The document discusses techniques for mining association rules from big data using compressed bitsets. It compares the performance of the ECLAT frequent pattern mining algorithm using different bitset compression methods like EWAH, CONCISE, Roaring and BitMagic. The experiments vary dataset size, minimum support, transaction length, item count and frequent pattern count to analyze the effect on runtime, memory usage and energy consumption. The results show that more sophisticated compression does not always perform better and that the best technique depends on the specific problem characteristics.

1. The Merits of Bitset Compression
Techniques for Mining Association
Rules from Big Data
Hamid Fadishei*, Sahar Doustian, Parisa Saadati
University of Bojnord, Iran
*fadishei@ub.ac.ir
TOPHPC 2017 24-26 APRIL, TEHRAN, IRAN 1

2. Outline
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 2

3. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 3
Big Data
Analytics
Outline

4. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 4
Big Data
Analytics
Set Math
Relies on
Outline

5. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 5
Big Data
Analytics
Set Math
Relies on
Compressed
Bitsets
Accelerated
by
Outline

6. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 6
Big Data
Analytics
Set Math
Relies on
Compressed
Bitsets
Accelerated
by
Some
Algorithm
Another
Algorithm
Yet Another
Algorithm
...
Implemented by
Outline

7. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 7
Big Data
Analytics
Set Math
Relies on
Compressed
Bitsets
Accelerated
by
Some
Algorithm
Another
Algorithm
Yet Another
Algorithm
...
Implemented by
Comparative
Study
Outline

8. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 8
Big Data
Analytics
Set Math
Relies on
Compressed
Bitsets
Accelerated
by
Some
Algorithm
Another
Algorithm
Yet Another
Algorithm
...
Implemented by
Comparative
Study
Outline
Focus of
this paper!

9. Frequent Pattern Mining
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 9

10. Seeking Frequent Patterns in Big Data
Problem of finding the itemsets whose occurrence count is more than a predefined “support” [2]
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 10

11. Seeking Frequent Patterns in Big Data
Problem of finding the itemsets whose occurrence count is more than a predefined “support”
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 11
Transaction Basket Items
1 Bread, Beer, Diaper
2 Beer, Milk, Diaper
3 Beer, Diaper, Milk, Nuts
4 Bread, Milk, Diaper
5 Beer, Diaper

12. Seeking Frequent Patterns in Big Data
Problem of finding the itemsets whose occurrence count is more than a predefined “support”
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 12
Transaction Basket Items
1 Bread, Beer, Diaper
2 Beer, Milk, Diaper
3 Beer, Diaper, Milk, Nuts
4 Bread, Milk, Diaper
5 Beer, Diaper

13. Algorithms for Frequent Pattern Mining
Many algorithms, some of the most well known ones are:
◦ Apriori [3] Scans data multiple times to generate the itemsets of length 1, then 2, then 3, and so on
◦ FPGrowth [4] Constructs a tree and recursively extracts patterns from it
◦ ECLAT [5] Utilizes vertical representation of dataset
The present study uses ECLAT
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 13

14. Vertical Presentation of Data
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 14
Item Transactions
Bread 1, 4
Beer 1, 2, 3, 5
Milk 2, 3, 4
Diaper 1, 2, 3, 4, 5
Nuts 3
Transaction Items
1 Bread, Beer, Diaper
2 Beer, Milk, Diaper
3 Beer, Diaper, Milk, Nuts
4 Bread, Milk, Diaper
5 Beer, Diaper

15. Bitset Encoding of Vertical Form
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 15
Item Transactions
Bread 1, 4
Beer 1, 2, 3, 5
Milk 2, 3, 4
Diaper 1, 2, 3, 4, 5
Nuts 3
1 2 3 4 5
Bread 1 0 0 1 0
Beer 1 1 1 0 1
Milk 0 1 1 1 0
Diaper 1 1 1 1 1
Nuts 0 0 1 0 0
Transaction Items
1 Bread, Beer, Diaper
2 Beer, Milk, Diaper
3 Beer, Diaper, Milk, Nuts
4 Bread, Milk, Diaper
5 Beer, Diaper

16. ECLAT Algorithm
Repeatedly appends patterns and calculates support by counting ones
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 16
1 2 3 4 5
Bread 1 0 0 1 0
Beer 1 1 1 0 1
Milk 0 1 1 1 0
Diaper 1 1 1 1 1
Nuts 0 0 1 0 0

17. ECLAT Algorithm
Repeatedly appends patterns and calculates support by counting ones
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 17
1 2 3 4 5
Bread 1 0 0 1 0
Beer 1 1 1 0 1
Milk 0 1 1 1 0
Diaper 1 1 1 1 1
Nuts 0 0 1 0 0

18. ECLAT Algorithm
Repeatedly appends patterns and calculates support by counting ones
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 18
1 2 3 4 5
Bread 1 0 0 1 0
Beer 1 1 1 0 1
Milk 0 1 1 1 0
Diaper 1 1 1 1 1
Nuts 0 0 1 0 0
1 2 3 4 5
Beer, Milk 1 0 0 1 0

19. ECLAT Algorithm
Repeatedly appends patterns and calculates support by counting ones
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 19
1 2 3 4 5
Bread 1 0 0 1 0
Beer 1 1 1 0 1
Milk 0 1 1 1 0
Diaper 1 1 1 1 1
Nuts 0 0 1 0 0
1 2 3 4 5
Beer, Milk 1 0 0 1 0
Beer, Diaper 1 1 1 0 1

20. ECLAT Algorithm
Repeatedly appends patterns and calculates support by counting ones
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 20
1 2 3 4 5
Bread 1 0 0 1 0
Beer 1 1 1 0 1
Milk 0 1 1 1 0
Diaper 1 1 1 1 1
Nuts 0 0 1 0 0
1 2 3 4 5
Beer, Milk 1 0 0 1 0
Beer, Diaper 1 1 1 0 1

21. ECLAT Algorithm
Most time consuming parts are set operations
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 21

22. ECLAT Algorithm
Most time consuming parts are set operations
◦ Calculating intersections
◦ Calculating cardinalities
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 22

23. ECLAT Algorithm
Most time consuming parts are set operations
◦ Calculating intersections by ANDing bitsets
◦ Calculating cardinalities by counting ONEs
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 23

24. Bitset Compression
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 24

25. Bitset Compression Methods
Bitset compression techniques
◦ Many of them
◦ The present study focuses on 4 of them
◦ EWAH [7]
◦ CONCISE [8]
◦ Roaring [9]
◦ BitMagic [10]
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 25

26. Bitset Compression Methods
Bitset compression techniques
◦ Many of them
◦ The present study focuses on 4 of them
◦ EWAH [7]
◦ CONCISE [8]
◦ Roaring [9]
◦ BitMagic [10]
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 26
Simple, Based on RLE

27. Bitset Compression Methods
Bitset compression techniques
◦ Many of them
◦ The present study focuses on 4 of them
◦ EWAH [7]
◦ CONCISE [8]
◦ Roaring [9]
◦ BitMagic [10]
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 27
Simple, Based on RLE
More sophisticated

28. Bitset Compression: EWAH
EWAH uses RLE compression
It defines two type of words
◦ Marker – for sparse parts
◦ Dirty – for uncompressible parts
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 28

29. Bitset Compression: CONCISE
Similar to EWAH but defines an addition:
◦ Ability to define a single-bit exception inside the marker word
◦ Tends to reduce memory usage
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 29

30. Bitset Compression: Roaring
Roaring is more sophisticated
Uses a notion of hybrid containers
◦ Sparse parts are stored in array containers
◦ Dense parts in bitmap containers
Containers are organized as a two-level tree
◦ Small root that can usually fit into CPU cache
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 30
Array container
0
62
124
186
248
310
.
.
.
.
61938
Bits:0X0000&Cardinality:1000
Bitmap container
1
0
1
0
1
0
.
.
.
.
0
Bits:0X0002&Cardinality:2
15
Array of containers
Array container
0
1
2
3
4
5
.
.
.
.
99
Bits:0X0001&Cardinality:100

31. Bitset Compression: BitMagic
Another container-based bitset compression technique
Simliar to Roaring, but more simplistic
◦ Does not use array containers
◦ Does not exploits binary search for calculating intersections
◦ Does not use heuristics to decide which container type should be used for results of operations
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 31

32. Experiments
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 32

33. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 33
ECLAT
(eclat.c, etc)
Bitset interface
(Compile-time binding)
(wrapper.h)
EWAH
(wrapper_ewah.cpp)
BitMagic
(wrapper_bm.cpp)
Roaring
(wrapper_roaring.c)
CONCISE
(wrapper_concise.cpp)
Energy and
Performance
Monitoring
(stats.c)
Real-world
Dataset
Params
IBM Quest
Generator
Synthetic
Dataset
Output
Results
Experimental Framework

34. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 34
ECLAT
(eclat.c, etc)
Bitset interface
(Compile-time binding)
(wrapper.h)
EWAH
(wrapper_ewah.cpp)
BitMagic
(wrapper_bm.cpp)
Roaring
(wrapper_roaring.c)
CONCISE
(wrapper_concise.cpp)
Energy and
Performance
Monitoring
(stats.c)
Real-world
Dataset
Params
IBM Quest
Generator
Synthetic
Dataset
Output
Results
Experimental Framework

35. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 35
ECLAT
(eclat.c, etc)
Bitset interface
(Compile-time binding)
(wrapper.h)
EWAH
(wrapper_ewah.cpp)
BitMagic
(wrapper_bm.cpp)
Roaring
(wrapper_roaring.c)
CONCISE
(wrapper_concise.cpp)
Energy and
Performance
Monitoring
(stats.c)
Real-world
Dataset
Params
IBM Quest
Generator
Synthetic
Dataset
Output
Results
Experimental Framework

36. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 36
ECLAT
(eclat.c, etc)
Bitset interface
(Compile-time binding)
(wrapper.h)
EWAH
(wrapper_ewah.cpp)
BitMagic
(wrapper_bm.cpp)
Roaring
(wrapper_roaring.c)
CONCISE
(wrapper_concise.cpp)
Energy and
Performance
Monitoring
(stats.c)
Real-world
Dataset
Params
IBM Quest
Generator
Synthetic
Dataset
Output
Results
Experimental Framework
Instacart (~3M transactions)
Kosarak (~1M transactions)

37. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 37
ECLAT
(eclat.c, etc)
Bitset interface
(Compile-time binding)
(wrapper.h)
EWAH
(wrapper_ewah.cpp)
BitMagic
(wrapper_bm.cpp)
Roaring
(wrapper_roaring.c)
CONCISE
(wrapper_concise.cpp)
Energy and
Performance
Monitoring
(stats.c)
Real-world
Dataset
Params
IBM Quest
Generator
Synthetic
Dataset
Output
Results
Experimental Framework
Server under test:
Core i5, 32GB RAM

38. Studying the effects of…
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 38
Dataset
Size
Minimum
Support
Transaction
Length
Item
Count
Frequent
Pattern
Count

39. …on
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 39
Run
Time
Memory
Usage
Energy
Consumption

40. Effect of dataset size
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 40

41. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 41

42. Effect of minsup
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 42

43. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 43

44. Effect of transaction length
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 44

45. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 45

46. Effect of item count
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 46

47. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 47

48. Count of frequent itemsets
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 48

49. TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 49

50. Conclusions and Lessons Learnt
Different bitset compression techniques can exhibit dramatically different behavior.
There is no always-best advice for selecting the proper technique.
More sophisticated does not always mean better
Devising an advisory layer can be a promising future work.
◦ A framework that predicts the possibly best lower level technique at runtime
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 50

51. Thanks, many thanks!
TOPHPC 2019 23-25 APRIL, TEHRAN, IRAN 51
Espidan village, Bojnord, Iran