macks22 / gi Goto Github PK

Implements Sequtur (online) and Re-Pair (off-line) grammar induction algorithms for Grammarviz 2.0 and SAX-VSM-G. This code is released under GPL v.2.0.

[1] Nevill-Manning, C.G. and Witten, I.H., "Identifying Hierarchical Structure in Sequences: A linear-time algorithm", Journal of Artificial Intelligence Research, 7, 67-82, (1997).

[2] Larsson, N.J.; Moffat, A., "Offline dictionary-based compression", Data Compression Conference, 1999. Proceedings. DCC '99 , vol., no., pp.296,305, 29-31 Mar 1999.

If you are using this implementation for you academic work, please cite our Grammarviz 2.0 paper:

[Citation] Senin, P., Lin, J., Wang, X., Oates, T., Gandhi, S., Boedihardjo, A.P., Chen, C., Frankenstein, S., Lerner, M., GrammarViz 2.0: a tool for grammar-based pattern discovery in time series, ECML/PKDD Conference, 2014.

The code is written in Java and I use maven to build it:

$ mvn package
[INFO] Scanning for projects...
[INFO] ------------------------------------------------------------------------
[INFO] Building GI
[INFO]    task-segment: [package]
...
[INFO] Building jar: /media/Stock/git/jmotif-GI.git/target/jmotif-gi-0.3.1-SNAPSHOT.jar
[INFO] ------------------------------------------------------------------------
[INFO] BUILD SUCCESSFUL
[INFO] ------------------------------------------------------------------------

Following the original Eibe Frank's java implementation the code is built using global (static) variables:

String TEST3_STRING = "a b a b c a b c d a b c d e a b c d e f";

SAXRule r = SequiturFactory.runSequitur(TEST3_STRING);

System.out.println(SAXRule.printRules());

which prints the following output:

Number	Name	Level	Occurr.	Usage	Yield	Rule str	Expaneded	Indexes
0	R0	0	0	0	0	R1 R2 R3 R4 R4 f 	a b a b c a b c d a b c d e a b c d e f	[]
1	R1	1	5	2	2	a b 	a b 	[0, 2, 5, 9, 14]
2	R2	1	4	2	3	R1 c 	a b c 	[2, 5, 9, 14]
3	R3	1	3	2	4	R2 d 	a b c d 	[5, 9, 14]
4	R4	1	2	2	5	R3 e 	a b c d e 	[9, 14]

My own addition allows to retrieve the Sequitur rules as an iterable collection of GrammaRuleRecords and to map them back to the discretized time series:

GrammarRules rules = r.toGrammarRulesData();
GrammarRuleRecord rec = rules.get(4);
ArrayList<RuleInterval> intervals = rec.getRuleIntervals();
...

I've implemented RePair from scratch and it uses the same GrammaRules / GrammaRuleRecord data structures as for Sequitur, so it can be plugged into Grammarviz seamlessly:

String TEST_STRING = "abc abc cba XXX abc abc cba";

RePairGrammar rg = RePairFactory.buildGrammar(TEST_STRING);

System.out.println(rg.toGrammarRules());

which yields:

	R0 -> R2 XXX R2 
    R1 -> abc cba  : abc cba, [1, 5]
    R2 -> abc R1  : abc abc cba, [0, 4]

Thanks to the algorithm's design, I was able to parallelize RePair. However, the cost of inter-tread communications and synchronization was the majot showstopper, so the current new implementation (>0.8.5) is single-threaded (but you can still get the parallel one -- it is tagged "old_repair" in the version control).

The both implemented GI algorithms, Sequitur and RePair, demonstrate a somewhat similar performance with minor differnces. Specifically:

• Sequitur implementation is slower than RePair
• Sequitur tends to produce more rules, but Sequitur rules are less frequent than RePair rules
• Sequitur rule-corresponding subsequences vary in length more
• Sequitur rules usually cover more points than RePair
• Sequitur rule coverage depth is lower than that of RePair

All these may affect the performance of the upstream time series analysis algorithms such as SAX-VSM-G, Grammarviz, and RRA. Here is the table with some numbers collected by running Sequitur and RePair using sliding window of size 150, PAA 6, and the alphabet 4. I used the EXACT numerosity reduction in these runs.

1.0.0

• more tests & Grammarviz3.0 release.

0.8.6

• pre-1.0 release with improved RePair implementation.

0.0.1 - 0.8.5

• initial code development, parallel repair implementation lifecycle.