counting the number of discordant pairs. To do this we order tion on number of users looked at the films by vote and apply the following formulae N=∑∑D(,) D(,r) ifr> otherwise Figure 1: Effect of stimulation rate on neighbourhood and reviewers Where n is the overlap size and r, is the rank of film i as ecommended by the neighbourhood. Note that i here refers The graphs show averaged results over five runs at each to the antigen rank of the film, not the film ID. No is the stimulation rate. The bars show standard deviations. In order to have a fair comparison, the Simple Pearson parameters cost of a bubble sort to reconcile the two lists. d is set to one (neighbourhood and number of reviewers looked at)match if the rankings are discordant the Ais values for each rate. In figure 2, we show the prediction error, number of recommendations, number of Mean number of reviewers. This is the num ber overlaps and recommendation accuracy for each algorithm reviewers looked at before the ais stabilised Note that low prediction error values are better, whereas for the other measures we are looking for high values Mean number of neighbours This is the final number of neighbours in the stabilised AIS 3 EXPERIMENTS Experiments were carried out on a Pentium 700 in JavaTM JDK13 Each run involved looking at up to 15,000 as/ 256MB RAM, running Windows 2000. The AlS was coded reviewers(20% of the Each Movie data set, randomly chosen) to provide predictions and recommendations for 100 users Averaged statistics are then taken for each run. Runtimes ranged from 5 to 60 minutes, largely dependent on the Effect of stimulation on recommendation accuracy number of reviewers Experiments on Simple AlS Initial experiments concentrated on a simple AlS, with diotypic effects. The goal was to find a good stimulation ii rate, but also to ensure that the baseline'system operates i similarly to a Simple Pearson predictor (SP). Therefore, we set the suppression rate to zero, and varied only the stimulation rate, i.e. the weighting given to antigen binding 0.2406 Other parameters had been fixed by preliminary experiments EfLect of stimulation on number of recommendations Effect of Stimultion on Neighbourhood size Sumulationcounting the number of discordant pairs. To do this we order the films by vote and apply the following formulae: ( ) ( ) ( )    > = = − = − ∑ ∑= + = otherwise if r r D r r N D r r n n N i j i j n i n j i D i j D 0 1 , , )6( 1 4 1 1 1 τ Where n is the overlap size and ri is the rank of film i as recommended by the neighbourhood. Note that i here refers to the antigen rank of the film, not the film ID. ND is the number of discordant pairs, or, equivalently, the expected cost of a bubble sort to reconcile the two lists. D is set to one if the rankings are discordant. Mean number of reviewers. This is the number of reviewers looked at before the AIS stabilised. Mean number of neighbours: This is the final number of neighbours in the stabilised AIS. 3 EXPERIMENTS Experiments were carried out on a Pentium 700 with 256MB RAM, running Windows 2000. The AIS was coded in JavaTM JDK1.3. Each run involved looking at up to 15,000 reviewers (20% of the EachMovie data set, randomly chosen) to provide predictions and recommendations for 100 users. Averaged statistics are then taken for each run. Runtimes ranged from 5 to 60 minutes, largely dependent on the number of reviewers. Experiments on Simple AIS Initial experiments concentrated on a simple AIS, with no idiotypic effects. The goal was to find a good stimulation rate, but also to ensure that the ‘baseline’ system operates similarly to a Simple Pearson predictor (SP). Therefore, we set the suppression rate to zero, and varied only the stimulation rate, i.e. the weighting given to antigen binding. Other parameters had been fixed by preliminary experiments. Effect of Stimulation on Neighbourhood size 0 10 20 30 40 50 60 70 80 90 100 0 0.2 0.4 0.6 0.8 1 Stimulation Rate Neighbourhood Size Effect of stimulation on number of users looked at 0 5000 10000 15000 0 0.2 0.4 0.6 0.8 1 Stimulation Rate Num ber of users looked at Figure 1: Effect of stimulation rate on neighbourhood and reviewers. The graphs show averaged results over five runs at each stimulation rate. The bars show standard deviations. In order to have a fair comparison, the Simple Pearson parameters (neighbourhood and number of reviewers looked at) match the AIS values for each rate. In figure 2, we show the prediction error, number of recommendations, number of overlaps and recommendation accuracy for each algorithm. Note that low prediction error values are better, whereas for the other measures we are looking for high values. Effect of stimulation on prediction error 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0 0.2 0.4 0.6 0.8 1 Stimulation Rate M ean Absolute E rror AIS (av) SP (av) Effect of stimulation on recommendation accuracy 0.35 0.4 0.45 0.5 0.55 0 0.2 0.4 0.6 0.8 1 Stimulation Rate Recommendation Accuracy (Kendall's Tau) AIS (av) SP (av) Effect of Stimulation on number of recommendations 0 200 400 600 800 1000 1200 0 0.2 0.4 0.6 0.8 1 Stimulation Rate Number of recommendations AIS(av) SP (av)