Your proven path to success in Machine Learning Interviews – developed by FAANG engineers. Unlock ML loops at top companies with a System Design approach.

System design is an important component of any ML interview. Being able to efficiently solve open-ended machine learning problems is a key skill that can set you apart from other engineers and increase the level of seniority at which you’re hired.

This course helps you build that skill, and goes over some of the most popularly asked interview problems at big tech companies. You’ll walk step-by-step through solving these problems, focusing in particular on how to design machine learning systems rather than just answering trivia-style questions. 

Once you’re done with the course, you’ll be able to not just ace the machine learning interview at any tech company, but impress them with your ability to think about systems at a high level. If you have a machine learning or system design interview coming up, you’ll find the course tremendously valuable.

Grokking the Machine Learning Interview

The ad prediction component has to make predictions for the final set of candidate selected ads. It needs to be robust and adaptive and should be able to learn from massive data volume.
 
Let's go over the best setup and models for this problem.

# Modeling approach
Ads are generally short-lived. So, our predictive model is going to be deployed in a dynamic environment where the ad set is continuously changing over time. 

Given this change in an ad set, keeping the model up to date on the latest ads is important. In other words, model performance will degrade with each passing day if it isn't refreshed frequently.  




### Online learning
If we have to plot the log loss of the model, it might look like the graph on the right. Here we are assuming that the model is trained on day one and we are observing log loss on six consecutive days. We can observe the degradation of prediction accuracy (increase of log loss) because of the increased delay between the model training and test set. 

So, keeping models updated with the latest data is important for ad prediction.  


One approach to minimize this loss could be refreshing the model more frequently e.g. training it every day on new data accumulated since the previous day.

However, a more practical and efficient approach would be to keep refreshing the model with the latest impressions and engagements after regular internals (e.g. 15 mins, 30 mins, etc.). This is generally referred to as **online learning** or **active learning**. 

In this approach, we train a base model and keep adding new examples on top of it with every ad impression and engagement. From an infrastructure perspective, we now need a mechanism that collects the recent ad data and merges it with the current model.

The following figure shows some key components that are critical for enabling online learning. We need a mechanism that generates the latest training examples via an online joiner. Our training data generator will take these examples and generate the right feature set for them. The model trainer will then receive these new examples to refresh the model using stochastic gradient descent. This forms a tightly closed loop where changes in the feature distribution or model output can be detected, learned on, and improved in short successions. Note that the refresh of the model doesn't have to be instantaneous, and we can do it in same batches at a certain frequency, e.g., every 30 mins, 60 mins etc.



Let's have a look at how the most relevant ads will be predicted from a set of selected ads.

Introduction

Practical ML Techniques/Concepts

Search Ranking

Feed Based System

Recommendation System

Self-Driving Car: Image Segmentation

Entity Linking System

Ad Prediction System

Ad Prediction

Modeling approach

Online learning