Mushroom Classification Project

Author: Lindsay Foster Date: November 2025
Course: Applied Machine Learning – Midterm Project

---

Overview

This project applies machine learning classification techniques to predict whether a mushroom is edible or poisonous based on its physical characteristics.
The analysis follows a structured workflow including data exploration, preprocessing, feature selection, model training, evaluation, and comparison of multiple classifiers.

Project Files

• Jupyter Notebook: classification_foster.ipynb
• Peer Review: peer_review.md

Imports Used

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.svm import SVC
from sklearn.metrics import classification_report, confusion_matrix
from sklearn.model_selection import GridSearchCV

---

Dataset

Source: UCI Mushroom Dataset
Description:
The dataset contains 8124 mushroom samples described by 22 categorical features such as cap shape, odor, gill color, and habitat.
The target variable (class) indicates whether a mushroom is edible (e) or poisonous (p).

Example Features:

• cap-shape
• cap-surface
• odor
• gill-color
• habitat
• class (target)

---

Project Goals

• Load, inspect, and clean the mushroom dataset.
• Explore distributions and feature importance.
• Encode categorical variables for modeling.
• Train and evaluate multiple classification algorithms:
  • Decision Tree
  • Random Forest
  • Logistic Regression
• Compare performance using metrics such as:
  • Accuracy
  • Precision
  • Recall
  • F1-score
  • Confusion Matrix

---

Setup & Run Instructions

Clone the Repository

git clone https://github.com/yourusername/mushroom-classification.git
cd mushroom-classification

Create and Activate Virtual Environment

.venv\Scripts\activate

Install Dependencies

pip install -r requirements.txt

Launch Notebook: classification_foster

Load and Inspect Data

• Loaded dataset into pandas DataFrames for features (X) and target (y).
• Displayed the first 10 rows of the dataset and verified feature names and target variable.
• Check for missing values and duplicates.
• Findings All columns have no missing values except stalk-root, which has 2,480 missing entries (~30% of rows). There are no duplicate rows.
• Fill missing 'stalk-root' values with 'unknown'.
• Summary Statistics Most features are categorical with a small number of unique values. Target variable class is balanced between edible and poisonous mushrooms.

Data Exploration and Preparation

Patterns and Anomalies Observed

• Most features in the Mushroom dataset are categorical, with a small set of unique values per feature.
• The target variable (poisonous) is roughly balanced, which is ideal for classification tasks.
• Some categories are rare (e.g., certain odor or cap-color values), which could be highly predictive.
• The stalk-root column had 30% missing values, which needed attention to avoid biasing the model.

Features That Stand Out

• odor: Certain odors strongly correlate with poisonous mushrooms and are likely key predictors.
• cap_look : A new feature created by combining cap-shape and cap-color, capturing patterns not evident from individual features.
• stalk-root : After imputing missing values, this feature adds predictive information.

Prep Steps

1. Handling Missing Values
   • Imputed missing values in stalk-root with "unknown" to preserve all rows.
2. Feature Engineering
   • Created cap_look by combining cap-shape and cap-color to summarize the mushroom’s overall appearance.
3. Encoding Categorical Variables
   • Applied one-hot encoding using pd.get_dummies() to convert categorical features into numeric columns.
4. Scaling / Normalization (Optional)
   • Not necessary for tree-based models (Decision Tree, Random Forest). Required only for models sensitive to feature scale, such as Logistic Regression or MLP.

Feature Selection and Modification

• Selected key features likely to contribute to predictive power: cap_look, odor, stalk-root.
• One-hot encoding expanded categorical features into binary columns for better interpretability and modeling.
• Combined features like cap_look were created to capture interactions and patterns that improve model performance.

Reflection

• During data exploration, I identified key predictive features (odor and cap_look) and addressed missing values in stalk-root by imputing "unknown".
• One-hot encoding transformed categorical variables into numeric form suitable for modeling. Creating combined features, like cap_look, allowed the model to detect patterns that individual features alone might not capture.

Feature Selection and Justification

• Features: cap_look, odor, stalk-root

• Target: poisonous (encoded as 0/1)

• Rationale: These features are meaningful, interpretable, and likely to improve model performance.

• Preprocessing: Combined features, imputed missing values, and one-hot encoding applied.

• I selected cap_look, odor, and stalk-root as input features because they are directly related to mushroom characteristics that indicate edibility or toxicity.

• I created a heatmap to discern a relationship between cap_look and odor to see if there were any crossovers. All mushrooms with odor n of all cap looks were not poisonous.

• Selected features: 'cap_look', 'odor', 'stalk-root' for predictive power.

• 'odor' is highly indicative of poisonous mushrooms.

• 'cap_look' combines cap shape and color to capture visual patterns.

• 'stalk-root' provides structural information and missing values were imputed as 'unknown'.

• One-hot encoding converts categorical data into numeric form suitable for modeling.

• These features are expected to improve prediction accuracy and model interpretability.

Train a Model

• I split data into a 70% training and a 30% test set, maintaining class proportions (stratify=y).
• I trained a Decision Tree classifier to predict whether a mushroom is poisonous or edible.

Split the Data

• The dataset was split into training (70%)** and test (30%) sets using train_test_split.
• The split was stratified by the target variable poisonous to maintain class balance between edible and poisonous mushrooms.
• A Decision Tree classifier was trained using Scikit-Learn's fit() method on the training data (X_train, y_train).
• Input features used for training:
  • cap_look (combined cap shape + color)
  • odor
  • stalk-root

Evaluate Performance

• The model was evaluated using common classification metrics:
  • Accuracy – proportion of correctly classified mushrooms.
  • Precision – proportion of predicted poisonous mushrooms that are truly poisonous.
  • Recall – proportion of actual poisonous mushrooms that were correctly identified.
  • F1-score – harmonic mean of precision and recall.
• Additionally, I generated a confusion matrix and a full classification report to inspect detailed performance.

Results

The Decision Tree model achieved high accuracy due to the strong predictive power of features. The confusion matrix highlights how well the model distinguishes between edible and poisonous mushrooms.
Overall, the model demonstrates the effectiveness of our feature selection and preprocessing steps.

Reflection:

This model performed very well. The model performance metrics are all close to 99%.

• Accuracy: 0.9979 → about 99.8% of mushrooms were correctly classified.
• Precision: 0.9991 → of all mushrooms predicted as poisonous, 99.9% were actually poisonous. High precision means few false positives (few edible mushrooms were wrongly labeled as poisonous).
• Recall: 0.9966 → of all actual poisonous mushrooms, 99.66% were correctly identified. High recall means few false negatives (few poisonous mushrooms were missed).
• F1-score: 0.9979 → combines precision and recall. Close to 1.
• Confusion matrix shows:
  • TP (1262): edible mushrooms correctly classified
  • FP (1): edible mushrooms incorrectly classified as poisonous
  • FN (4): poisonous mushrooms incorrectly classified as edible
  • TN (1171): poisonous mushrooms correctly classified
• Only 5 misclassifications out of 2438 test samples.

Improve the Model or Try Alternates (Implement a Second Option)

To evaluate whether a different modeling approach could improve performance, we trained a Random Forest classifier as an alternative to the Decision Tree.

Train an Alternative Model

• Random Forest was trained using 100 estimators (n_estimators=100) with default hyperparameters.
• Same input features were used as in the Decision Tree:
  • cap_look (cap shape + color)
  • odor
  • stalk-root

Compare Performance

Metric	Decision Tree	Random Forest
Accuracy	0.9979	0.9984
Precision	0.9991	0.9992
Recall	0.9966	0.9974
F1-score	0.9979	0.9983

• Confusion Matrices show very few misclassifications in both models, indicating that the dataset is highly predictable with these features.
• Random Forest slightly improved metrics over the Decision Tree, demonstrating the benefit of ensemble methods for stability and minor performance gains.

Insights

Both models perform exceptionally well due to the highly predictive nature of features like odor and cap_look.
Random Forest offers slightly better recall, reducing the chance of missing poisonous mushrooms.
The comparison supports that ensemble methods can provide small but meaningful improvements on top of simple tree-based models.

Summary of Findings

• The classification models were able to predict whether a mushroom is poisonous or edible with extremely high accuracy (>99%).
• Key features driving predictions:
  • Odor – most predictive of poisonous mushrooms
  • Cap_look – combination of cap shape and color
  • Stalk-root – structural characteristic
• Both Decision Tree and Random Forest performed almost perfectly due to the strong predictive features.

Challenges Faced

• Handling the missing values in stalk-root required careful preprocessing; changed the missing values to "unknown".
• Encoding categorical features into a numerical format with one-hot encoding increased the number of features, which could become cumbersome for larger datasets.
• Ensuring proper stratified splits to maintain class balance required attention.

Learning from the Project

• Learned how to preprocess categorical data, handle missing values, and perform one-hot encoding.
• Gained experience in training, evaluating, and comparing multiple classifiers for a real-world dataset.
• Developed skills in interpreting performance metrics like accuracy, precision, recall, F1-score, and confusion matrices.
• Reinforced the importance of feature selection and thoughtful preprocessing to achieve high-performing machine learning models.