Homework III

Due Thursday, November 15

Note the semi-open ended nature of the questions. This is a warmup to the Term Project in that sense.

Please make sure you are using Version 1.0.4 of rectools.

Problem A

This problem will involve the Czech dating agency data:

• Whose ratings are more predictable, men's or women's? Do an empirical investigation.
• Separate the data into male and female user subsets.
• Place your code in a file ProbA.R and your analysis in ProbA.tex. As usual, submit those files plus your .pdf, as well as figure files if any.

Problem B

Let N_i denote the number of ratings by User i. Typically in RS applications, covariates may not help prediction if most N_i are reasonably large; in such a situation, the user's ratings will tell us all we need to know about the person, so that the covariates become superfluous.

However, the covariates may help in predicting the cases with small N_i. In this problem you will investigate this.

• Use the InstEval data. This is a good testbed, as it has a lot of small N_i. The covariates are then the material in columsn 4 through 19.
• Use kNN, with k set to a value you found to work well in Homework I.
• Separate all users having N_i less than or equal to some threshold m. Train the model (i.e. run formUserData()) on the rest of the data, then predict this small-N_i set, using MAPE as the accuracy criterion. Do this both with and without covariates.
• As a check: I found that covariates gave a substantial improvement in the case m = 5, using k = 5.
• Place your code in a file ProbB.R and your analysis in ProbB.tex. As usual, submit those files plus your .pdf, as well as figure files if any.