Toronto Datathon 2024
Date: 3/2/2025
Team Name: ByteBuilders
Team Members:
- Zhongze (August) Zheng [GitHub]
- Zhengyu (Joey) Wang [GitHub]
- Zixiang (Terry) Huang [GitHub]
- Yuqing (Janice) Liu [GitHub]
-
Problem Statement: The Toronto real estate market is highly dynamic, influenced by factors such as neighborhood amenities, public transit proximity, economic conditions, and market trends. However, predicting real estate prices accurately remains a significant challenge due to market complexity, missing data, and potential biases. The goal of this project is to develop a machine learning model that can accurately predict real estate prices across Toronto's diverse neighborhoods by analyzing various influencing factors.
-
Solution Summary: To address this challenge, our project utilizes data science and machine learning techniques to develop a real estate price prediction model. Our solution involves:
- Comprehensive Data Analysis – Identifying key factors affecting property prices, such as location, building age, and amenities.
- Feature Engineering & Model Development – Implementing a robust machine learning model (e.g., XGBoost, Random Forest) to generate accurate price predictions.
- Visualization & Insights – Presenting findings through interactive data visualizations to help homebuyers, investors, and real estate agents make informed decisions.
- Performance Evaluation & Optimization – Ensuring high accuracy through hyperparameter tuning and handling missing data effectively.
-
Datathon Theme Alignment: This project aligns with the Toronto Datathon themes by:
- Providing actionable insights into the real estate market – Helping users understand how market trends, location, and property attributes affect housing prices.
- Leveraging data science and AI – Using machine learning models to analyze housing data, predict future prices, and uncover market patterns.
- Supporting informed decision-making – Offering insights for homebuyers, investors, and urban planners to optimize their strategies.
-
Key Features:
- Data-Driven Price Prediction – A machine learning model trained on real estate data to forecast property prices accurately.
- Location-Based Analysis – Evaluating the impact of neighborhood amenities, public transport access, and crime rates on real estate prices.
- Market Trend Insights – Identifying key trends in price fluctuations, the influence of economic conditions, and property type comparisons.
- Interactive Visualizations – Dynamic dashboards displaying real estate trends, pricing distribution, and geographic insights for better decision-making.
- Scalable & Reproducible Model – Well-documented machine learning pipeline that can be expanded for future real estate analyses.
-
Dataset Source: The dataset used in this project includes real estate listings and market data for Toronto. The data comes from:
- Competition-provided dataset – Official Datathon dataset containing property details, pricing, and neighborhood features.
- Publicly available datasets – Supplementary data sources such as:
- Overpass Turbo API (OpenStreetMap) OpenStreetMap. (n.d.). Overpass Turbo Query Service. Retrieved March 1, 2025, from https://overpass-turbo.eu/
- Toronto Open Data Portal – Neighbourhood Crime Rates City of Toronto. (2025). Neighbourhood crime rates dataset. Toronto Open Data Portal. Retrieved March 1, 2025, from https://open.toronto.ca/dataset/neighbourhood-crime-rates/
-
Data Cleaning & Preprocessing: To ensure a high-quality dataset, we performed comprehensive data preprocessing, including:
- Handling Missing Values – Imputed missing values using median imputation for numerical features (e.g., building age, lot size) and mode imputation for categorical variables.
- Feature Engineering – Created new variables such as distance to nearest subway station, crime rate per neighborhood, and property price per square foot.
- Outlier Detection & Removal – Identified extreme property prices using IQR-based filtering to remove unrealistic listings.
git clone https://github.com/zzz403/SDSS-Datathon-Project.git
cd datathon-projectpip install -r requirements.txtstreamlit run src/app.py-
Model Used: We experimented with multiple models and selected the most effective one based on performance evaluation:
-
XGBoost (Extreme Gradient Boosting) - Final model
-
K-Nearest Neighbors (KNN) - Used for region classification
-
Linear Regression (Baseline Model) - For performance comparison
-
Random Forest (Initial Experimentation) - Tried but not selected
-
-
Hyperparameter Tuning: To optimize our model, we performed Grid Search and Cross-Validation on key hyperparameters:
| Hyperparameter | Tuned Value | Purpose |
|---|---|---|
learning_rate |
0.08 | Controls step size for each boosting iteration |
max_depth |
8 | Limits tree depth to prevent overfitting |
subsample |
0.8 | Uses a fraction of the training data per boosting round |
colsample_bytree |
0.8 | Selects a subset of features per tree |
- Performance Metrics:
To evaluate model effectiveness, we used the following key metrics:
| Model | R² Score |
|---|---|
| XGBoost | 0.90 |
| Random Forest | 0.85 |
| KNN | 0.75 |
| Linear Regression | 0.68 |
For a comprehensive report, please see our official Report.
| Criterion | Score (Out of 20%) | Explanation |
|---|---|---|
| Relevance | ✅ | Clearly aligned with [Datathon theme] |
| Feasibility | ✅ | Solution is practical and implementable |
| Creativity | ✅ | Unique approach and visualization |
| Code Quality | ✅ | Code is well-structured and documented |
| Executability | ✅ | The project runs without major issues |
Challenges:
- How to deal with a bunch of raw data in file to find the correlation between each variable and make a future prediction on the targeted data is the big problem we are facing. Therefor, the first thing we did was data cleaning and feature engineering.
- Which model is best fit? That's need a lot of accuracy comparison and evaluation on each model. We tried to use different types of models，like supervised-learning and clustering model, and evaluate by different types of graph.
Future Improvements:
- To enhance model diversity and robustness, we propose using the Random Subspace Method for future adjustments, which could mitigate imbalanced feature of our model and improve generalization.
- The sample size of our model is small. To generate a more accurate prediction, work on higher size of sample can achieve our goal!
This project is licensed under the GNU License. For more details, please refer to the LICENSE file.
Information regarding third-party licenses can be found in the THIRD-PARTY file.