Why the First Spam-Filtering Strategy Is Hard to Pick
Question: In designing a new email spam-filtering system for a university inbox, why is it difficult to choose the best first development path, even for an experienced machine learning engineer? How does knowledge of the email domain affect that choice?
Sample answer: Choosing an initial path for a spam filter is difficult because there are many reasonable ways to proceed, such as changing features, trying different models, or adjusting thresholds. Even a skilled practitioner can have trouble deciding which option is most promising at the start. If the practitioner does not understand the email environment well, the choice becomes even harder because it is less clear which errors matter most and which clues are useful.
Key points:
- A new system usually has several plausible development paths.
- Even experienced machine learning practitioners may find it hard to select the best first step.
- Limited knowledge of the application domain makes the decision substantially more difficult.
Rubric: The response should state that there are multiple possible development directions, explain that even experienced practitioners can struggle to choose the best initial path, and note that weaker domain knowledge makes the decision harder.
0
1
Tags
Machine Learning
Deep Learning
Machine Learning Strategy
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Yearning @ DeepLearning.AI
Related
Building a Spam Filter with a Large Labeled Message Set
Text Content Analysis as a Spam-Filtering Direction
Using Mail Routing Metadata for Spam Detection
What lesson does the fraud-detection example highlight about starting a new ML project?
Choosing an Initial Direction for a Spam Filter
Andrew Ng says it is even _____ to pick a first direction for a new ML project when you do not know the domain well.
Match each way to improve an email-filtering system with the kind of information it mainly uses.
Order the steps for choosing and improving a first version of a fraud-detection system.
Why is a team lead’s admission that choosing the first model for a document classifier is hard educationally useful?
Is the best starting point for a new ML project always obvious to experienced practitioners?
Choosing among early project directions
Match each statement about making an initial choice in a new ML project to its practical meaning.
Why a Fast Baseline Can Be Better Than Waiting for the Perfect Plan
Why the First Spam-Filtering Strategy Is Hard to Pick
Judging the Best First Path for an Email Spam Filter
Domain Familiarity and Choosing a First Modeling Direction