AUTOML_DOCUMENTATION.md

🚀 DataVision AutoML - Complete Documentation

Table of Contents

1. Overview
2. Training Modes
3. Supervised Learning Algorithms
4. Unsupervised Learning Algorithms
5. NLP Algorithms
6. Deep Learning Architectures
7. Charts & Visualizations
8. API Endpoints

---

Overview

DataVision AutoML is a production-grade machine learning platform supporting:

• 30+ ML Algorithms (Classification, Regression, Clustering)
• Auto GPU/CPU Detection (CUDA, ROCm, Metal)
• Advanced Data Cleaning (Smart Imputation, Outlier Detection)
• Feature Selection (Variance, Correlation, Mutual Info, RFE)
• Bayesian Hyperparameter Optimization (Optuna)
• Ensemble Methods (Stacking, Voting, Blending)
• 20+ Visualization Charts

---

Training Modes

1. Fast Mode (Default)

• Speed: ~30-60 seconds
• Algorithms: Top 5 performing algorithms
• CV Folds: 3-fold cross-validation
• Use Case: Quick prototyping, small datasets

2. Ultra Mode

• Speed: 2-10 minutes
• Algorithms: All 15+ algorithms
• CV Folds: 5-fold stratified cross-validation
• Features: Hyperparameter tuning, ensemble stacking
• Use Case: Production models, maximum accuracy

3. Multi-Mode Training

• Combines: Traditional ML + NLP + Deep Learning
• Parallel Training: Runs selected modes simultaneously
• Best Model Selection: Cross-mode comparison

---

Supervised Learning Algorithms

Classification Algorithms

Algorithm	Parameters	Use Case	Chart Support
XGBoost	n_estimators=100-500, max_depth=3-10, learning_rate=0.01-0.3, subsample=0.6-1.0, colsample_bytree=0.6-1.0, min_child_weight=1-10, gamma=0-5, reg_alpha=0-1, reg_lambda=0-1	Large datasets, tabular data	✅ All
LightGBM	n_estimators=100-500, max_depth=3-15, learning_rate=0.01-0.3, num_leaves=20-150, subsample=0.6-1.0, colsample_bytree=0.6-1.0, min_child_samples=10-100, reg_alpha=0-1, reg_lambda=0-1	Fast training, large data	✅ All
CatBoost	iterations=100-500, depth=4-10, learning_rate=0.01-0.3, l2_leaf_reg=1-10, border_count=32-255, random_strength=0-10	Categorical features, GPU	✅ All
Random Forest	n_estimators=100-500, max_depth=5-30, min_samples_split=2-20, min_samples_leaf=1-10, max_features='sqrt'/'log2'/0.5-1.0, bootstrap=True/False	General purpose, interpretable	✅ All
Gradient Boosting	n_estimators=100-300, max_depth=3-8, learning_rate=0.01-0.2, subsample=0.6-1.0, min_samples_split=2-20, min_samples_leaf=1-10	Balanced performance	✅ All
Extra Trees	n_estimators=100-500, max_depth=5-30, min_samples_split=2-20, min_samples_leaf=1-10, max_features='sqrt'/'log2'/0.5-1.0	Fast random splits	✅ All
Hist Gradient Boosting	max_iter=100-500, max_depth=3-15, learning_rate=0.01-0.3, min_samples_leaf=10-50, l2_regularization=0-1	Fast, native categorical	✅ All
AdaBoost	n_estimators=50-200, learning_rate=0.01-1.0, algorithm='SAMME'/'SAMME.R'	Boosting weak learners	✅ All
Bagging	n_estimators=10-50, max_samples=0.5-1.0, max_features=0.5-1.0, bootstrap=True/False	Variance reduction	✅ All
Logistic Regression	C=0.001-100, penalty='l1'/'l2'/'elasticnet', solver='lbfgs'/'saga'/'liblinear', max_iter=100-1000, l1_ratio=0-1	Linear baseline, interpretable	✅ All
SVC (SVM)	C=0.1-100, kernel='rbf'/'linear'/'poly'/'sigmoid', gamma='scale'/'auto'/0.001-10, degree=2-5, class_weight='balanced'/None	Non-linear boundaries	✅ All
Linear SVC	C=0.1-100, penalty='l1'/'l2', loss='hinge'/'squared_hinge', max_iter=1000-10000	Large-scale linear	✅ All
K-Neighbors	n_neighbors=3-15, weights='uniform'/'distance', metric='euclidean'/'manhattan'/'minkowski', p=1-2, leaf_size=20-50	Instance-based	✅ All
Gaussian NB	var_smoothing=1e-9-1e-6	Text, probabilistic	✅ All
Multinomial NB	alpha=0.01-1.0, fit_prior=True/False	Text classification	✅ All
Complement NB	alpha=0.01-1.0, norm=True/False	Imbalanced text	✅ All
Bernoulli NB	alpha=0.01-1.0, binarize=0.0-1.0	Binary features	✅ All
MLP Classifier	hidden_layer_sizes=(64,32)-(256,128,64), activation='relu'/'tanh', solver='adam'/'sgd', alpha=0.0001-0.01, learning_rate='constant'/'adaptive', max_iter=200-1000, early_stopping=True	Neural network	✅ All
Decision Tree	max_depth=3-20, min_samples_split=2-20, min_samples_leaf=1-10, criterion='gini'/'entropy', splitter='best'/'random'	Interpretable rules	✅ All
LDA	solver='svd'/'lsqr'/'eigen', shrinkage='auto'/0-1	Dimensionality reduction	✅ All
QDA	reg_param=0-1	Quadratic boundaries	✅ All
SGD Classifier	loss='hinge'/'log_loss'/'perceptron', penalty='l1'/'l2'/'elasticnet', alpha=0.0001-0.01, max_iter=1000-5000, learning_rate='optimal'/'constant'/'adaptive'	Large-scale online	✅ All
Passive Aggressive	C=0.1-10, max_iter=1000-5000, loss='hinge'/'squared_hinge'	Online learning	✅ All

Regression Algorithms

Algorithm	Parameters	Use Case
XGBoost Regressor	Same as classifier	General regression
LightGBM Regressor	Same as classifier	Fast regression
CatBoost Regressor	Same as classifier	Categorical features
Random Forest Regressor	Same as classifier	General purpose
Gradient Boosting Regressor	Same as classifier	Balanced
Extra Trees Regressor	Same as classifier	Fast
Ridge	alpha=0.01-100, solver='auto'/'svd'/'cholesky'/'lsqr'	L2 regularization
Lasso	alpha=0.01-100, max_iter=1000-10000	L1 regularization, feature selection
ElasticNet	alpha=0.01-100, l1_ratio=0-1, max_iter=1000-10000	Combined L1+L2
SVR	C=0.1-100, kernel='rbf'/'linear'/'poly', epsilon=0.01-0.5, gamma='scale'/'auto'	Non-linear
Linear SVR	C=0.1-100, epsilon=0-0.5, max_iter=1000-10000	Large-scale linear
K-Neighbors Regressor	Same as classifier	Instance-based
MLP Regressor	Same as classifier	Neural network
Bayesian Ridge	alpha_1=1e-6, alpha_2=1e-6, lambda_1=1e-6, lambda_2=1e-6	Probabilistic
Huber Regressor	epsilon=1.0-2.0, alpha=0.0001-0.01, max_iter=100-500	Robust to outliers
Poisson Regressor	alpha=0-1, max_iter=100-500	Count data
Quantile Regressor	quantile=0.1-0.9, alpha=0-1	Quantile estimation
Lasso Lars	alpha=0.01-100, max_iter=500-1000	Feature selection

---

Unsupervised Learning Algorithms

Clustering Algorithms

Algorithm	Parameters	Charts Generated	Use Case
KMeans	n_clusters=2-15 (auto-detect), init='k-means++'/'random', n_init=10-30, max_iter=300-500, tol=1e-4, algorithm='lloyd'/'elkan'	Elbow, Silhouette, Scatter, Distribution, Heatmap, PCA, 3D, t-SNE, UMAP, Pairplot, Boxplot, Violin, Correlation, Radar	General clustering
DBSCAN	eps=auto-detect (k-distance), min_samples=max(3, n//100) (auto), metric='euclidean'/'manhattan', algorithm='auto'/'ball_tree'/'kd_tree'	K-Distance, Scatter, Distribution, Silhouette, Heatmap, PCA, 3D, t-SNE, UMAP, Pairplot, Boxplot, Violin, Correlation, Radar	Noise detection, arbitrary shapes
Hierarchical (Agglomerative)	n_clusters=2-15, linkage='ward'/'average'/'complete'/'single', metric='euclidean', compute_full_tree='auto'	Dendrogram, Silhouette, Scatter, Distribution, Heatmap, PCA, 3D, t-SNE, UMAP, Pairplot, Boxplot, Violin, Correlation, Radar	Hierarchical structure
GMM (Gaussian Mixture)	n_components=2-15, covariance_type='full'/'tied'/'diag'/'spherical', n_init=10, max_iter=200, init_params='k-means++'	BIC/AIC Chart, Silhouette, Scatter, Distribution, Heatmap, PCA, 3D, t-SNE, UMAP, Pairplot, Boxplot, Violin, Correlation, Radar	Probabilistic clustering
Spectral	n_clusters=2-15, affinity='nearest_neighbors'/'rbf', n_neighbors=10-30, assign_labels='cluster_qr'/'kmeans', gamma=1.0	Affinity Matrix, Silhouette, Scatter, Distribution, Heatmap, PCA, 3D, t-SNE, UMAP, Pairplot, Boxplot, Violin, Correlation, Radar	Graph-based clustering

Clustering Charts (17 Total)

Chart	Description	Size (figsize)	When Generated
cluster_scatter	PCA 2D scatter with centroids	(10, 8)	Always
elbow_method	Silhouette scores vs k	(10, 6)	KMeans/Hierarchical/GMM/Spectral
silhouette_plot	Per-sample silhouette coefficients	(10, 8)	n_clusters≥2, n_samples≥50
cluster_distribution	Bar chart of cluster sizes	(10, 6)	Always
cluster_heatmap	Feature means per cluster	(12, 8)	n_features≥3, n_clusters≥2
pca_variance	Explained variance per component	(14, 5)	n_features≥3
cluster_3d	PCA 3D scatter	(12, 10)	n_features≥3, n_samples≥50
dendrogram	Hierarchical tree structure	(14, 8)	Hierarchical only
gmm_bic_aic	BIC/AIC model selection	(10, 6)	GMM only
dbscan_kdist	K-distance graph for eps	(10, 6)	DBSCAN only
spectral_affinity	Affinity matrix heatmap	(10, 8)	Spectral only, n_samples≤500
tsne	t-SNE 2D visualization	(10, 8)	100≤n_samples≤3000, n_features≥3
umap	UMAP 2D visualization	(10, 8)	100≤n_samples≤5000, n_features≥3
pairplot	Feature pairwise scatter	auto	2≤n_features≤6, n_samples≤2000
boxplots	Feature distribution by cluster	(4*n_features, 6)	n_features≥2, n_clusters≥2
violin_plots	Violin distribution by cluster	(5*n_features, 6)	n_features≥2, n_clusters≥2, n_samples≥50
correlation_heatmap	Feature correlation matrix	(10, 8)	n_features≥3
radar_chart	Cluster comparison radar	(10, 10)	n_features≥3, n_clusters≥2

Clustering Metrics

Metric	Range	Description
silhouette_score	[-1, 1]	Higher is better, cluster cohesion vs separation
calinski_harabasz_score	[0, ∞)	Higher is better, variance ratio
davies_bouldin_score	[0, ∞)	Lower is better, average similarity
reliability_score	[0, 100]	Production intelligence quality score

---

NLP Algorithms

Text Classification Algorithms

Algorithm	Parameters	Vectorizer	Use Case
TF-IDF + Logistic Regression	C=0.1-100, max_iter=500	TF-IDF: max_features=5000-20000, ngram_range=(1,2)/(1,3), min_df=2-5, max_df=0.9-0.95	General text
TF-IDF + SVM	C=0.1-100, kernel='linear'/'rbf'	Same as above	High accuracy
TF-IDF + Naive Bayes	alpha=0.01-1.0	Same as above	Fast, probabilistic
TF-IDF + Random Forest	n_estimators=100-300	Same as above	Feature importance
TF-IDF + XGBoost	Same as classifier	Same as above	Best accuracy
BOW + Classifiers	Same as TF-IDF	CountVectorizer: same params	Simple baseline
N-gram Models	Same as TF-IDF	ngram_range=(2,3)/(1,4)	Phrase patterns

NLP-Specific Charts

Chart	Description	When Generated
confusion_matrix	Class prediction matrix	Always
word_cloud	Important words visualization	Text data
class_distribution	Target class frequencies	Always
roc_curve	ROC-AUC per class	Multi-class
precision_recall	PR curve per class	Multi-class

---

Deep Learning Architectures

MLP (Multi-Layer Perceptron) Architectures

Architecture	Hidden Layers	Parameters	Use Case
mlp_small	(64, 32)	~5K params	Small data, fast
mlp_medium	(128, 64, 32)	~15K params	Balanced
mlp_large	(256, 128, 64)	~50K params	Complex patterns
mlp_wide	(512, 256)	~150K params	High-dimensional
mlp_deep	(128, 128, 128, 128)	~70K params	Deep representation

MLP Training Parameters

Parameter	Values	Description
activation	'relu', 'tanh', 'logistic'	Activation function
solver	'adam', 'sgd', 'lbfgs'	Optimizer
alpha	0.0001-0.01	L2 regularization
learning_rate	'constant', 'adaptive', 'invscaling'	LR schedule
learning_rate_init	0.001-0.01	Initial LR
max_iter	200-1000	Max epochs
early_stopping	True	Validation-based stopping
validation_fraction	0.1	Validation split
n_iter_no_change	10	Early stopping patience
batch_size	32-256, 'auto'	Mini-batch size

Deep Learning Charts

Chart	Description	When Generated
learning_curve	Loss vs epochs	Training history
confusion_matrix	Prediction matrix	Classification
feature_importance	Input weights	All
activation_distribution	Layer activations	Debug

---

Charts & Visualizations

Supervised Learning Charts (Classification)

Chart Key	Description	Size	Parameters
confusion_matrix	True vs Predicted heatmap	(10, 8)	cmap='Blues', annot=True, fmt='d'
feature_importance	Top 15 features bar chart	(12, 8)	Sorted descending
roc_curve	ROC curves per class	(10, 8)	lw=2, AUC in legend
precision_recall	PR curves per class	(10, 8)	AP in legend
class_distribution	Target class bar chart	(10, 6)	With percentages
learning_curve	Train/Val score vs size	(10, 6)	CV=5
calibration_curve	Predicted vs actual prob	(10, 8)	Perfectly calibrated line
lift_curve	Cumulative gains	(10, 8)	Baseline comparison
correlation_heatmap	Feature correlations	(12, 10)	cmap='coolwarm'
shap_summary	SHAP feature values	(12, 10)	Requires SHAP

Supervised Learning Charts (Regression)

Chart Key	Description	Size	Parameters
actual_vs_predicted	Scatter with ideal line	(10, 8)	alpha=0.5
residual_plot	Residuals vs predicted	(10, 6)	Zero line
residual_distribution	Residual histogram+KDE	(10, 6)	Normal curve overlay
feature_importance	Top 15 features	(12, 8)	Sorted descending
error_by_range	MAE per target bin	(10, 6)	10 bins
qq_plot	Quantile-quantile	(8, 8)	Normal reference
learning_curve	Train/Val RMSE vs size	(10, 6)	CV=5
cook_distance	Influential points	(10, 6)	Threshold line

Chart Generation Parameters

# Common matplotlib settings
plt.style.use('seaborn-v0_8-whitegrid')
plt.rcParams['figure.dpi'] = 150
plt.rcParams['savefig.dpi'] = 150
plt.rcParams['savefig.bbox'] = 'tight'
plt.rcParams['savefig.facecolor'] = 'white'

# Color palettes
CLASSIFICATION_COLORS = ['#2563eb', '#16a34a', '#dc2626', '#ea580c', '#9333ea', 
                         '#0891b2', '#db2777', '#d97706', '#0d9488', '#4f46e5']
REGRESSION_COLORS = ['#3b82f6', '#ef4444']
CLUSTER_COLORS = ['#2563eb', '#16a34a', '#dc2626', '#ea580c', '#9333ea', 
                  '#0891b2', '#db2777', '#d97706', '#0d9488', '#4f46e5',
                  '#84cc16', '#06b6d4', '#f43f5e', '#8b5cf6', '#14b8a6']

---

API Endpoints

Supervised Learning

Endpoint	Method	Mode	Description
/api/v1/automl/production_train	POST	Fast	Quick training with top algorithms
/api/v1/automl/train	POST	Standard	Full training pipeline
/api/v1/automl/ultra_train	POST	Ultra	Maximum accuracy mode
/api/v1/automl/multi_mode/train	POST	Multi	Combined Traditional+NLP+Deep
/api/v1/automl/predict	POST	-	Single prediction
/api/v1/automl/batch_predict	POST	-	Batch predictions
/api/v1/automl/saved-result	GET	-	Load saved model results
/api/v1/automl/stop_training	POST	-	Cancel ongoing training

NLP

Endpoint	Method	Description
/api/v1/automl/nlp/train	POST	Text classification training
/api/v1/automl/nlp/predict	POST	Text prediction

Deep Learning

Endpoint	Method	Description
/api/v1/automl/deep_learning/train	POST	Neural network training
/api/v1/automl/deep_learning/predict	POST	Deep learning prediction

Unsupervised Learning

Endpoint	Method	Description
/api/v1/ml/clustering	POST	Clustering analysis
/api/v1/ml/clustering/predict	POST	Predict cluster for new data
/api/v1/ml/clustering/download-model/{user_id}	GET	Download PKL model
/api/v1/ml/clustering/download-data/{user_id}	GET	Download clustered CSV

---

Production Intelligence

Reliability Score (0-100)

Computed based on:

• Data quality (missing values, outliers)
• Model validation (CV performance, overfitting check)
• Feature quality (variance, correlation)
• Sample size adequacy

Validation Warnings

Warning Type	Trigger
small_dataset	n_samples < 100
high_cardinality	categorical unique > 50%
class_imbalance	minority class < 10%
missing_values	missing > 20%
potential_leakage	feature corr > 0.95 with target
low_variance	feature variance ≈ 0

---

File Outputs

Saved Artifacts

File	Location	Content
best_model.pkl	storage/users/{user_id}/models/	Trained model + metadata
cleaned_data.csv	storage/users/{user_id}/files/	Preprocessed data
multimode_metadata.json	storage/users/{user_id}/models/	Training configuration
clustering_model.pkl	storage/users/{user_id}/models/	Clustering model + scaler
clustered_data.csv	storage/users/{user_id}/files/	Data with cluster labels

---

Version Info

• Engine Version: 7.0
• Last Updated: February 2026
• Supported Python: 3.11+
• Key Dependencies: scikit-learn>=1.3.0, xgboost>=2.0.0, lightgbm>=4.2.0, catboost>=1.2.0, optuna>=3.4.0