Q24: (6 points) Assuming we aim to build a more advanced recommendation system for an online bookstore using matrix factorization-based methods, similar to the one that won the Netflix prize. Suppose the global mean rating of books is 3.6 stars. Bob, a loyal customer, has rated 400 books, and his average rating is 0.4 stars higher than the global average rating. Meanwhile, Pride and Prejudice is a book in the bookstore that has 200,000 ratings, with an average rating that is 0.5 stars lower than the global average. What would be a baseline estimate of Bob’s rating for Pride and Prejudice? (each 2 points) Illustrate how you arrived at your answer. (2 points)

Q11: (3 points) Answer True or False only: Reinforcement learning is to model sequential decision data. One of the most famous methods of policy gradient is DQN.

Q23 (6points) In an online shopping application scenario, you want to design a recommendation system for complementary or bundle item recommendations (e.g., laptop + laptop bag, pasta + pasta sauce). Which of the following would be a better recommendation system? Choose one and explain the reason(2 points). a) User-user collaborative filteringb) Item-item collaborative filteringc) Content-based recommendation  In one sentence, justify your answer (4 points).

Q22(20 points): Let us consider a one-dimensional space. We wish to perform a hierarchical clustering of the points 2, 6, 10, 14, and 18. Show what happens at each step until there are two clusters and give these two clusters. Your answer should be a table with a row for each step; the row should contain: The members of the new cluster formed  (5 points). Its centroid  (5 points). If you are merging a cluster C1 = {x, y, z} with centroid c1, and a cluster C2 = {p, q} with centroid c2, use the centroid–centroid distance |c2 – c1| to measure the distance between C1 and C2. For the merged cluster, you should report both: The members of the new cluster (5 points). The centroid obtained after merging (5 points).  

Q23 (6points) In an online shopping application scenario, you want to design a recommendation system for complementary or bundle item recommendations (e.g., laptop + laptop bag, pasta + pasta sauce). Which of the following would be a better recommendation system? Choose one and explain the reason(2 points). a) User-user collaborative filteringb) Item-item collaborative filteringc) Content-based recommendation  In one sentence, justify your answer (4 points).

Q3: (3 points) Answer True or False only: In AdaBoosting, subsequent decision stumps are trained by reweighing misclassified samples of the previous decision stumps. In this way, weak learners are combined parallelly to create a strong learner.

Q13: (3 points)In AdaBoosting, there are 8 data instances (i.e., samples). Before updating the weights of the eight samples, the weights of the eight samples are. After updating the weights of the eight samples, the new weights of the eight samples are. Which data instances (samples) is/are correctly classified by the decision stump of this iteration?

Q3: (3 points) Answer True or False only: In AdaBoosting, subsequent decision stumps are trained by reweighing misclassified samples of the previous decision stumps. In this way, weak learners are combined parallelly to create a strong learner.

Q28: (6 points)What are the three loss function names of RankSVM, RankBoost, and RankNet discussed in the lectures?

Q9: (3 points) Answer True or False only: In RL, the agent learns from labeled data, which we call trajectories, provided during training.