with m records and the other with n records (assume m < n). Under a basic approach, m×n record comparisons will be needed.
Assume there are no duplicates. What is the maximum number of record matches? What is the corresponding number of non-matching comparisons required in this circumstance?
What is the smallest possible number of comparisons?
What is the smallest possible number of comparisons that will return all matches? What is the value of b?
What is the largest possible number of comparisons?
If records are evenly allocated to b blocks, how many comparisons will be needed? What is the advantage with a large b? What is the advantage with a small b?
What are the reasons two records could be linked incorrectly?
Suppose a false positive (FP) is two records that are linked by the system, which a human believes should not have been linked. Suppose a false negative (FN) is when two records are not linked by the system, which a human believes should have been linked. A true positive (TP) is two records linked by the system which a human believes should have been linked and a true negative (TN) is two records not linked by the system which a human believes should not have been linked.
– What are the relative sizes of the categories TP, TN, FP, FN? In practice, how might one calculate these sizes?
1
3. Consider the following dataset:
User Iron Man Superman Batman Spiderman Ant-Man Wonder Woman
Anne 3 3 3.5 2.5 3 Bob 4 3.5 2.5 4 3 3
Chris 3 3 3 4
Dave 3.5 2 4 2.5
Eve 3 3 2 5
Frank 2.5 4 5 3.5 5 Gary 4.5 3 4 2
Use the Item-based recommender systems approach discussed in lectures to predict Frank’s rating for Batman.
Use the User-based recommender systems approach to predict Frank’s rating for
Batman
Identify the advantages and disadvantages of each approach
4. Recommender systems are sometimes criticised for over-recommending popular items to users and under-recommending rarer items. Why do you think this happens? How might it be addressed?
2
