Performance Over Time Sample Clauses
Performance Over Time. At each timestamp, the (smoothed) DoP and Error are shown in Figure 4.10. As expected, Jtrs provides the strongest protection of privacy, while JutiL and DPLS DPMM−G util DPMM−G trs DPLS DPMM−G util DPMM−G trs 120 80 DoP DPLS DPMM−G util DPMM−G trs
(a) DoP on GeoLife
(b) DoP on Gowalla
(c) Error on GeoLife
(d) Error on Gowalla Figure 4.10: Performance over time.
Performance Over Time. In order to show the performance of a release mechanism as a user moves over time, including how ∆X changes, how often drift happens and how accurate is the perturbed location, we first run a set of experiments for a single test trajectory with popular M learned from all users. We selected a random test trajectory from Geolife dataset consisting of 500 timestamps. We tested both PIM and LM at each timestamp with ϵ = 1 and δ = 0.01. Each method was run 20 times and the average is reported. Figure 3.4a shows the original trajectory in map and state (grid) coordinates; Figures 3.4b and 3.4c show the released (perturbed) locations at each timestamp. We can see that the released locations of PIM is closer to the true location, compared with LM. Size of ∆X. From Figure 3.4d we see that the size of ∆X does not increase dramatically, instead it maintains at stable level after a few timestamps. The reason is that by selecting the δ-location set the inference mechanism only boost probabilities of locations in ∆X. Then the probabilities of other locations decay gradually. Thus a stable δ-location set can be maintained.