Introductory Discussion

Strategy for Prediction of Total Scooter Rentals By Hour

to predict count, we make two models, each trained with same training set and input X, each predicting one of the dependent variables casual, and registered
this is necessary because registered rentals and casual rentals depend stroingly on disjount sets of inputs
while both estimators show there is a dependence on the cyclic time varibles and obvious weather patterns (weekends result in larger rentals of both kinds... good weather results in more ridership, both rental totals increase overall as time moves forward on the scale of years, and are periodic in day, week, and year) ...
But, there are some major differences
- registered rentals are less sensitive to ambient weather and the existence of a holiday, or other non-working day
- registered rentals stronly depend on the interaction of the workweek with thet time-of-day (rush-hour!)
- casual rentals depend more strongly variables of circumstance and display larger variance

See eda*.ipynb for interactive exploration leading to the following strategies.

Removal of Correlated inputs

we are dropping atemp because temp and atemp have $r^2=0.97$

Atemp has corrupted values and is otherwise highly correlted with temp.

Procedure

EDA
Data Preprocessing
Model interaction
Results

Findings from EDA

Independent Variable Interactions

season, month, year all interact a little
workingday and hour interact very strongly. To see why, set SHOW_ALL_PLOTS=True and run the cells below, or check the EDA notebooks.
some other interactions are built, see get_dx: function

Model Choice and Encoding

a linear model may only perform well if non-monatonic and categorical vartiables are one-hot-encoded.
for example, season itself becomes four one-hot features
temperature is a continuous variable but the boxplot analysis shows the impact is NOT exactly monatonic, much less linear.

Preprocessing

Encoding Select Interactions `weekdaytime`,`workdayhour`,`month*year`, etc

We capture the time interactions by creating several additionalvariables by multuplying (concatenating the labels of) year, month, hour, season

workingday_hr
- on working days, there is a huge impact on the registered rentals at 8am and 5pm.
- the impact is less pronouces on casual rentals but is present.
- this is a crucial interaction.
- here we encode "{workingday},{hour}" as a feature for a total of 2*24 = 48 features in the one-hot-encoded resulting X matrix.
holiday is tempting, but holiday interactions are redundant with workingday interactions
over months, years, the total (EMA) number of rentals rises!
- month -> f'{year},'{month'}'
- week-> f'{year},{week}' etc
hr_season
- the time of sunup, sundown is probably relevant to the impact of time of day on the counts
- here we encode "{hour},{season}" as a feature
yr_month
- the month alone is not as helpful because of the ever increasing popularity of the scooter
- it is relevant if it is the first or second May, for example
- here we encode "{year},{month}" as a feature
yr_season
- likewise, to separate season from the upward popularity, encode "{year},{season}"
- could also try "{month},{season}" but dont want to overfit for a first pass

Accounting for Linear + Cyclic Time Relationships

As time goes on (see below), the number of rentals increases
dependent variables are periodic in time
we define some important time intervales, $\Delta t_{i} \in { hour, week, month, year } $
The number of rentals is also cyclic in the season and other categoies which interact with cyclic time
cyclic time variables are demanded, we will require weekday, month, hour, etc and perhaps some interactions
can account for interactions by combining strings to make additional labels e.g. str(val1)+','+str(val2)
- ( can get interactions using a processing tool, but the number I want to include is smallish )
- this has sufficed

Preprocessing Procedure

Use of scaling, and encoders

geerate (by hand) interacting features
- can use polynomial feature generation, but the resulting N features is large
- can simply concatenate string producing interacting categorical variable, then encode
the process to produce X is the following:
1. one-hot encode all the categorical featues
  - can use sklearn.feature_extraction.DictVecgtorizer or pandas.get_dummies
  - benefit of the former is the ability to get the inverse transform
  - benefit of the latter is preservation of the feature names(!)
sklearn.preproessing.MinMaxScaler is used after all the encoding and partitioning is done

Derived data description

break up into dataframes for continuous/categorical X, y

Y data:
- all continuous
- columns: ['casual', 'registered', 'count']
- name: d_y
- name of scaled data: d_ycont_scaled
- scaler: x_cont_scaler
X_data
- continuous:
  - columns: ['temp', 'humidity', 'windspeed']
  - name: d_xcont
  - name of scaled data: d_xcont_scaled
  - scaler: y_cont_scaler
- categorical:
  - columns: ['holiday', 'workingday', 'weather', 'hr_season', 'yr_month', 'yr_season']
  - name: d_xcat

Results:

0       ridge   32.462056      casual      20.291495
1     svm lin   38.933862      casual      39.226996
2         svm   43.894072      casual      59.804135
3  linear_reg   32.416716      casual      19.918105
4  elasticnet   34.523040      casual      23.875527
5       ridge   84.813232  registered      15.907984
6     svm lin  118.068750  registered      42.838540
7         svm  141.407167  registered      74.503286
8  linear_reg   85.012282  registered      15.453266
9  elasticnet  100.391269  registered      25.298674

Coinfilp Problem Statement

The attachment contains:

Data dictionary
Training dataset for model building
Testing dataset for submission

Problem statement:

Given is the hourly rental data of an e-scooter rental company. The goal is to predict the total number of e-scooters rented per hour. We recommend using Python for this assignment and are free to use any packages and IDE. Submissions are evaluated based on how the candidate is able to explore the data using charts, correlation matrix, etc.; and prediction using various modeling methods, the Root Mean Squared Logarithmic Error (RMSLE) as potential evaluation metrics.

Some guidance:

Pls submit your analysis/code in Python Notebook, you can show the following results:

data pre-processing, feature generation
data exploration analysis and visualization
modeling - algorithms choices and optimization
final model performance analysis, feature importance
Based on the parameters provided in the testing data, you can use the trained model above (or retrain the model with different set of variables) to predict counts, and add the "count" column in the testing data and submit it.

Data Description

datetime - hourly date + timestamp

season - 1 = spring, 2 = summer, 3 = fall, 4 = winter

holiday - whether the day is considered a holiday

workingday - whether the day is neither a weekend nor holiday

weather - 1: Clear, Few clouds, Partly cloudy, Partly cloudy 2: Mist + Cloudy, Mist + Broken clouds, Mist + Few clouds, Mist 3: Light Snow, Light Rain + Thunderstorm + Scattered clouds, Light Rain + Scattered clouds 4: Heavy Rain + Ice Pallets + Thunderstorm + Mist, Snow + Fog

temp - temperature in Celsius

atemp - "feels like" temperature in Celsius

humidity - relative humidity

windspeed - wind speed

casual - number of non-registered user rentals initiated

registered - number of registered user rentals initiated

count - number of total rentals

Altoidnerd/coinflip