Text Results Report

News Category text classification report comparing scalable ML baselines, reduced-feature pipelines, and transformer fine-tuning on held-out test performance.

EDA Results

Overall Comparison

Traditional ML Baselines

1

Input Text

Headline and short description are combined into one document

2

Vectorize

Word TF-IDF, character TF-IDF, and hashing-based sparse features

3

Classify

Scalable linear classifiers and ensemble baselines for sparse text features

4

Evaluate

Accuracy, macro-F1, weighted-F1, precision, and recall on the test split

Evaluation protocol: train/test split only. These baselines are used for direct held-out comparison, without a separate validation stage.

Feature Reduction Pipeline Grid

1

Feature Extraction

Bag-of-Words and TF-IDF representations with unigram and bigram terms

2

Dimensionality Reduction

Chi-square feature selection and TruncatedSVD projections

3

Classifier

Linear LR/SVC/SGD models, with MLP evaluated on dense SVD features

4

Grid Ranking

Candidate pipelines are ranked by macro-F1 on the held-out test set

Evaluation protocol: train/test split only. The grid is presented as an empirical comparison of reduced-feature pipelines, not as validation-based model selection.

BERT Fine-Tuning Grid

1

Input Text

Combined headline and description, truncated or padded to 128 tokens

2

Tokenizer

Checkpoint-specific WordPiece tokenization for BERT-family encoders

3

Encoder

BERT and DistilBERT encoders are fine-tuned end to end

4

Pooling + Head

CLS, mean, or pooler-style representation with dropout and a linear head

5

Evaluate

The best validation checkpoint is evaluated once on the final test split

Evaluation protocol: train/validation/test split. Validation macro-F1 is used for checkpoint selection before reporting final test metrics.

Best Model Per-Class Report

Per-Class Metrics

Error Samples