4 Recommender Systems
4.1 Collaborative filtering
Given the following user-item interaction matrix:
| p1 | p2 | p3 | p4 | p5 | p6 |
u1 | ? | ? | 2 | ? | 4 | 5 |
u2 | 2 | ? | ? | 5 | ? | 4 |
u3 | ? | 4 | ? | 5 | 2 | ? |
u4 | ? | 2 | 5 | ? | 4 | ? |
1. Apply the basic user-based collaborative filtering (without considering bias) with cosine similarity. Give the top-1 recommended item to user u2. [3 marks]
2. In the lecture, we have discussed how to model the rating bias including (i) the bias over all transactions; (ii) the bias of a user; and (iii) the bias of an item in collaborative filtering. What is the rating of user u1 to item p2 if all biases are considered? [3 marks]
Note: The predicted ratings should round to one decimal place.
4.2 Evaluation of recommender system
A recommender system generates a ranked list of items for a specific user u as (p3, p10, p5, p7, p1, p9, p2, p4, p6, p8). The ranked list contains all items that haven’t been purchased by the user in the training data. We find that the user only buys items p10 and p1 in the test data.
1. Compute the AUC for user u. [1 mark] 2. If top-3 items are returned to the user, what are the values of Precision@3 and Recall@3? [2 marks]
Note: The AUC, Precision@3 and Recall@3 should round to 2 decimal places.
4.3 Application of Recommendation Algorithm
A start-up company plans to build a system for recommending training courses to users. The company has 100,000 users and 500 courses. Each user
in the database has a complete profile with age, gender, educational back- ground, and working experience. Each course has a short description of its content. Ninety percent of the users have taken one course, and the remaining users have taken more than two courses.
The data scientists in the company are considering three recommendation al- gorithms that we have learnt in the lectures: (a) user-based collaborative fil- tering; (b) item-based collaborative filtering; and (c) content-based approach.
- Which of (a) and (b) is more appropriate for the above system? Explain the reason. [2 marks]
- If you were one of the data scientists, which of the above methods will you use for recommending new courses to users? Explain the reason.[2 marks]
- The company wants a single model to support (i) recommending existing courses to an arbitrary group of users; (ii) recommending existing courses to new users; and (iii) recommending new courses to existing users. Which of the above methods can be used? If none of them apply, give a solution based on the methods we learnt in the lectures. [3 marks]
