Page 36 - Proceedings of the 2018 ITU Kaleidoscope
P. 36

2018 ITU Kaleidoscope Academic Conference






                       C6     C7     C8      C9    C10
                C1      0     1,19    0      0      0
                C2      0     0,24   0,71    0      0
                C3     1,66   2,14   2,38    0      0
                C4      0      0      0      0      0
                C5     0,48    0     0,24    0     2,14
                C6      0      0      0      0      0
                C7     1,66   1,90    0     0,24    0
                C8      0      0     0,48    0     0,95
                C9      0      0      0     3,33    0
                C10    2,38    0      0      0      0
           Table  2  shows  the  distribution  of  observations  among
           clusters, and compares the first versus the second execution.
           There are five cells with frequencies above 6%, and eight
           cells above 5%. Depending on the cutoff, there may be five
           or eight stable classes, and this is similar to what we have
           interpreted from Figure 1.
                                                                          Figure 2 - Clusters’ centroids
                  Table 3 - Observed F values for K groups.


                          K             F value
                           2                  84,31
                           3                 194,52
                           4                 162,58
                           5                  38,78
                           6                  25,46
                           7                  52,52
                           8                 145,45
                           9                 103,23
                          10                  32,79

           Even though the variables in this study do not follow an exact
           Gaussian  distribution,  we  estimated  the  values  of  the  F
           statistic.  Notice  in  Table  3  that  in  all  cases  the  value  is
           superior  to  10. The  statistic  is  below  the  empirical  cutoff
           when K=19. We continue the analysis based on the number   Figure 3 - Clusters’ centroids between 88 MHz and 108
           of clusters determined by the empirical methods.                         MHz

                       4.2 Interpretation of the clusters     In  Figure  2,  we  have  noted  at  first  sight  that  one  of  the
                                                              clusters, C3, groups the cases of all subscribers who hired the
           We performed this analysis with four to seven clusters and   Internet  service  only  (usually  our  subscribers  hire  both
           decided that six would be enough. When    varied, we saw   broadband access and video services). We continue to plot
           on the 3G-downlink band that when there were fewer classes,  the  frequency  bands  that  are  of  main  interest  for  ingress
           at least one of them grouped mixed patterns. Setting    = 7,   detection.
           the patterns that we identified were the same as the ones seen
           with     = 6 .  The  extra  cluster  did  not  add  any  new   According to Figure 3, the cluster C4 centroid is consistently
           information to the analysis.                       above  all  the  rest,  and  C3  is  consistently  below.  In  this
                                                              frequency  range,  we  could  assume  there  was  radio  signal
           Next, we present the results obtained after executing the k-  ingress.
           means algorithm with K=6.











                                                           – 20 –
   31   32   33   34   35   36   37   38   39   40   41