Prediction of Bicycle Rentals in Washington, D.C.

The current study deals with the exploration and preparation of a dataset concerning bicycle rentals in Washington D.C, in order to predict the total number of rental bicycles per hour. The dataset is structured as follows:

• instant: record index
• dteday: date
• hr: hour (0 to 23)
• weathersit: Weather situation
• temp: Normalized temperature in Celsius. The values are divided to 41 (max)
• atemp: Normalized feeling temperature in Celsius. The values are divided to 50 (max)
• hum: Normalized humidity. The values are divided to 100 (max)
• windspeed: Normalized wind speed. The values are divided to 67 (max)
• casual: count of casual users
• registered: count of registered users
• cnt: count of total rental bikes including both casual and registered. This is the target to predict.

This study has effectively prepared a dataset for its prediction with a model, which in turn has been selected after a thorough search within linear and non-linear algorithms, simultaneously evaluating their performance with different parameter choices. With respect to the exploratory data analysis, the following can be extracted:

• 'weathersit', 'temp', 'atemp', and 'hum' all present NaN values, at different rows
• 'weathersit' is the only categorical feature, with 4 possible categories
• 'temp' and 'atemp' have a fairly centered distribution, whereas 'cnt' and 'windspeed' are right-skewed and 'hum' is left-skewed.
• 'windspeed' and 'cnt' have outliers for large values of the variable, whereas 'hum' has outliers at 0.
• 'atemp' and 'temp' are very strongly correlated with each other, at the same time being correlated significantly with 'cnt'. 'hum' is negatively correlated with 'cnt', and 'hr' follows a non-monotonic relation with 'cnt', with a peak in bicycle rentals around mid-afternoon.

With regards to the data engineering of the dataset, the following steps have been followed:

• From 'dteday', columns indicating the season, whether the day was in the weekend, and whether the day was a holiday, have been generated.
• For 'weathersit', four intuitive categories have been defined, imputing NaN values based on the most common weather condition of each month.
• For 'hr', the numerical value has been re-scaled to the range from 0 to 1, the hours with daylight have been labeled, as well as rush hours.
• For 'hum', outliers at 0 and NaN values have been imputed with the average of each month, and the variable has been normalized to reduce its skweness.
• For 'windspeed', outliers for large values and NaN values have been imputed with the average of each month, and the variable has been normalized to reduce its skweness.
• For 'atemp' and 'temp', both variables have been combined into 'wtemp', imputing outliers within one variable with the corresponding value of the other.
• For the target variable 'cnt', a normalization has seen fit to reduce its skewness, although this transformation is not included in the final preprocessing pipeline.

In order to select and train an adequate model:

• A simple linear regression, a k-nearest neighbors regression, a random forest regression, and a gradient boosting regression have been attempted with default hyperparameters. The model with the best R2 in this preliminary examination was the random forest regressor.
• Based on the three best-performing models, an extensive grid search has been performed, once again confirming the best performance of the random forest regressor.
• As a result, the most relevant hyperparameters of the RandomForestRegressor have been further examined, revealing a flawed selection of hyperparameters in the previous grid search. The optimum model corresponded to n_estimators = 50, max_depth = 10, and min_samples_split = 10.
• Predicting with this optimum model resulted in R2 = 0.856, with a satisfactory prediction plot.

To further improve this model, the following alternatives can be explored:

• Could all numerical variables be completely normalized, such that their distributions are centered at 0? This would imply normalizing 'wtemp', and not re-scaling any of the other numerical variables ('hum', 'windspeed', etc.) after their normalization.
• Could other more complex models be used to predict 'cnt' with greater success? Would a neural network return a better prediction?
• Could one split the dataset into multiple sub-datasets, training a different model to each split? For example, one could have a model for holidays and another for work-days.
• Could PCA be used to fit a model based on a smaller number of transformed, combined features, resulting in a better model?