Managing Corporate Portal Usage with Recommender Systems

More detailed descriptive statistics of this corporate portal are provided in Sect. 4.

http://​www.​last.​fm.

http://​www.​netflix.​com.

http://​www.​facebook.​com.

For instance, different departments often have different wordings for comparable issues. This would need to be reflected in service descriptions and user preferences. Moreover, complex and specialized services often cannot be easily textually described.

Each prediction p _j for an item j is compared to the original rating r _j with \(e_{\mathrm{MAE}} = \frac {1}{M} \sum_{j=1}^{M} |p_{j} - r_{j}|\), and \(e_{\mathrm{MSE}} = \sqrt{\frac {1}{M} \sum_{j=1}^{M} (p_{j} - r_{j})^{2}}\), respectively, where M is the number of predictions.

The statistic is based on a total of 466 new users in the period from December 2009 to July 2010. Before this time period, a new user could not be unambiguously distinguished from a repeated user. New users that were later assigned to the treatment group are not considered.

The ANOVA is equivalent to the well-known t-test if there is only one grouping variable that can take exactly two different values. Since we consider a full factorial design with three grouping variables (rating rule, user similarity, prediction algorithm), each of which can take two or more different values, the ANOVA is the more appropriate test. Similar to the t-test, the null hypothesis is that the mean prediction quality is the same across the different levels of the grouping variable. Under the null hypothesis, the ANOVA test statistic has an F-distribution. The ANOVA derives an F-value for each grouping variable, called the main effects, as well as an F-value for all possible combinations of the grouping variables, called the interaction effects. Given the degrees of freedom of the grouping variable (i.e., the number of values of the grouping variable minus one) and the residual degrees of freedom of the model (i.e., number of observations minus the sum of the degrees of freedom of the grouping variables minus one: here 89) , the p-value (significance level) can be computed.

To obtain the Bonferroni adjusted p-value, the uncorrected p-value is multiplied by the total number of comparisons. If the answer exceeds 1.0, the corrected p-value is reported as 1.0.