ML_Lablator

A Streamlit app for interactive Exploratory Data Analysis and Machine Learning model training on any tabular dataset you upload — no code required.

Features

• Upload a CSV/Excel file and browse it (Home page).
• EDA page: missing-value / duplicate / outlier detection and cleanup, dtype fixes (including automatic date-column detection), manual data selection (by row/column slice, filter, text search, or pandas query), descriptive statistics, and a large library of Seaborn/Plotly charts.
• Model Training page: build a preprocessing + model pipeline (scaling, encoding, imputation), train any of 24 classification / regression / clustering algorithms with a UI-generated hyperparameter form, evaluate with train/test split and optional k-fold cross validation, and inspect confusion matrix / ROC curve / learning curve / actual-vs-predicted / cluster scatter plots.

Project layout

main_page.py                 # st.navigation entry point
app/pages/
    home_page.py              # file upload / landing page
    page_2.py                 # Exploratory Data Analysis
    page_3.py                 # Model Training
Core/
    eda.py                     # EDA, handle_MissingValue, handle_outliers, data_manipulation
    Preprocessor.py            # DataPreprocessor: scaling/encoding/imputation + train/test split
    model_trainer.py           # ModelPipelineBuilder + ModelParameterFactory (hyperparameter UI)
    evaluator.py                # Evaluator: fit/predict/evaluate/cross-validate any task type
components/
    charts.py                  # seaborn_chart, plotly_charts — reusable plot builders
    metrics_plots.py           # MetricPlot — confusion matrix / ROC / learning curve / clusters
Untils/
    Until.py                    # save_data / save_model helpers

Setup

python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt
streamlit run main_page.py

Changelog — bug-fix & refactor pass (2026-08-09)

This pass focused on: (1) fixing real, verified bugs, (2) documenting every public class/function that lacked a docstring, and (3) turning two already-half-built features into working ones, without changing the app's overall structure or UI flow.

Bugs fixed

#	File	Issue	Fix
1	Untils/Until.py	"Data\Proccessed_data" is a Windows-only literal path (backslash isn't a separator on Linux/macOS), so save_data silently wrote to the wrong folder off Windows.	Rebuilt with os.path.join; also removed unused imports.
2	Core/eda.py — change_dtype_datetime64	check_date_object returns either False or (True, col_name). The caller did ch, data_col = check_date_object(...) unconditionally, which raises TypeError: cannot unpack non-iterable bool object whenever no date column is found.	Now checks isinstance(result, tuple) before unpacking, and always returns the DataFrame (previously fell through to None in that branch).
3	Core/eda.py — check_dtype_column	Fell through to an implicit return None for any dtype that isn't object or number (e.g. bool, datetime64). Callers treat the result as a 0/1 flag (== 0 → text filter, else → numeric filter), so None silently took the numeric branch and broke on boolean/date columns.	Non-object dtypes now explicitly return 1.
4	Core/model_trainer.py — ModelParameterFactory.PARAMS_FACTORY	The "Extra Tree" classifier had no entry in this mapping (only its regressor twin did), even though its hyperparameter function extra_tree was already implemented. Selecting it in the UI silently trained with library defaults and no visible controls.	Added the missing "Extra Tree": extra_tree entry.
5	Core/model_trainer.py — logistic_regression, elasticnet	l1_ratio sliders allowed 0.0–20.0; scikit-learn requires [0.0, 1.0] and raises ValueError above that.	Sliders now capped at 1.0.
6	Core/model_trainer.py — Neural_Network_regressor	An alpha slider was given key="" (empty string), risking a duplicate/empty widget key collision.	Given a unique key.
7	Core/evaluator.py — Evaluator.evaluate	AUC was always computed as roc_auc_score(y_test, predict_proba[:, 1]), which only works for binary classification; any 3+-class target raised ValueError.	Branches on the number of classes: binary keeps the original behavior, multiclass uses roc_auc_score(..., multi_class="ovr", average="weighted").
8	app/pages/page_3.py	y.ravel() was called on a pandas.Series in three places. Series.ravel() was removed in modern pandas (crashes with AttributeError on pandas ≥ 2.x, confirmed against pandas 3.0).	Replaced with y.to_numpy().ravel().
9	app/pages/page_2.py	A caught exception (except Exception as E) discarded E and showed a generic message, unlike every other error handler on the page, making failures hard to diagnose.	The message now includes {E}.
10	app/pages/page_2.py	4 f-strings had no placeholders (harmless but misleading — flagged by ruff as F541).	Cleaned up.
11	Core/Preprocessor.py — get_config_df	The local dict was named data, shadowing the data: DataFrame parameter. It happened to work (Python evaluates the right-hand side before rebinding the name) but was confusing and fragile to future edits.	Renamed to config_rows.

All fixes were verified with a small functional test script (constructing real pipelines, evaluating a 3-class classifier, converting a date column, etc.) — see the bottom of this section for the kind of checks that were run.

Previously unfinished features, now completed

• MetricPlot.clustering_visualization was an empty pass stub — the Model Training page showed a metrics table for K-Means/DBSCAN but no plot. It's now implemented: a 2D scatter of the clustered points (PCA-projected when there are more than 2 features), colored by cluster, with DBSCAN's noise points (label -1) shown separately. Wired into page_3.py under a new "Show Cluster Visualization" expander.
• Learning curve confidence bands: plot_learning_curve was already computing train_std/val_std (flagged as unused by ruff) but never plotting them. It now draws a shaded ±1 standard-deviation band around each curve, which is the standard way to show cross-validation noise on a learning curve.

Known limitation intentionally left as-is

Core/eda.py's handle_MissingValue.fill_SimpleImputer (a fully implemented, documented method for imputing missing values with scikit-learn's SimpleImputer) is not called anywhere in page_2.py — the EDA page currently only offers row/column deletion for missing values, not imputation. This wasn't touched in this pass since it's a feature gap rather than a bug, but see the suggestions below.

Suggested next features

1. Wire up fill_SimpleImputer in the EDA page — it already exists and is documented; add a third option next to "delete by row/column" for "impute" with a strategy selector.
2. Model persistence — Untils/Until.save_model is fully implemented but the "Save model" section in page_3.py is commented out. Re-enable it so trained pipelines can be downloaded as .joblib and reloaded later for inference on new data.
3. Batch/inference page — a new page that loads a saved .joblib pipeline and lets the user upload new data (or fill a form) to get predictions, closing the loop from "train" to "use."
4. Feature importance / SHAP view — for tree-based models (RandomForest, XGBoost, LightGBM, ExtraTrees), show built-in feature importances; optionally add SHAP summary plots for deeper explainability.
5. Multiclass ROC — plot_roc_curve currently only handles binary targets (predict_proba[:, 1]); extend it to one-vs-rest curves per class for multiclass problems, mirroring the AUC fix already made in Evaluator.
6. Session-based experiment tracking — keep a running table (in st.session_state) of every model trained in the session with its metrics, so users can compare runs side by side instead of only seeing the latest one.
7. Config export/import — let a user export the full pipeline configuration (get_config_df output) as JSON/YAML and re-import it to reproduce a run without re-clicking every widget.
8. Data profiling report export — a "Download EDA report" button that renders the key EDA findings (missing values, duplicates, outliers, describe table) to a PDF/HTML, useful for sharing with non-technical stakeholders.
9. Class imbalance handling — given your recurring interest in RandomUnderSampler/SMOTE, this would fit naturally as an optional preprocessing step alongside scaling/encoding on the Model Training page.