Predict Damage to a building - ML Challenge on Hackerearth

Determining the degree of damage that is done to buildings post an earthquake can help identify safe and unsafe buildings, thus avoiding death and injuries resulting from aftershocks. Leveraging the power of machine learning is one viable option that can potentially prevent massive loss of lives while simultaneously making rescue efforts easy and efficient.

In this challenge we are provided with the before and after details of nearly one million buildings after an earthquake. The damage to a building is categorized in five grades. Each grade depicts the extent of damage done to a building post an earthquake.

Goal of the Project:

our task is to build a model that can predict the extent of damage that has been done to a building after an earthquake.

Author:

• Brungi Vishwa Sourab

Achievements about this project:

• Rank 92 out of 7540 on public leaderboard.

Data Variables Description:

Variable	Description
area_assesed	Indicates the nature of the damage assessment in terms of the areas of the building that were assessed
building_id	A unique ID that identifies every individual building
damage_grade	Damage grade assigned to the building after assessment (Target Variable)
district_id	District where the building is located
has_geotechnical_risk	Indicates if building has geotechnical risks
has_geotechnical_risk_fault_crack	Indicates if building has geotechnical risks related to fault cracking
has_geotechnical_risk_flood	Indicates if building has geotechnical risks related to flood
has_geotechnical_risk_land_settlement	Indicates if building has geotechnical risks related to land settlement
has_geotechnical_risk_landslide	Indicates if building has geotechnical risks related to landslide
has_geotechnical_risk_liquefaction	Indicates if building has geotechnical risks related to liquefaction
has_geotechnical_risk_other	Indicates if building has any other geotechnical risks
has_geotechnical_risk_rock_fall	Indicates if building has geotechnical risks related to rock fall
has_repair_started	Indicates if the repair work had started
vdcmun_id	Municipality where the building is located

Libraries Used:

numpy
pandas
matplotlib
sklearn
keras

Preprocessing Steps done:

1. Merging Datasets:

In the competition, we were given with four datasets: train.csv, test.csv, Building_Ownership_Use.csv, Building_Structure.csv. Our first task was to merge the datasets. First of all, I had merged Building_Ownership_Use and Building_Structure datasets on building_id column (lets call this as buildings dataset). Then, merged train dataset with buildings dataset on building_id using left join. After Merging train with buildings dataset, I did the same with the test dataset.

Merging the datasets is done using pd.merge method.

2. Dealing with Null Values:

Machine Learning models can not accept null values while training the model. So, we have to either remove the null values or fill them. We can use several methods to fill null values. But in numerical datasets, the mean & mode are most favourite methods.

Null values are filled using fillna() method in python

3. One-Hot Encode the data:

Machine Learning models performs fast and accurate if the data is one-hot encoded.(No official statement but through personal experience). So, I have encoded both train and test datasets. Before encoding, there are 56 columns in train and test (I have already separated target from train) and after encoding, the number of columns in test and train are 97.

One Hot Encoding is done using pd.get_dummies() function. After encoding, we need to align the columns in train and test datasets. This is done using align() method in python.

4. Scaling the data:

We have converted the data to numerical. But, the numerical ranges of each columns are different. This tends our model to overfit easily. We scaled all the columns in our data to the range 0-1 (Both Inclusive).

Scaling is done using MinMaxScaler() method in sklearn.preprocessing library.

Models:

Neural Network:

Built a Sqeuential model with 1 input, 4 hidden & output layers in Keras. relu is used as activation function in input and hidden layers and a softmax with 5 classes for the output layer. The code used to build the architecture mentioned is

keras_model= Sequential()

keras_model.add(Dense(256,input_shape=X_train.shape[1:],activation='relu'))
keras_model.add(Dense(128,activation='relu'))
keras_model.add(Dense(64,activation='relu'))
keras_model.add(Dense(32,activation='relu'))
keras_model.add(Dense(16,activation='relu'))
keras_model.add(Dense(5,activation='softmax'))

We can see the summary of our model using summary() method in keras. Our model Summary was:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_1 (Dense)              (None, 256)               25088     
_________________________________________________________________
dense_2 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_3 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_4 (Dense)              (None, 32)                2080      
_________________________________________________________________
dense_5 (Dense)              (None, 16)                528       
_________________________________________________________________
dense_6 (Dense)              (None, 5)                 85        
=================================================================
Total params: 68,933
Trainable params: 68,933
Non-trainable params: 0
_________________________________________________________________

The above model was trained for 10 epochs and this model gave an accuracy of 71.88%.

Random Forest Model:

The model that gave me the top score is a Random Forest Model. The hyperparameters used are:

n_estimators=400
min_samples_split=3
random_state=120

I trained this model with minimal tuning of hyperparameters because of the computational time and costs.

After training the model, I have checked the importances of features. The top 10 features with importance percentages are:

Variable	Importance Percentage
height_ft_post_eq	12.04
count_floors_post_eq	11.13
condition_post_eq_Not damaged	5.49
age_building	5.07
plinth_area_sq_ft	4.98
ward_id_x	4.72
ward_id_y	4.6
area_assesed_Both	4.4
condition_post_eq_Damaged Not used	3.7
area_assesed_Building removed	3.56

Thank You for your time

with ❤️, Brungi Vishwa Sourab