Source Data was sourced from the UCI Machine Learning Repository Seoul-Bike-Sharing-Demand
Description: The repo contains a brief explanation on why it is important understand bike demand, particularly to ensure the stability of supply and improve mobiltiy comfort in rapidly urbanising contexts.
Objective: Overall objective of the analysis is to understand/predict bike demand based on specific variables
Data Loading: Data was loaded into the Googl Colab notebook using the'!wget' command and directly read into a pandads dataframe using 'pd.read_csv'. A copy of the data was made, so we can go back to a clean original version of the data at any point.
Variables and Data Types: Using 'data.sample(10)', a subset of the data was printed and the variables were explored. 'data.shape' revealed that the dataset contains 14 attributes and 8760 observations. Using 'data.info()', the info about the dataset was explored. 10 of the attributes are of the integer or float data type while 4 are objectives or categorical data types. 'data.isnull().sum()' revealed that there is no missing observation for any of the variables.
Independent and Dependent Variables: All other attributes asides the bike demand (Rented Bike Count) are the predictors, while Rented Bike Count (Bike Demand) is the dependent variable
Correlations: Grouped into varaibles that are positively correlated with the bike demand (i.e as these variables increase or improve, bike demand increases) and variables that are negatively correlated with bike demand (i.e as these variables increase or the situation worsens, bike demand decreases). These variables can be further categorised as being strongly correlated or not strongly correlated.
Variables Selection and Renaming: All varaibles were selected except the date variable (at least for now). And the variables were renamed to improve readability.
Categorical to Integers via Dummy Variables: The categorical variables include Seasons; Spring, Autumn, Summer, Winter, Functioning Day; Yes or No and Holiday Yes or No. These categorical variables were mapped into integers using the 'pd.get_dummies' function of pandas. To prevent the dummy variable trap (multicolinearity), the first column of each of the dummy variables were dropped.
Correlations After Variable Mapping: revealed 'Time of Day', 'Solar Radiation', 'Temperature', 'Summer' and a 'Working Day' had a strong positive correlation with increase bike demand.
Standardisation and Data Split: As the variables in the dataset had varying orders of magnitude, it was important to standardise the dataset using the sci-kit learns' StandardScaler function to make the mean 0 and the variance 1. After that, the data was split into 2 parts, the train and the test split.
Regression Models: Several regression models were used. They were fitted and evaluated based on the coefficient of determination (R2), difference between the actual bike demand (y_test) and predicted bike demand (y_pred), mean squared error and the root mean squared error.
-
Linear Regression: only 54% of the variation in the bike demand can be explained by the variation in the input features. Exploration of the actual and predicted values indicates a wide variance. MSE significantly high and RMSE in the same region.
-
Decision Tree Regressor: Actual and predicted values for bike demand are reasonably close, with 72% of the variations in bike demand explanable by variation in the input features. MSE is 0, indicating a perfect prediction model and RMSE reasonably within the range.
-
Random Forest Regressor: Actual and predicted values for bike demand also reasonably close. 86% of the variation in bike demand explanable by the variation in the input features. MSE is 0, indicating a perfect prediction of the model and RMSE reasonably close.
-
Support Vector Machine Regressor: Performed poorly than all the other models with only about 51% of the variations in bike demand explanable by the variation in the input features. Although the MSE is 0, the difference between the actual and the predicted values were significant.
Hyperparameter Tuning with GridSearch CV and Feature Importance: Based on the performance of the models,the hyperparameter tuning was done using the Random Forest Regressor and an array of parameters in a grid. The grid search revealed the max optimal features to be 8 and the number of estimators to be 30 based on the mean scores of the param categories. the feature importance also computed and ranked.
Pickling: the best estimator from the grid search consists of the best model this is pickled and packaged for deployment.
Heroku: Bike Demand App