CTR Anomaly Detection
Introduction
Click-through rate (CTR) is the ratio of users who click on a specific link to the number of total users who view a page, email, or advertisement. It is commonly used to measure the success of an online advertising campaign for a particular website as well as the effectiveness of email campaigns
The main goal of the repository is to introduce a simple approach of CTR Anomaly Detection.
Depending on the available data, there are could be several approaches for anomaly detection. If data is labeled (each point in time has a label anomaly / not anomaly, then supervised learning approaches can be used. Thus, it is a classification task where logistic regression, random forest, SVM, boosting, RNN, etc. can be applied. Here you have to pay attention to:
- Data imbalance because usually there are just a few anomalies (less than 5% of all available data) and
- Cross validation through time.
In case labels are not provided, other unsupervised techniques should be applied. There are also several methods for anomaly detection, but the general approach is quite similar: predict a value and compare it with a realized value. And if the residual is more or less than some threshold, then it is an anomaly.
There are several methods to make a prediction: moving average (simple, weighted) taking into account the last few hours or days or weekdays, etc., ARIMA model, Prophet, seasonal Hybrid Extreme Studentized Deviate technique and many others.
Requirements
- Python 3+
- pandas 1.2.1+
- numpy 1.19.2+
- seaborn 0.11.1+
- matplotlib 3.3.2+
- scipy 1.3.1+
Structure
There are two Notebooks:
- In the Exploratory Data Analysis notebook are given data description, variables preparation, and CTR calculation and visualization.
- In the Time series anomaly detection notebook, we are going to use a moving average for prediction and 1.5 standard deviations as a threshold.
Data
The underlying data is derived from a former Kaggle competition, that was about predicting “Click-Through-Rates” (CTR). CTR is defined as clicks divided by impressions and it measures how often advertisements are clicked relative to how often they are shown. The data is available in Kaggle competition
CTR over time
The data set includes a lot of dimensional fields with (categorical) information about the environment (device, location, etc) but we are only concentrating on CTR and time series relevant fields, such as
- click: 0/1 for non-click/click
- hour: format is YYMMDDHH, so 14091123 means 23:00 on Sept. 11, 2014 UTC.
Results
The simple Moving Approach works quite well for anomaly detection.
By varying the MA period in the plot_anmolies()
function, you can try to find the optimal period value. But given the nature of the data, I recommend using a period as a multiple of 24. If you would like to find less and more robust anomalies, use period 48, 72, or more.