--- license: mit library_name: keras pipeline_tag: tabular-classification tags: - tabular-classification - binary-classification - keras - tensorflow - scikit-learn - mlp - missing-values - education metrics: - accuracy - f1 model-index: - name: mlp-tabular-binary-classifier results: - task: type: tabular-classification dataset: name: base_classificacao_binaria (held-out test split) type: custom metrics: - type: accuracy value: 0.81 --- # MLP tabular binary classifier (Keras) A small Multilayer Perceptron (16-8-1) for binary classification of noisy tabular sensor data, trained with a leakage-safe preprocessing pipeline (per-column imputation + scaling fitted on the training split only). Built as part of an Artificial Intelligence course at UDESC (Brazil). - **Code, notebooks and full write-up:** https://github.com/lucasomac0/ann-classification-machine-learning - **Paper (SBC format, Portuguese):** `docs/relatorio.pdf` in the GitHub repository ## Model description | | | |---|---| | Architecture | Dense(16, ReLU, L2 0.001) → Dense(8, ReLU, L2 0.001) → Dense(1, sigmoid) | | Loss / optimizer | Binary cross-entropy / Adam | | Training | batch size 64, up to 100 epochs, early stopping on `val_loss` (patience 15, best weights restored), 20% of the training set held out for validation | | Preprocessing | `ColumnTransformer`: mean / median / KNN (k=5) / mode imputation chosen per column, then `StandardScaler`; the categorical `environment` column is ordinal-encoded (cold=0, mild=1, hot=2) | | Output | probability of the positive class; predicted class = probability ≥ 0.5 | ## Intended use and limitations This is an **educational / portfolio** model. It is not intended for real decisions. - The dataset has 10,000 samples, 10 numeric features, 1 categorical feature, 4%–15% missing values per column and outliers. Feature-target linear correlations are weak (about 0.2–0.3 in absolute value). - Accuracy is around 81% and did not improve with more capacity or different preprocessing (see below). The paper discusses this as consistent with an information limit in the features, but that is a hypothesis, not a proof. - Each variant was trained once and evaluated on one test split, so differences of about 1 percentage point are within sampling noise. ## How to use The repository expects two files: the trained network (`model.keras`) and the fitted preprocessor (`preprocessor.joblib`). New data must have the same columns as the training data, with `environment` already ordinal-encoded. ```python import joblib import keras import pandas as pd from huggingface_hub import hf_hub_download repo = "lucasomac/mlp-tabular-binary-classifier" model = keras.saving.load_model(hf_hub_download(repo, "model.keras")) preprocessor = joblib.load(hf_hub_download(repo, "preprocessor.joblib")) X_new = pd.DataFrame(...) # same columns as the training features proba = model.predict(preprocessor.transform(X_new)).ravel() pred = (proba >= 0.5).astype(int) ``` ## Evaluation Held-out test set: 2,000 samples (991 class 0, 1,009 class 1), stratified split, test set used only once for the final evaluation. Version 1.2 is the model published here. | Version | Change | Accuracy (95% Wilson CI) | Sensitivity | Specificity | F1 | |---|---|---|---|---|---| | 1.1 | 16-8-1, L2, StandardScaler, ordinal encoding | 81.60% (79.8–83.2) | 81.67% | 81.53% | 81.75% | | **1.2** | one-hot encoding of `environment` | **81.00%** (79.2–82.7) | 82.56% | 79.41% | 81.43% | | 1.3 | RobustScaler | 80.55% (78.8–82.2) | 81.17% | 79.92% | 80.81% | | 1.4 | 64-32-16, BatchNorm, Dropout 0.2, swish (no L2) | 81.30% (79.5–82.9) | 80.77% | 81.84% | 81.34% |