Linear Regression with Prediction Features in Mobile Application

Problem Statement

New York City bike is a sharing services by Citi Bike. The company's goal is to crate a digital experience for a new pay-as-you-go rental options. Therefore, in the development of a new mobile application, they plan to add feature that enable customers to know the predicted average trip time and cost estimation for the ride if they provide their departure and arrival station.

Project Objectives

The goal of this project is to develop a predictive model that enable customers to know the predicted average trip time and cost estimation for the ride. By accurately predicting trip outcomes of a bike rental, customers can make informed decisions, thereby increasing customer's benefits and enhancing overall digital experience.

Dataset

The ML datasets have been collected and stored in a BigQuery public dataset bigquery-public-data under the name new_york_citibike dataset.

The public dataset has 2 BigQuery tables. The one used is citybike_trips table.

• The tripduration is our label -- indicates the duration expressed in seconds if each bike rental.
• Other fields are potential features -- contains more than 58 million records.

Methodology

The project follows a standard machine learning pipeline, which includes the following steps:

• Data Preprocessing: Cleaning the dataset, handling missing values, tranform datetime variables, and feature scaling.

• Model Selection: Using Power Query on BigQuery to split dataset into 3 sets to perform model selection and model evaluation. Months were used as a targeted measurement to split dataset. After that, trying out various machine learning algorithms and selecting the best-performing model based on evaluation metrics.

• Model Evaluation: Assessing the model's performance using appropriate evaluation metrics and fine-tuning as necessary.

Training Linear Regression model - Develop an effective asset with good performance

Model Optimization: Improve the performance of our ML model multiple times with Loss function

Mean Squared Error -- avgerage squared distance between actual values and predicted values of the label

1.  trip_duration_by_stations - MSE = 111.2176

2. trip_duration_by_stations_and_day - MSE = 98.0522

3. trip_duration_by_stations_day_age - MSE = 110.8004

Conclusion

The result of the SELECT COUNT query returns a value of :

1,548,371 Predictions, with a difference between the predicted, the actual value being Less Than 15 Minutes.

Considering that the total size of the prediction_table table is 1,728,078, we can say that:

 In 89.6% Of Cases, our machine learning model is able to predict:

- The Trip Duration with a gap that's Less Than 15 Minutes.
- The Exact Fare of The Ride with Mean Absolute Error of 6,8 mins.

Actionable Recommendations

From the conclusion, here are some recommendations we provide for the new on-demand and pay-as-you-go pricing model strategy:

Recommendations	Description For The Model In Mobile Application
Pricing Model Strategy	Pricing start with quarterly fare for pricing model with recurring payments
Prediction Trip User Stories	when a customer picks up a bike from a start station and specifies its destination on the mobile application on a specific day of the week.
Prediction Trip Duration Confident	Our model will be able to predict the trip duration with a with a gap less than 15 minutes.
Model Performance Confident	Our model will be able to predict the exact fare of the ride with a mean absolute error of about 6.8 minutes
Price Prediction Confident	We confident that 89.6% the application will provide a good estimation of the price to our customers

Author

• Nguyen Dang Xuan Linh - GitHub Profile