AI House Price Prediction

• This project aims to build an AI-based predictive model for house prices using machine learning techniques.
• Traditional models like Comparative Market Analysis (CMA) and hedonic pricing struggle with non-linear interactions between property attributes and market trends, leading to inaccurate pricing.
• The project uses AI models to improve pattern recognition, feature importance, and predictive accuracy.
• The project benefits buyers, sellers, investors, and financial institutions.
• Real estate price prediction is achieved using machine learning models like Random Forest, XGBoost, and Artificial Neural Networks (ANNs).
• These models utilize large datasets to discover complex correlations between property characteristics and market fluctuations.
• The project addresses challenges encountered due to data quality issues and computational limitations.
• The project adjusts for data quality problems and computational restrictions.
• The project seeks to bridge the gap between traditional pricing models and AI valuation techniques for more accurate and efficient house price predictions.

Project Progress Overview

• Key project stages are dataset selection, preprocessing, and initial model development.
• The project identifies, acquires, and reviews datasets such as Zillow House Price Dataset, Kaggle real estate data, and UK Land Registry data for quality and completeness.
• Numerical and categorical variables are encoded, missing values are handled, numerical features are normalized, and outliers are removed.
• A baseline model (Linear Regression) is developed for performance evaluation.
• Model selection and training have started with Decision Trees and Random Forest models currently evaluated.

Dataset Selection and Justification

• Datasets selected include Zillow House Price Dataset, Kaggle Real Estate Data, and UK Government Land Registry data.
• The selected datasets contain complete real estate pricing information suitable for training machine learning models.
• Zillow Research (2024) provides historical home price trends, property attributes, and location-based pricing variations across the US.
• The Kaggle dataset contains real estate data with feature engineering and model optimization, based on the UK Land Registry (2024).
• UK Land Registry dataset contains transactional price records and property sales, and is internationally applicable.
• Datasets are diverse, available, and covered in history.
• Zillow and Kaggle provide U.S. market insights.
• the UK Land Registry does cross-regional analysis.
• Key dataset variables include square footage, number of bedrooms, crime rates, school ratings, and economic indicators.
• Challenges include missing values, inconsistencies, and biases.
• Data gaps exist, especially regarding property features needing imputation.
• Inconsistencies arise from differences in regional reporting methods.
• Past price trends contain potential biases caused by socioeconomic inequality.
• Methods such as imputation, categorical encoding, and normalization are applied to preprocessing to improve data quality and reliability for AI-based house price prediction.

Project Development

• The project development phase involves setting up the computational environment, performing data preprocessing, and implementing the initial AI model.
• Python is used, with Jupyter Notebook, Scikit-learn, TensorFlow, Pandas, and NumPy as main tools.
• Jupyter Notebook provides an interactive coding environment for experimenting with preprocessing and modeling techniques.
• Data preprocessing is critical for model accuracy.
• Steps that have been completed:
• Handling Missing Values: Imputation techniques like mean/median filling for numerical features and mode-based imputation for categorical data are used.
• Categorical Encoding: Location and property type features are transformed via one-hot encoding for AI models.
• Normalization and Scaling: Min-Max scaling and other normalization techniques are used to ensure all numerical features are on a consistent scale.
• Outlier Removal: IQR (Interquartile Range) to eliminate extreme values with negative impacts on model performance.
• A baseline model based on Linear Regression is developed for comparison.
• Feature importance is analyzed using Decision Trees to determine influential factors on house prices.
• Advanced models like XGBoost and Random Forest are under development, with early results suggesting improved prediction accuracy.
• Hyperparameter optimization techniques are employed to balance bias and variance, and models are fine-tuned.

Preliminary Results

• Linear Regression is used for initial model performance evaluation.
• Root Mean Squared Error (RMSE) was 72,500.
• Mean Absolute Error (MAE) was 48,200.
• R2 score of 0.67 indicates the model explains 67% of the variance in house prices, but could do better.
• Location, square footage and number of bedrooms are leading predictive variable.
• Models is overfitting because it produces good results on training data, but not so for generalization
• Key challenges include overfitting in tree-based models and hyperparameter tuning for better performance.
• Future steps include pruning techniques, cross-validation, and hyperparameter tuning to improve generalization.
• Other ensemble models like Random Forest and XGBoost will be considered to enhance predictive accuracy.

Challenges Encountered and Adjustments Made

• Challenges include data quality and computational limitations.
• Property attributes missing values include square footage, number of bedrooms, and location-based economic indicators.
• Mean/mode imputation techniques and K Nearest Neighbors (KNN) imputation used for filling data.
• Data normalization and resampling techniques were used to solve issues of dataset bias, such as socioeconomic factors leading to bias in historical housing data.
• GPU acceleration for fast model training was integrated due to the cloud nature(Google Colab).
• Feature engineering techniques, like log transformations and polynomial creation of features were introduced to increased model performance and reducing risk of overfitting
• Adjustments to handling data has improved, model computational efficiency has been optimized, and the overall model robustness has increased.
• Models hyperparameters will be legalized to get higher accuracy.

Conclusion

• Dataset selection, preprocessing, and initial model development have been successful.
• The project was able to make it to its end state.
• Cloud-based solutions and advanced preprocessing techniques have addressed challenges like data inconsistencies and computational limitations.
• Optimization of models, hyperparameter tuning and comparative analysis will improve reliability while making comparisons to traditional pricing models.