Factorization Machines Quiz

Answers and explanations appear as you take the quiz.

1. How does a factorization machine represent a pairwise feature interaction?
   • By a dot product between the features' latent vectors, scaled by their values
   • By a separate unrestricted parameter for every possible feature pair
   • By sorting feature names alphabetically
   • By replacing all linear terms with a decision tree
2. Why can low-rank factors help in sparse interaction data?
   • Every pair must occur equally often
   • Related features share factor parameters across combinations
   • The method fills missing labels with known negatives
   • It removes the need for training observations
3. What does a plain linear model need to represent a user-item interaction?
   • A target encoded as a timestamp
   • Only a larger intercept
   • An explicit cross-feature or interaction term
   • A clustering silhouette score
4. How can the pairwise FM term be computed efficiently?
   • Use only the largest feature value
   • Build a dense parameter for every feature pair
   • Sort the sparse matrix by target value
   • Use a squared-sum identity instead of enumerating all feature pairs
5. What do learned latent dimensions mean by default?
   • They are causal user traits
   • They are predictive parameters without guaranteed human-readable meanings
   • They correspond one-to-one with feature names
   • They are calibrated probabilities
6. What limitation follows from a second-order FM formulation?
   • It requires dense examples for every pair
   • It cannot include linear feature effects
   • It represents pairwise interactions but not arbitrary higher-order effects directly
   • It predicts only binary outcomes
7. Why evaluate cold-start users or items separately?
   • They have known labels by definition
   • Their factors are always more accurate
   • They eliminate exposure bias
   • They may lack learned factors or sufficient interaction history
8. What does an FM score fail to prove about a learned interaction?
   • That the interaction is causal
   • That the model used feature values
   • That factors were optimized
   • That the prediction can be evaluated