100% | 100% | 100% | 100%% | 100% |
All models exhibited excellent performance in predicting classes on the testing set. The Decision Tree Classifier was the only model that misclassified one instance. While such high scores might initially suggest overfitting, this is unlikely in this case. The dataset is relatively small and straightforward, making achieving high accuracy expected.
To help visualize any errors, we also plotted the confusion matrix for each model. Since only the Decision Tree Classifier misclassified one instance, all confusion matrices are identical except for the one corresponding to the Decision Tree Classifier.
This last one is the confusion matrix of the Decision Tree classifier. As it might be observed, the error happened when the model classified a versicolor observation as virginica.
Through this project, we analyzed and understood the distinct characteristics of three iris flower species: Iris setosa, Iris virginica, and Iris versicolor. We identified that Iris setosa was linearly separable from the other two, making its classification relatively straightforward. While Iris virginica and Iris versicolor also exhibited different characteristic distributions, they shared some overlapping values, which presented a greater classification challenge.
To address this, we built several classification models (Decision Tree, Random Forest, SVM, KNN, and MLP) and evaluated their performance using a range of metrics (Accuracy, Balanced Accuracy, Precision, Recall, and F1-score). All models performed exceptionally well, achieving nearly perfect scores, with most metrics reaching 100%. The only exception was the Decision Tree Classifier, which misclassified a single sample in our test dataset.
To further refine these models, future work could involve hyperparameter optimization using techniques like GridSearch or RandomSearch. If possible, evaluate the model's performance on an independent, unseen dataset to assess its true generalization ability.
This project will not involve the deployment of any models in a production environment. The deployment phase will consist of storing the serialized models within the designated models folder. The pickle library will be utilized for serializing and saving the trained models to disk.
This approach allows for convenient storage and potential future reuse of the trained models. However, it's important to note that this is not a complete deployment solution. For production deployment, additional considerations such as model serving, versioning, monitoring, and security would be necessary.