Multimodal House Price Prediction for New Jersey: Redfin-Based Dataset and Model Artifacts

Multimodal House Price Prediction for New Jersey: Redfin-Based Dataset and Model Artifacts Description Overview This dataset supports a multimodal machine learning project for predicting residential sale prices in New Jersey using both tabular and text features derived from Redfin listings. It contains: Raw and cleaned Redfin export data for New Jersey home sales (2015–2025), Train/validation/test splits for multimodal modeling, A unified text field combining listing description, address, and nearby schools, Text encoder artifacts (SentencePiece tokenizer, encoder weights, precomputed text embeddings), Metadata files documenting feature sets, targets, and data preparation parameters. The associated code and notebooks implementing the full pipeline (scraping, EDA, data preparation, tabular baselines, text encoder, and multimodal TabNet fusion) are available in the companion GitHub repository. Data contents The Zenodo archive mirrors the local project structure described in the README. Key components include: Raw data (data/raw/) redfin_nj_sold_2015-01-01_to_2025-11-24.csv: Raw Redfin export for New Jersey residential sales. FEDFUNDS_151001_to_251001.csv: Daily Federal Funds Effective Rate time series, used as a macroeconomic feature. Processed tabular data (data/processed/) redfin_nj_sold_2016plus_basic_clean.csv: Cleaned tabular dataset (2016+ sales) after EDA and filtering. train_multimodal.csv, val_multimodal.csv, test_multimodal.csv: Final multimodal splits used in all modeling notebooks. multimodal_prep_summary.json: JSON summary documenting: Target column names (sold_price, log_sold_price = log1p(sold_price)), Numeric and categorical tabular feature lists, Split sizes and random seed, Key preprocessing parameters. Text encoder and features (data/processed/multimodal_features/) sentencepiece_bpe/sp_bpe_16000.model, sentencepiece_bpe/sp_bpe_16000.vocab: SentencePiece BPE tokenizer (vocab size 16k) trained on listing text. sentencepiece_bpe/train_text.txt: Text corpus used to train the tokenizer. text_encoder_best.pt, text_encoder_spbpe_best_vocab16000.pt: Trained text encoder weights (Transformer-based model predicting normalized log sale price from text). txt_features_meta_spbpe_vocab16000_d128.json: Metadata describing the text embeddings (tokenization config, embedding dimension, etc.). txt_features_train_spbpe_vocab16000_d128.csv, txt_features_val_spbpe_vocab16000_d128.csv, txt_features_test_spbpe_vocab16000_d128.csv: Precomputed text embeddings (one vector per listing) aligned with the train/val/test splits. multimodal_dl_prep_params.json: Additional parameters used for multimodal deep learning preparation (e.g., normalization constants, encoder hyperparameters). Intended use This dataset is intended for: Reproducing the results of the Multimodal House Price Prediction (New Jersey) project, Benchmarking new models on a real-world multimodal regression task, Methodological research on: Combining tabular and text features, Multimodal fusion architectures (e.g., TabNet + text embeddings), Effects of macroeconomic variables (Fed Funds rate) on house price modeling. The recommended way to use this dataset is to: Download and extract the archive into the root of the GitHub project, so that data/ matches the layout described above. Run the notebooks in order: 01.query.ipynb (optional if using provided raw data), 02_eda.ipynb, 03_data_prep.ipynb, 04a_LR_baseline.ipynb, 04_TabNet_tabularONLY.ipynb, 05_text_encoder.ipynb, 06_TabNet_fusion.ipynb. Use multimodal_prep_summary.json as the primary reference for feature sets and targets when building additional models. Please refer to the GitHub README for environment setup, dependency details, and a step-by-step reproduction guide.