mikel-brostrom / Housing_Price_Prediction

California housing price prediction with NN, Random Forest and Linear Regression

Housing_Price_Prediction

The idea of this project was to create a predictor on the california housing dataset. The data pertains to the houses found in a given California district and some summary stats about them based on the 1990 census data.

The data

The data is comprised of 8 attributes

• MedInc median income in block

• HouseAge median house age in block

• AveRooms average number of rooms

• AveBedrms average number of bedrooms

• Population block population

• AveOccup average house occupancy

• Latitude house block latitude

• Longitude house block longitude

as well as the target, the housing price

Training

train.py runs the training for three different models: NN, linear regression and random forest on the scikit-learn california housing dataset:

python3 train.py

Training output example:

...

Train Epoch: 150 [0/18576 (0%)] Loss: 0.130984
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 37/37 [00:00<00:00, 143.57it/s]
5
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 102.93it/s]
        NN
r2 score: 0.8232532827300671
MAE score: 0.2833885734455647
        Linear regressor
r2 score: 0.6098033978087847
MAE score: 0.4635741867691994
        Random forest
r2 score: 0.8138137169848451
MAE score: 0.2837869675577879
...

The network

We use a stack fully connected layers with ReLU. The r2 score and MAE was used for evaluating the models

Conclusion

The neural network trained on the standardized signals gave the best model with an R2 score of 82.4. The models trained on the first two principal componentes gave a poor result even if they accounted for ~96% of the data variance.