This shows you the differences between two versions of the page.

Link to this comparison view

blogs:pub2016:divide_and_conquer [2016/09/30 11:11] (current)
Line 1: Line 1:
 +====== Divide and conquer ======
 +H. M. W. Verbeek, W. M. P. van der Aalst, and J. Munoz-Gama, [[http://​bpmcenter.org/​wp-content/​uploads/​reports/​2016/​BPM-16-06.pdf|Divide and conquer]], BPMCenter.org,​ BPM Center Report BPM-16-06, 2016.
 +===== Abstract =====
 +Process ​ mining ​ has  been  around ​ for  more  than  a
 +decade ​ now,  and  in  that  period ​ several ​ discovery ​ and  replay
 +algorithms have been introduced that work fairly well on average-
 +sized  event  logs.  Nevertheless, ​ these  algorithms ​ have  problems
 +dealing ​ with  big  event  logs.  If  the  algorithms ​ do  not  run  out
 +of  memory, ​ they  will  run  out  of  time,  because ​ the  problem
 +handed ​ to  them  is  just  too  complex ​ to  be  solved ​ in  reasonable
 +time  and  space. ​ For  this  reason, ​ a  generic ​ approach ​ has  been
 +developed which allows such big problems to be decomposed into
 +a series of smaller (say, average-sized) problems. This approach
 +offers ​ formal ​ guarantees ​ for  the  results ​ obtained ​ by  it,  and
 +makes  existing ​ algorithms ​ also  tractable ​ for  larger ​ logs.  As  a
 +result, ​ discovery ​ and  replay ​ problems ​ may  become ​ feasible, ​ or
 +may  become ​ easier ​ to  handle. ​ This  paper  introduces ​ a  tool
 +framework, ​ called
 +Divide ​ and  Conquer
 +that  fully  supports ​ this
 +generic approach, which has been implemented in ProM 6. Using
 +this  novel  framework, ​ this  paper  demonstrates ​ that  significant
 +speed-ups ​ can  be  achieved, ​ both  for  discovery ​ and  for  replay.
 +This  paper  also  shows  that  a
 +decomposition ​ (that  is,
 +a  decomposition ​ into  as  many  subproblems ​ as  possible) ​ is  not
 +always preferable: Non-maximal decompositions may run as fast,
 +and generally provide better results.