brendon4488/CNN_Architecture_Comparisons

Comparative analysis of 12 CNN architectures for multi-modal property price prediction using Zimbabwe real estate data **CNN Architecture Comparison for Property Price Prediction** **Project Overview** A comprehensive comparative analysis of 12 CNN architectures to identify the optimal feature extraction backbone for property price prediction using Zimbabwe real estate data. Research Question Which convolutional neural network architecture provides the most discriminative visual features when combined with tabular property data for accurate price prediction? _Architectures Compared_ 12 strategically selected CNN models representing key evolutionary milestones: **Architecture Family Models** **ResNet** ResNet50, ResNet152 **EfficientNet** EfficientNetB0, EfficientNetB3 **DenseNet** DenseNet121, DenseNet201 **Mobile** MobileNetV2, NASNetMobile **Inception** InceptionV3, InceptionResNetV2 **Classic** VGG16, Xception **Dataset** 1,610 property listings from Zimbabwe market Structured features: Price, bedrooms, bathrooms, area measurements, location Visual data: Property photographs (466 actual + 1,144 median-imputed) Target variable: Property price (USD) **Methodology** **Data Preprocessing** Two-pass median imputation for missing images Feature engineering: Area ratios, price per sqm, boolean flags Text processing: Property type extraction from titles Complete data preservation strategy **Multi-Modal Fusion** Tabular features: 12 engineered attributes Visual features: CNN embeddings from each architecture Neural network with dual input streams Early stopping and consistent train/val/test splits **Results** **Performance Ranking** Rank Model Test MAE R² Features 1 DenseNet201 $631,166.94 -0.064 1,932 2 ResNet50 $631,169.31 -0.064 2,060 3 InceptionV3 $631,187.06 -0.064 2,060 12 InceptionResNetV2 $631,210.00 -0.064 1,548 **Key Findings** Minimal performance variation between architectures ($43 range) Negative R² scores across all models DenseNet family performed slightly better on average Visual features provided limited predictive value beyond tabular data