next up previous
Next: Evaluation Criteria and Experimental Up: Collaborative Filtering Previous: Collaborative Filtering


Datasets

We evaluated dependency networks and Bayesian networks on three datasets: (1) Nielsen, the dataset described in Section 3.3, (2) MS.COM, which records whether or not users of microsoft.com on one day in 1996 visited areas (``vroots'') of the site (available on the Irvine Data Mining Repository), and (3) MSNBC, which records whether or not visitors to MSNBC on one day in 1998 read stories among the most popular 1001 stories on the site. In each of these datasets, users correspond to cases and items possibly viewed correspond to variables. The MSNBC dataset contains 20,000 users sampled at random from the approximate 600,000 users that visited the site that day. In a separate analysis on this dataset, we found that the inclusion of additional users did not produce a substantial increase in accuracy. Table 2 provides additional information about each dataset. All datasets were partitioned into training and test sets at random. The train/test split for Nielsen was the same as for the density-estimation experiment described in Section 3.3. The learning algorithms for dependency networks and Bayesian networks and their parameters described in Section 3 were used here.

Table 2: Number of users, items, and items per user for the datasets used in evaluating the algorithms.
  Dataset
  MS.COM Nielsen MSNBC
Training cases 32711 1637 10000
Test cases 5000 1637 10000
Total items 294 203 1001
Mean items per case 3.02 8.64 2.67
in training set      


next up previous
Next: Evaluation Criteria and Experimental Up: Collaborative Filtering Previous: Collaborative Filtering
Journal of Machine Learning Research, 2000-10-19