Effect of stimulation on number of overlaps Effect of suppression on number of reviewers looked at Figure 2: Effect of stimulation rate on prediction and recommendation Figure 3: Effect of suppression rate on neighbourhood size and reviewers It can be seen that the simple AIS gives broadly similar We then tested the effect of suppression on the Als prediction performance to the Simple Pearson. The MAE performance. Here we fixed the baseline rate at stimulation measurements from different runs are not normally only(no suppression), and took measurements relative to this distributed,so a non-parametric statistic is appropriate. We baseline. Again, it should be noted that the first graph shows performed a Wilcoxon analysis, which showed that the prediction error(hence, a good result is low) fference between prediction errors of SP and Als is zero with 95% confidence. In addition, the choice of an appropriate stimulation rate did make a significant difference (a rate of 0.2 compared with 0.02 at the 95% level) For recommendation, the AlS performs better than the SP positive95% Wilcoxon analysis to assess significance. We itmomft t at stimulation rates above 0. 1. Again, we performed a excluded cases where a recommendation score was unavailable(due to an insufficient number of overlaps ) The sow number of recommendations and overlaps show similar trends though the als gives a more constant value. Again, some stimulation was beneficial In later experiments, the stimulation rate was fixed at one of the better values(0. 2, 0.3 or 0.5), in order to give us a Effect of suppression on recommendation accuracy good base to work on. These values give us generally good performance, while keeping a good neighbourhood size and still evaluating a reasonable number of reviewers 110.0% Experiments on the Idiotypic AlS Having fixed all the simple parameters, we tested of suppression for stimulation rates of 0. 2, 0.3 and surprisingly we found that suppression changed the d os. not of reviewers looked at and the num ber of neighbours Effect of suppression on neighbourhood siz Effect of suppression on number of overlaps “Effect of Stimulation on number of overlaps 0 10 20 30 40 50 60 0 0.2 0.4 0.6 0.8 1 Stimulation Number of overlaps AIS(av) SP (av) Figure 2: Effect of stimulation rate on prediction and recommendation. It can be seen that the simple AIS gives broadly similar prediction performance to the Simple Pearson. The MAE measurements from different runs are not normally distributed, so a non-parametric statistic is appropriate. We performed a Wilcoxon analysis, which showed that the difference between prediction errors of SP and AIS is zero with 95% confidence. In addition, the choice of an appropriate stimulation rate did make a significant difference (a rate of 0.2 compared with 0.02 at the 95% level). For recommendation, the AIS performs better than the SP at stimulation rates above 0.1. Again, we performed a positive 95% Wilcoxon analysis to assess significance. We excluded cases where a recommendation score was unavailable (due to an insufficient number of overlaps). The number of recommendations and overlaps show similar trends though the AIS gives a more constant value. Again, some stimulation was beneficial. In later experiments, the stimulation rate was fixed at one of the better values (0.2, 0.3 or 0.5), in order to give us a good base to work on. These values give us generally good performance, while keeping a good neighbourhood size and still evaluating a reasonable number of reviewers. Experiments on the Idiotypic AIS Having fixed all the simple parameters, we tested the effect of suppression for stimulation rates of 0.2, 0.3 and 0.5. Not surprisingly we found that suppression changed the number of reviewers looked at and the number of neighbours: Effect of suppression on neighbourhood size 0 10 20 30 40 50 60 70 80 90 100 0 0.2 0.4 0.6 0.8 1 Suppression rate Neighbourhood size Rate 0.2 Rate 0.3 Rate 0.5 Effect of suppression on number of reviewers looked at 0 2000 4000 6000 8000 10000 12000 14000 16000 0 0.2 0.4 0.6 0.8 1 Suppression Rate Number reviewers Rate 0.2 Rate 0.3 Rate 0.5 Figure 3: Effect of suppression rate on neighbourhood size and reviewers. We then tested the effect of suppression on the AIS performance. Here we fixed the baseline rate at stimulation only (no suppression), and took measurements relative to this baseline. Again, it should be noted that the first graph shows prediction error (hence, a good result is low). Effect of suppression on prediction error 70.0% 80.0% 90.0% 100.0% 110.0% 120.0% 130.0% 0 0.2 0.4 0.6 0.8 1 Suppression rate Mean absolute error (relative to baseline ) Rate 0.2 Rate 0.3 Rate 0.5 Effect of suppression on recommendation accuracy 80.0% 90.0% 100.0% 110.0% 120.0% 0 0.2 0.4 0.6 0.8 1 Suppression rate Recommendation accuracy (Kendall) relative to baseline Rate 0.2 Rate 0.3 Rate 0.5 Effect of suppression on number of overlaps 0.0% 20.0% 40.0% 60.0% 80.0% 100.0% 0 0.2 0.4 0.6 0.8 1 Suppression rate Number of overlaps (relative to baseline) Rate 0.2 Rate 0.3 Rate 0.5