Example - Optimize Baseline

• Find the system performance of the normalized set (baseline without child-rule, 87 SD-Rule)
  shell> GetSdRule 2013
> 1
> Test
  => Make sure manually remove all child-rules from 97-Rule and put to sdRules.stats.in
  > 0
  • The total valid SD-Pair instance count is 37,136
  • The cutoff SD-Rule for the optimized set is:

    Rule No.	A. Rate	Occr.	Yes	No	Tbd	SD-Rule	Status	Source	Notes	Sys A. Rate	Sys C. Rate	Sys. Perf
    60	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.76%	91.84%	1.8760

• Go through each parent-rule to replace parent-rule by child-rules and compare the results:
  • Decompose parent-rule
    • Decompose the parent-rule and find SD-pairs stats
      edit the sdRule.data by update the parent-rule
      shell>GetSdRule 2013
      7
      40 (min. coverage rate for further decomposition)
      35 (min. coverage rate for child candidate)
      also, child must has high accuracy rate than the root parent-rule
      Manually look through the output file sdRule.decompose.out and search for "<= Candidate", these candidates are child-rules match following criteria:
      • the accuracy rate is higher than parent-rule
      • the coverage rate is higher than 35% (or the specified number)
    • The decomposed candidate child-rules (shown in the 3rd column in the table below) are used to replace parent-rule for evaluating optimized SD-rule set.
  • Evaluate child-rules
    • Manually replace parent-rule with found candidate child-rules and find the system performance
shell>GetSdRule 2013
1
Test
37136 (this is the total valid SD-Pair from parent only rule, baseline)
...
    • Combine all cases with better system performance. A better system performance should have:
      • More rules included (ID)
      • High Sys. Perf (System performance)

  The iterative results are shown in the following table

  ID	Parent-Rule	Candidate Child-Rules	Rule No.	A. Rate	Occr.	Yes	No	Tbd	SD-Rule	Status	Source	Notes	Sys A. Rate	Sys C. Rate	Sys. Perf	Notes
  0	Parent-rule only (Baseline)	No child-Rule	60	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.76%	91.84%	1.8760
  1.1	0|2063|2016|47|$|adj|ity$|noun|97.72%|100.00%	1|935|926|9|c$|adj|city$|noun|99.04%|45.32% 1|725|707|18|l$|adj|lity$|noun|97.52%|35.14%	61	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.77%	90.81%	1.8658	Worse
  1.2	0|2063|2016|47|$|adj|ity$|noun|97.72%|100.00%	2|934|926|8|ic$|adj|icity$|noun|99.14%|45.27% 1|725|707|18|l$|adj|lity$|noun|97.52%|35.14%	61	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.77%	90.81%	1.8658	Worse
  2.1	0|1311|954|357|$|noun|al$|adj|72.77%|100.00%	1|666|550|116|n$|noun|nal$|adj|82.58%|50.80%	64	62.65%	332	208	124	0	$|noun|ist$|noun	2013	ORG_RULE	SELF	95.10%	94.16%	1.8926	Better
  2.2	0|1311|954|357|$|noun|al$|adj|72.77%|100.00%	2|614|526|88|on$|noun|onal$|adj|85.67%|46.83%	64	62.65%	332	208	124	0	$|noun|ist$|noun	2013	ORG_RULE	SELF	95.17%	94.09%	1.8926	Better
  2.3	0|1311|954|357|$|noun|al$|adj|72.77%|100.00%	3|571|491|80|ion$|noun|ional$|adj|85.99%|43.55%	65	60.66%	183	111	72	0	ar$|adj|e$|noun	2013	ORG_RULE	SELF	95.01%	94.30%	1.8931	Better (Best)
  2.4	0|1311|954|357|$|noun|al$|adj|72.77%|100.00%	4|466|402|64|tion$|noun|tional$|adj|86.27%|35.55%	65	60.66%	183	111	72	0	ar$|adj|e$|noun	2013	ORG_RULE	SELF	95.04%	94.06%	1.8910	Better
  3.1	0|263|227|36|a$|noun|an$|adj|86.31%|100.00%	No candidate child-rule found														Same
  4.1	0|273|1|272|a$|noun|an$|noun|0.37%|100.00%	1|136|1|135|ia$|noun|ian$|noun|0.74%|49.82%	60	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.76%	91.84%	1.8760	Same
  5.1	0|138|120|18|a$|noun|ar$|adj|86.96%|100.00%	1|115|105|10|la$|noun|lar$|adj|91.30%|83.33%	60	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.78%	91.80%	1.8758	Worse
  5.2	0|138|120|18|a$|noun|ar$|adj|86.96%|100.00%	2|69|65|4|ula$|noun|ular$|adj|94.20%|50.00%	60	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.79%	91.70%	1.8748	Worse
  6.1	0|306|303|3|ance$|noun|ant$|adj|99.02%|100.00%	No candidate child-rule found														Same
  7.1	0|2429|2410|19|ation$|noun|e$|verb|99.22%|100.00%	1|1007|1006|1|sation$|noun|se$|verb|99.90%|41.46% 1|1222|1222|0|zation$|noun|ze$|verb|100.00%|50.31%	61	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.78%	91.35%	1.8714	Worse
  7.2	0|2429|2410|19|ation$|noun|e$|verb|99.22%|100.00%	2|985|985|0|isation$|noun|ise$|verb|100.00%|40.55% 1|1222|1222|0|zation$|noun|ze$|verb|100.00%|50.31%	61	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.78%	91.30%	1.8708	Worse
  8.1	0|280|275|5|ency$|noun|ent$|adj|98.21%|100.00%	No candidate child-rule found														Same
  9.1	0|1004|1000|4|sis$|noun|tic$|adj|99.60%|100.00%	1|329|327|2|esis$|noun|etic$|adj|99.39%|32.77% 1|366|366|0|osis$|noun|otic$|adj|100.00%|36.45% 1|214|214|0|ysis$|noun|ytic$|adj|100.00%|21.31%	62	73.68%	19	14	5	0	a$|noun|iasis$|noun	2013	ORG_RULE	SELF	95.75%	91.59%	1.8735	Worse

From above table, only case-2 provides better results when replacing parent-rule by child-rules. In addition, the case 2.3 has the best system performance of 1.8931 (with system accuracy rate of 95.01% and system coverage rate of 94.30%) and thus is used as our optimized set of SD-Rules for the original SD-rules. The diagram below shows the system accuracy rate vs. system coverage rate. The optimized cutoff point is right around the intersection of these two curves to include 65 (out of 87) SD-rules.