Mushroom data creation, curation, and simulation to support classification tasks.

Area of Science:

• Mycology
• Machine Learning
• Data Science

Background:

• Accurate mushroom identification is crucial for public safety.
• Existing datasets may lack comprehensiveness for robust machine learning models.
• Developing reliable classification rules for edible and poisonous mushrooms is a key challenge.

Purpose of the Study:

• To create and curate the largest attribute-based dataset for mushroom classification.
• To evaluate machine learning algorithms for predicting mushroom edibility.
• To provide a reproducible workflow and FAIR data for future research.

Main Methods:

• Utilized natural language processing on a mushroom identification textbook to generate primary data.
• Included simulated and hypothetical entries for comprehensive data simulation.
• Evaluated Naive Bayes, Logistic Regression, Linear Discriminant Analysis (LDA), and Random Forests (RF).

Main Results:

• Random Forests (RF) achieved perfect five-fold Cross-Validation accuracy (1.0) and F2-score.
• The developed dataset demonstrated that mushroom classification is not linearly separable.
• The dataset comprises 173 species from 23 families, making it the largest available.

Conclusions:

• The curated dataset and RF model offer a powerful tool for mushroom edibility prediction.
• The non-linear separability highlights the complexity of mushroom identification.
• The study provides a reproducible and FAIR-compliant resource for mycological research.