Given the quality of prediction is reasonably good with small model size, we focus on the run-time and throughput of the system. We recorded the time required to generate predictions for the entire test set and plotted them in a chart with varying model size. We plotted the run time at different x values. Figure 8 shows the plot. Note here that at x=0.25 the whole system has to make prediction for 25,000 test cases. From the plot we observe a substantial difference in the run-time between the small model size and the full item-item prediction case. For x=0.25the run-time is 2.002 seconds for a model size of 200 as opposed to 14.11 for the basic item-item case. This difference is even more prominent with x=0.8 where a model size of 200 requires only 1.292 seconds and the basic item-item case requires 36.34seconds.
These run-time numbers may be misleading as we computed them for different train/test ratios where the work-load size i.e., number of predictions to be generated is different (recall that at x=0.3 our algorithm uses 30,000 ratings as training data and uses the rest of 70,000 ratings as test data to compare predictions generated by the system to the actual ratings). To make the numbers comparable we compute the throughput (predictions generated per second) for the model based and basic item-item schemes. Figure 8 charts these results. We see that for x=0.3 and at a model size of 100 the system generates 70,000 ratings in 1.487 seconds producing a throughput rate of 47,361whereas the basic item-item scheme produced a throughput of 4961only. At x=0.8 these two numbers are 21,505 and 550respectively.