Web Traffic Forecasting

This Repository is to analyse the web traffic of a website for a particular time interval using Time Series Analysis and forecast the same for the future. The website chosen for analysis here is Wikipedia. The scripts are in iPython Notebook (ipynb) format.

This Analysis adapts the flow of the Data Science Life Cycle.

Initial Setup

git clone https://github.com/San411/Time-Series-Analysis.git
pip3 install -r requirements.txt

Data Mining

We take the wikipedia page views for 145k articles for the year 2015-2017 on a daily basis. This will form our dataset for training. First download the dataset provided by Google from Kaggle. This should not take long. Save the data as 'data_2015-17.csv' in the 'Datasets' folder. To get the validation data use the wikipedia pageview API to request for further article pageviews for the year 2018-19 on a Daily basis. Refer Pageview API documentation. The REST API returns the JSON data which is further stored in a file to access later. Create a folder named 'json_files' in Datasets Now execute the data_mining.ipynb. The files will take long to download. The JSON data will look like this.

{"project": "en.wikipedia",
  "article": "!vote", 
  "granularity": "daily",
  "timestamp": "2018010100",
  "access": "all-access", 
  "agent": "all-agents",
  "views": 9}

Data Cleaning and Exploration

Execute the 2 scripts in the following order:

data_cleaning.ipynb
data_exploration.ipynb 

The cleaned data will be saved in the datasets folder.

View Plot for a sample dataset after exploration

Feature Engineering

Results of Rolling Statistics

After performing the Augmented Dickey Fuller Test the flattened data is stored in the datasets folder.

ARIMA

Mean Absolute Error : 28.318

Mean Squared Error : 2110.667

Root Mean Squared Error : 45.941

FACEBOOK - PROPHET

Mean Absolute Error : 220.883

Mean Squared Error : 75844.554

Root Mean Squared Error : 275.398

SMAPE Score : 42.966