Classical ML with Classical Data: Linear Regression on Real Estate Price Prediction in Bengaluru

Code repository: https://github.com/rukshar69/bengaluru-house-prices

For tabular data, classical ML methods like Linear Regression can work quite well. In this article, I implement this model on a classical problem: predicting prices of properties in Bengaluru(India). There are 2 steps to this:

• Data preparation
• Training

We get the data on Bengaluru house prices from Kaggle. The data contains about 13k rows and 9 columns about property prices. The columns are:

area_type, availability, location, size, society, total_sqft, bath, and balcony price

The price is the target variable here.

Data Processing

The aim is to create a simplified version of the data for linear regression.

• 5 columns: 'location', 'size', 'total_sqft', 'bath', 'price' are kept
• Any row with nan values is dropped
• The size column, referring to the number of bedrooms, is processed to construct a new column bhk. The size column contains string values like 2 BHK. We take only the number value and insert it in the bhk column. The size column is then dropped.
• Values in total_sqft were found to have range values like 1133 - 1384. So, the column is modified to have only float values. For the previously mentioned range values, the average is taken and the range value is replaced with the average float value. Cases like 34.46Sq. Meter are dropped to keep things simple.
• A new feature price_per_sqft is created through dividing the price column by total_sqft

Dimensionality Reduction in the Location Column

• There are 1287 unique locations mentioned in the location column. The distribution of location values is very skewed

• Given a large number of locations don't have many datapoints, we need to apply a dimensionality reduction technique here to reduce the number of locations. locations having less than 10 rows are tagged as other locations. So, the number of categories is reduced by a lot. When using one-hot encoding, it will help having fewer dummy columns. Now, the number of unique locations is 241.

Outlier Removal

• We consider that a normal bedroom size is 300 sqft. We remove properties where per bhk size is less than 300 sqft. We now have about 12.5k rows.
• The price per sqft data reveals a significant price disparity, ranging from a minimum of 267 rupees to a maximum of around 175,000 rupees. To address this variation, we identify and remove outliers within each location using the mean and standard deviation. We keep properties for a particular location if the price per square foot is within 1 standard deviation of the mean for that location. We now have about 10.2k rows.
• Let's consider another condition. For the same location, the price of n bed apt should be greater than the mean of n-1 bed apt. The datapoints failing to meet the condition are the outliers and will be removed. So for a given location, we build a dictionary of stats(mean, std, etc.) of price per sqft per bhk
• Now we remove those n BHK apartments whose price_per_sqft is less than the mean price_per_sqft of n-1 BHK apartment. We now have about 7.3k rows.

• We can see for Rajaji Nagar and Hebbal, some of the 3 BHK properties with per sqft price less than the mean per sqft price of 2 BHK properties have been removed.
• We now consider a condition where an apartment with n bhk should have no more than n+2 bathrooms. It would be quite absurd or erroneous to have apartments where for n bhks there are more than n+2 baths. Such apartments are thus considered outliers and are removed from the dataset.

After such thorough cleaning, we move on to training a Linear Regression model using this clean and slimmed-down dataset.

Training

• There are 4 features: location, total_sqft, bath, and bhkThe price column is the target variable.
• One-hot Encoding is performed for location, a string categorical feature. For each location, we get a new binary column. Its value is 1 if the datapoint belongs to that location otherwise it's 0. However, we drop the other location column since any datapoint belonging to the other category will have 0s in all other location columns.
• The test set is 20% of all datapoints.
• The coefficient of determination/R^2 value for the trained model is about 86% on the test set, pretty good for a simplified dataset.