Attention-based multimodal deep learning for Turkish real estate price prediction using structured, text, and image data
Neural Computing and Applications, cilt.38, sa.15, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 38 Sayı: 15
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s00521-026-12360-0
- Dergi Adı: Neural Computing and Applications
- Derginin Tarandığı İndeksler: Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, Index Islamicus, INSPEC, zbMATH, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Automated valuation, Deep learning, Multimodal learning, Natural language processing, Real estate price prediction, Transformers
- Atatürk Üniversitesi Adresli: Evet
Özet
Accurate real estate valuation in volatile markets presents significant challenges. Turkey exemplifies this complexity: 60%+ annual inflation persists, yet prior studies relied predominantly on tabular data—with limited use of text descriptions and property photos. We developed a multimodal framework integrating three branches across 74,815 residential properties spanning 74 Turkish provinces (91.4% national coverage). For structured data, an Attention-based Transformer DNN processes 214 features spanning spatial, physical, socioeconomic, location quality, demographic, temporal, and derived property attributes. Turkish text receives dual treatment: BERTurk semantic embeddings combined with 13 statistical features (character/word/sentence counts, lexical diversity metrics, structural patterns). Interior photographs undergo context-guided Swin Transformer analysis, conditioning visual processing on five property scalars (net square meter, building age, floor number, room count, and bathroom count) to improve valuation performance. Learnable missing tokens handle incomplete data—approximately 7.1% of listings lack complete room photographic coverage. Multiple fusion strategies underwent systematic evaluation, including learned weighting, attention-based, and stacking ensemble strategies. Results: full trimodal integration reaches R² = 0.9029 on held-out test data, surpassing the structured-only branch (R² = 0.8875) and prior Turkish benchmarks (R² = 0.80). The improvement corresponds to an approximately 14% reduction in unexplained variance relative to the structured branch. Ablation studies across all seven modality combinations (S, T, V, S + T, S + V, T + V, S + T+V) quantify individual contributions—each modality captures complementary valuation-relevant signals. Text alone achieves R² ≈ 0.870, strong for qualitative descriptions. Context-guided vision adds complementary aesthetic and visual-quality signals when fused, improving the best structured-text configuration from R² = 0.8973 to R² = 0.9029 on held-out test data. For high-inflation market contexts, this framework offers a practical, deployable valuation solution.