educative.io

Cascade of classifiers that is commonly used in AdTech

Courses
,
#1

Hi all,

There is a disclaimer at the beginning of the section “Ad click prediction” of the Machine Learning Design course that says

For the sake of simplicity, we will not focus on the cascade of classifiers that is commonly used in AdTech.

I’d like to know more about the cascade of classifiers used in AdTech. I would appreciate if someone could indicate me some references.

Course: https://www.educative.io/courses/machine-learning-system-design
Lesson: https://www.educative.io/courses/machine-learning-system-design/N8K35PGw4nm

#2

Hi @Pablo1 !!
The cascade of classifiers is a technique commonly used in AdTech (Advertising Technology) for tasks such as ad click prediction, ad fraud detection, and ad targeting. It involves the use of multiple classifiers organized in a sequential manner to make predictions.

The primary goal of the cascade of classifiers is to improve the overall performance of the prediction model by effectively filtering out negative instances early in the classification process, thus reducing the computational overhead required for subsequent stages. This approach is particularly useful in scenarios where the majority of instances can be quickly classified as negative, allowing the classifier to focus more computational resources on the remaining instances that are likely to be positive.

Here’s a high-level overview of how the cascade of classifiers typically works:

  1. Stage 1: The first classifier in the cascade is designed to be highly efficient and fast, but with relatively high error rates. It performs a quick preliminary classification on the input instances and separates them into two categories: “likely negative” and “potential positive.” The goal of this stage is to filter out as many negative instances as possible.

  2. Stage 2: The instances classified as “potential positive” in the first stage are passed to the second classifier, which is typically more complex and accurate but slower than the first classifier. It further refines the classification by identifying true positives and discarding false positives. The output of this stage consists of instances classified as “positive” and “negative.”

  3. Stage 3: If necessary, subsequent stages can be added to the cascade, each with increasing complexity and accuracy. The number of stages and the specific classifiers used depend on the application and performance requirements.

The cascade of classifiers approach allows for a trade-off between computational efficiency and accuracy. By using a series of classifiers with increasing complexity, the overall performance can be improved while still maintaining reasonable computational efficiency. The classifiers in each stage can be trained independently or in a collaborative manner, depending on the specific design choices and data availability.
I hope it helps. Happy Learning :blush: