Data analysis

Loading the data


						import pandas as pd
						path = "./ "
						train_ori = pd.read_csv(path + "train.csv ")
						test_ori  = pd.read_csv(path + "test.csv ")
						sample = pd.read_csv(path + "sample_submission.csv ")

Loading the data


						test.head()

EDA

Feature distribution

Common features

Feature distribution

Common features

Feature distribution

Common features

Feature distribution

Numerical features

Feature distribution

Numerical features

Feature distribution

Numerical features

Pressure over time

Inference

Since the given data is time series data, we prefer using a recurrent neural network for the purpose,
LSTM network to be more specific

Inference

Reason being, RNNs have a ability to remmeber data rather than to train and predict upon data once and for all

Classic feed forward network

as opposed to RNN

Network choice

Many to one

Feature engineering

| Feature            | Definition  |
                        | -----------        | ----------- |
                        | R                  | str(R)      |
                        | C                  | str(C)      |
                        | RC                 | R+C         |
                        | u_in_lag           | shift-1(u_in) * breath_id_lagsame   |
                        | u_in_lag2          | shift-2(u_in) * breath_id_lag2same  |
                        | u_out_lag2         | shift-2(u_out) * breath_id_lag2same |
                        | area               | cumulative_sum(time_step * u_out)   |
                        | cross              | u_in * u_out       |
                        | cross2             | time_step * u_out  |
                        | breath_id_lag      | shift-1(breath_id) |
                        | breath_id_lag2     | shift-2(breath_id) |
                        | breath_id_lagsame  | np.select([df['breath_id_lag']==df['breath_id']], [1], 0)  |
                        | breath_id_lag2same | np.select([df['breath_id_lag2']==df['breath_id']], [1], 0) |
                        | u_in_cumsum        | cumulative_sum(u_in)     |
                        | u_in_cummean       | u_in_cumsum/breath_count |

Data scaling


                        RS = RobustScaler()
                        train = RS.fit_transform(train)
                        test = RS.transform(test)

Modelling

Bidirectional ( LSTM ( 450 units) )

Bidirectional ( LSTM ( 350 units) )

Bidirectional ( LSTM ( 250 units) )

Bidirectional ( LSTM ( 150 units) )

Bidirectional ( LSTM ( 100 units) )

Input layer

Bidirectional ( LSTM ( 450 units) )

Bidirectional ( LSTM ( 350 units) )

Bidirectional ( LSTM ( 250 units) )

Bidirectional ( LSTM ( 150 units) )

Bidirectional ( LSTM ( 100 units) )

Input layer

Bidirectional ( LSTM ( 450 units) )

Bidirectional ( LSTM ( 350 units) )

Bidirectional ( LSTM ( 250 units) )

Bidirectional ( LSTM ( 150 units) )

Bidirectional ( LSTM ( 100 units) )

Dense( 50 )

Dense( 1 )

Result

Mean Absolute Error = 0.158

Runtime = 31315.8 seconds (8.7 hours)

Accelerator = TPU v3.8

Data analysis

Loading the data

Loading the data

EDA

EDA

Feature distribution

Feature distribution

Common features

Feature distribution

Common features

Feature distribution

Common features

Feature distribution

Numerical features

Feature distribution

Numerical features

Feature distribution

Numerical features

Pressure over time

Inference

Inference

Inference

Classic feed forward network

Classic feed forward network

as opposed to RNN

Network choice

Feature engineering

Data scaling

Modelling

Result

Comparing manual pressure to pressure generated by model

Thank You

Himanshu Negi

Shivanshu Semwal

Rishabh Bharadwaj

Mohd Amaan Khan

Naman Gairola