Multi-Task Learning for Real Estate
Prediction
Course: AI211/ CS503 Machine Learning
Instructor: Dr. Santosh Kumar Vipparthi
Due Date: 01-04-2025
1 Problem Scenario
You are a data scientist at BharatHomes, a real estate analytics firm focus-
ing on urban Indian markets. The dataset includes 79 features describing
residential properties across Mumbai, Delhi, Bangalore, and Chennai.
1.1 Key Features in the Indian Context
Make sure when you analyze the values for the further regularization and
optimization of your own choices. All the feature values must be in the
euclidean space to obtain your choices for the considered regularization and
optimization.
Feature Indian Contextualization
SalePrice Price in INR (lakhs/crores)
Neighborhood Mumbai: Bandra, Andheri; Delhi: Gurgaon, Noida
Condition1 Proximity to metro stations (e.g., Delhi Metro)
Utilities Water supply consistency (24/7 vs. tanker-dependent)
LandContour Flood-prone zones (e.g., Chennai’s coastal areas)
OverallQual Builder reputation (e.g., Tata, DLF)
GarageQual Security features (gated community vs. standalone)
PoolQC Presence of clubhouse amenities
Objective:
1. Regression: Predict SalePrice (INR).
2. Classification: Classify properties as ”Premium” (top 20% prices)
or ”Standard”.
1
2 Mathematical Formulation
Let the dataset be:
• Features: X ∈ Rn×79 (e.g., GrLivArea, OverallQual, Neighborhood)
• Targets:
– Regression: yr ∈ Rn (SalePrice)
– Classification: yc ∈ {0, 1}n (Premium and Standard)
2.1 Loss Functions
1. Regression: Mean Squared Error (MSE)
n
1 X (i) 2
Lr = yr − f (x(i) ) (1)
n i=1
2. Classification: Cross-Entropy
n
1 X (i)
Lc = − y log g(x(i) ) (2)
n i=1 c
3. Joint Loss
L(θ) = αLr + βLc + λ∥θ∥22 (3)
3 Optimization
3.1 Key Analysis
1. Convexity:
• Prove whether L(θ) is convex under linear models for both tasks.
• Discuss the implications of non-convexity in deep neural networks.
2. Hessian Conditioning:
• Let Hr = ∇2 Lr and Hc = ∇2 Lc . Derive the condition number of the
joint Hessian H = αHr + βHc .
• Show how κ(H) affects optimization convergence.
3. Regularization:
2
• Compare L1 (sparse feature selection) vs. L2 (smooth feature weight-
ing).
• Justify the choice of regularization strength λ using the bias-variance
trade-off.
4 Multi-Task Learning Model
1. Problem Formulation:
• Why is multi-task learning suitable for this problem?
• How does feature sharing between tasks improve model perfor-
mance?
2. Data Understanding:
• Compute the correlation between OverallQual and SalePrice in
Delhi.
• Handle missing values in YearBuilt using imputation techniques.
3. Model Selection:
• Justify using linear regression + logistic regression vs. decision
trees.
• Discuss the role of feature engineering in improving model accu-
racy.
5 Data and Statistical Understanding
• Predict the 25th, 50th, and 75th quantiles of SalePrice in Mumbai.
• Analyze the distribution of SalePrice (e.g., normality, skewness).
• Define the hypothesis space for both regression and classification tasks.
• Explain how the choice of hypothesis space affects model performance.
• Modify hyperparameters (α, β, λ) based on quantile analysis.
3
6 Submission Guidelines
1. Report (12 pages) should include:
• Proofs of convexity and Hessian bounds.
• Feature importance plots for OverallQual and YearBuilt.
• Regularization path analysis for λ.
• Quantile predictions and data distribution analysis.
• Hypothesis space definition and its impact on performance.
• Hyperparameter tuning results and justification.
2. Code: Python implementation with adaptive hyperparameter tuning
as mentioned with stepwise instructions in the sample report.
