travel-time-prediction

        Please do refer the `main.ipynb` file for detailed description

Estimation of ETA and ETD from the Train Travel Time Data

1. Importing Libraries

Importing all the necessary libraries for making this project, listed in requirements.txt

2. Load Data

Loading Data using Pandas

3. Correcting Inconsistencies, Data Wrangling, Feature Engineering and EDA

Finding Unique Values in every Feature
Dropping Redundant Columns
Imputing NaN
Dropping some feautures
Breaking datetime to columns namely year, month, date, hours, minutes, secs

 Unique Values in stations =  (349,)  
 Unique Values in trainCode =  (15,)
 Unique Values in trainStationId =  (349,)  
 Unique Values in sA =  (24480,)  
 Unique Values in sD =  (23505,)  
 Unique Values in aA =  (25594,)  
 Unique Values in aD =  (25496,)  
 Unique Values in distance =  (573,)  
 Unique Values in dayCount =  (3,)  
 Unique Values in ArrivalDelay =  (501,)  
 Unique Values in DepartureDelay =  (455,)

After finding the unique values in all the features we get to know that the data is from the time span runDate 2020-01-01 to 2020-02-19 among 337 stations with 15 trains trainCode . Train Stations having their 337 trainStationID with Arrival and Departure frequency of 25051 in this time span. Trains take around dayCount from 0 to 2

Since, stations are already coded with their trainStationID we can remove one of them

3.1 After Imputing the data

Now we know that ArrivalDelay, DepartureDelay were derived columns but we can use those columns to derive the missing values in scheduledArrival and scheduledDeparture.

For the missing values in test sA and sD; train sD. We will drop those values in both test and train.

3.2 Dropping some features

Since Arrival Delay is a derived column we can remove the actualArrival and actualDeparture from the train data

3.3 Breaking down the datetime values

Dropping the year 2020 from all the datetime formats
Dropping the second from datetime values.
Breaking all the date time values in test and train data.

Not taking year and seconds because of having the same value all along

4. Model Selection and Model Building

DecisionTreeRegressor
RandomForestRegressor

For 2 different models, one for DepartureDelay and other for ArrivalDelay In totality 4, 2 for DepatureDelay(DD), and 2 for ArrivalDelay(AD)

4.1 Feature Importance

For example this what I received on

Feature Importance:

trainCode -- 0.25
trainStationId -- 0.25
distance -- 0.14
dayCount -- 0.13
runDate_month -- 0.08
runDate_date -- 0.04
sA_month -- 0.03
sA_day -- 0.02
sA_minute -- 0.02
sA_hour -- 0.01
sD_month -- 0.01
sD_day -- 0.01
sD_hour -- 0.01
sD_minute -- 0.01

5. Evaluation Criteria

Checked whether feature Importance is bogus by removing the trainCode feature and got to know that regressor score becomes 0.1113 but when the trainCode is supplied the score becomes more than 93. But since DepartureDelay cannot be passed the score remains out to be 0.77763.

The score given by the regressors were used for the Evaluation of the model

anoushkrit/travel-time-prediction