MLA-C01 Practice Question 84: Factorization Machines Sparse
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 9 · question 4 of 10
An ad-tech company predicts whether a user will click an ad. Each impression is one-hot encoded from user ID (8 million values), ad ID (300,000 values), and publisher site (60,000 values), producing an extremely sparse, high-dimensional feature vector. Click behavior depends heavily on *pairs* of features, such as a user's response to a kind of site, and most user-and-site pairs never appear in the training data. The team wants a SageMaker AI built-in algorithm that suits this data.
Which algorithm should the team choose?
Show answer and explanation
Answer:
Factorization Machines extends a linear model to capture interactions between features within high-dimensional sparse datasets economically. AWS names click prediction and item recommendation as typical uses. Each feature gets a small learned factor vector, and a pair's interaction is estimated from those vectors. The model can therefore score user-and-site pairs that never appeared together in training, which is exactly the problem here (option B).
- A is wrong because a linear model scores each feature on its own. It cannot represent how a user reacts to a particular *kind* of site, which is where most of the signal is. - C is wrong because trees learn interactions only from combinations they actually split on in training. They cannot generalize to user-and-site pairs never observed, and with millions of sparse one-hot columns they are costly and prone to overfitting. - D is wrong because distances in a space with millions of sparse dimensions are nearly meaningless, and searching for neighbors among every past impression does not scale to ad-serving volumes.
Reference: Factorization Machines Algorithm