Tham khảo tài liệu 'data analysis machine learning and applications episode 2 part 5', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | ADSL customer segmentation combining several SOMs 349 Daily Activities Profiles log file Unknown Down P2P-up Web Down Customers 1 1 s 1 1 Binary Profile . Binary Profile . Binary Profile STEP 1 STEP 2 Typical Days STEP 3 Proportion of days spent in each typical days for the month STEP 4 IDEZ Typical Customers STEP 5 Fig. 6. The multi-level exploratory data analysis approach. The first step leads to the formation of 9 to 13 clusters of typical application days profiles depending on the application. Their behaviours can be summarized into inactive days days with a mean or high activity on some limited time periods early or late evening noon for instance and days with a very high activity on a long time segment working hours afternoon or night . Figure 7 illustrates the result of the first step for one application it shows the mean hourly volume profiles of the 13 clusters revealed after the clustering for the web down application the mean profiles are computed by the mean of all the observations that have been classified in the cluster the hourly volumes are plotted in natural statistics . The other applications can be described similarly. 350 Francoise Fessant et al. hours ----C1 .C2 .C3 C4 ----C5 ----C6 ----C7 - - -C8 I C9 C10 - -C11 - - C12 B C13 Fig. 7. Mean daily volumes of clusters for web down application The second clustering leads to the formation of 14 clusters of typical days . Their behaviours are different in terms of traffic time periods and intensity. The main characteristics are a similar activity in up and down traffic directions and a similar usage of the peer-to-peer and unknown applications in clusters. The usage of the web application can be quite different in intensity. Globally the time periods of traffic are very similar for the three applications in a cluster. 10 percent of the days show a high daily activity on the three applications 25 percent of the days are inactive days. If we project the other applications on the map days we can observe
