Overampling for ML

Machine Learning with samova.R

A guide to enhancing ML models using synthetic metagenomic data generation for better classification, regression, and data augmentation.

---

Why Use samova.R for Machine Learning?

Metagenomic datasets are often:
✅ Small – Limited samples for training robust models
✅ Imbalanced – Rare taxa or conditions are underrepresented
✅ Noisy – Technical biases and missing values

samova.R solves these issues by:
🔹 Generating biologically realistic synthetic samples
🔹 Balancing class distributions for better ML fairness
🔹 Providing augmented training data without overfitting

---

Workflow Overview

graph TD
    A[Real Data] --> B(Preprocess & Cluster)
    B --> C[Generate Synthetic Samples]
    C --> D{Machine Learning}
    D --> E[Train Model]
    D --> F[Validate Performance]

Loading

---

Step-by-Step Tutorial

1. Prepare Your Data

Load real metagenomic data (e.g., from GMrepo):

library(samovaR)
library(tidyverse)

# Download healthy/diseased gut microbiome data
healthy_data <- GMrepo_type2data(mesh_ids = "D006262") %>% 
  teatree_trim(threshold_amount = 1e-4)
diseased_data <- GMrepo_type2data(mesh_ids = "D003967") %>% 
  teatree_trim(threshold_amount = 1e-4)

2. Generate Synthetic Samples

Create augmented datasets for each class:

generate_samples <- function(data, n) {
  data %>%
    tealeaves_pack() %>%
    teabag_brew(min_cluster_size = 30) %>%
    concotion_pour() %>%
    samovar_boil(N = n)
}

syn_healthy <- generate_samples(healthy_data, 500)
syn_diseased <- generate_samples(diseased_data, 500)

3. Build ML Training Set

Combine real and synthetic data with labels:

# Create labeled datasets
healthy_df <- syn_healthy$data %>% 
  mutate(label = "healthy")
diseased_df <- syn_diseased$data %>% 
  mutate(label = "diseased")

# Merge and shuffle
ml_data <- bind_rows(healthy_df, diseased_df) %>% 
  sample_frac(1)  # Shuffle rows

4. Train a Classifier

Example using Random Forest:

library(randomForest)

# Split data
train_idx <- sample(1:nrow(ml_data), 0.7*nrow(ml_data))
train_data <- ml_data[train_idx, ]
test_data <- ml_data[-train_idx, ]

# Train model
rf_model <- randomForest(
  x = select(train_data, -label),
  y = as.factor(train_data$label),
  importance = TRUE
)

5. Evaluate Performance

# Predictions
preds <- predict(rf_model, test_data)

# Confusion matrix
caret::confusionMatrix(preds, as.factor(test_data$label))

---

Key Applications

1. Addressing Class Imbalance

---|--- |

# Check class distribution
table(ml_data$label)

2. Feature Importance Analysis

varImpPlot(rf_model, main = "Top Predictive Taxa")

3. Dimensionality Reduction (UMAP/PCA)

library(umap)
umap_results <- umap(select(ml_data, -label))
plot(umap_results$layout, col = as.factor(ml_data$label))

---

Pro Tips

🔥 Hyperparameter Tuning

# Use tidymodels for advanced tuning
library(tidymodels)
rf_spec <- rand_forest(mtry = tune(), trees = 1000) %>% 
  set_engine("randomForest") %>% 
  set_mode("classification")

📈 Compare With/Without Augmentation

Metric	Original Data	+samova.R Data
Accuracy	0.96	0.99
F1-Score	0.97	0.99

---

FAQ

❓ How many synthetic samples should I generate?
→ Start with 2-5x your original sample size. Monitor validation accuracy.

❓ Can I use this for regression tasks?
→ Yes! Replace classification labels with continuous values.

❓ How to handle overfitting?
→ Use regularization (e.g., LASSO) or reduce synthetic sample complexity.

---

Next Steps

1. Try the examples
2. Experiment with different ML models (XGBoost, SVM)
3. Contribute your use cases to the GitHub wiki!

💡 Found a bug? Open an issue on GitHub