Handle cold start, dropout, and training stability

Machine Learning deep dive

Answer the following conceptual questions (you may use equations and small examples).

A) Recommender systems: cold start

1. How do you handle cold start (new users and/or new items) in a recommendation system?
2. If a user is brand new with no historical behavior , what should your model output?
3. How do you handle bias introduced by missing history or exposure/selection effects?

B) Dropout

1. Explain the difference between training and inference behavior for dropout.
2. Why is there a scaling factor (e.g., dividing by 1-p or multiplying by 1/(1-p) ) during training in inverted dropout?
3. What happens if you enable dropout at test time ?

C) Optimization choices

1. What learning-rate scheduler did you use (e.g., step, cosine, plateau-based)? Why was it appropriate?
2. Explain gradient clipping and when it becomes necessary.

D) Learning theory in practice

1. Explain the bias–variance trade-off .
2. Explain double descent : why can test error decrease again after model capacity/training passes a certain threshold?