林智仁 (Chih-Jen Lin) 國立臺灣大學資訊工程學系特聘教授 林智仁教授畢業於台大數學系，之後前往美國密西根大學獲得碩士與博士學位。他的研究團隊多年來致力於機器學習之相關研究，所發展的資料分類軟體在世界上被廣泛使用，是台灣計算機科學界突出的重要成果。

1. Big-data Analytics: Challenges and
Opportunities
Chih-Jen Lin
Department of Computer Science
National Taiwan University
Talk at ðcÇ™Ñx}t, August 30, 2014
Chih-Jen Lin (National Taiwan Univ.) 1 / 54

2. Everybody talks about big data now, but it's not easy to
have an overall picture of this subject
In this talk, I will give some personal thoughts on
technical developments of big-data analytics. Some are
very pre-mature, so your comments are very welcome
Chih-Jen Lin (National Taiwan Univ.) 2 / 54

3. Outline
1 From data mining to big data
2 Challenges
3 Opportunities
4 Discussion and conclusions
Chih-Jen Lin (National Taiwan Univ.) 3 / 54

4. From data mining to big data
Outline
1 From data mining to big data
2 Challenges
3 Opportunities
4 Discussion and conclusions
Chih-Jen Lin (National Taiwan Univ.) 4 / 54

5. From data mining to big data
From Data Mining to Big Data
In early 90's, a buzzword called data mining
appeared
Many years after, we have another one called big
data
Well, what's the dierence?
Chih-Jen Lin (National Taiwan Univ.) 5 / 54

6. From data mining to big data
Status of Data Mining and Machine
Learning
Over the years, we have all kinds of eective
methods for classi

7. cation, clustering, and regression
We also have good integrated tools for data mining
(e.g., Weka, R, Scikit-learn)
However, mining useful information remains dicult
for some real-world applications
Chih-Jen Lin (National Taiwan Univ.) 6 / 54

8. From data mining to big data
What's Big Data?
Though many de

9. nitions are
available, I am considering
the situation that data are
larger than the capacity of a
computer
I think this is a main
dierence between data
mining and big data
So in a sense we are talking
about distributed data
mining or machine learning
(a), (b): distributed
systems
Image from Wikimedia
Chih-Jen Lin (National Taiwan Univ.) 7 / 54

10. From data mining to big data
From Small to Big Data
Two important dierences:
Negative side:
Methods for big data analytics are not quite ready,
not even mentioned to integrated tools
Positive side:
Some (Halevy et al., 2009) argue that the almost
unlimited data make us easier to mine information
I will discuss the

11. rst dierence
Chih-Jen Lin (National Taiwan Univ.) 8 / 54

12. Challenges
Outline
1 From data mining to big data
2 Challenges
3 Opportunities
4 Discussion and conclusions
Chih-Jen Lin (National Taiwan Univ.) 9 / 54

13. Challenges
Possible Advantages of Distributed Data
Analytics
Parallel data loading
Reading several TB data from disk is slow
Using 100 machines, each has 1/100 data in its
local disk ) 1/100 loading time
But having data ready in these 100 machines is
another issue
Fault tolerance
Some data replicated across machines: if one fails,
others are still available
Chih-Jen Lin (National Taiwan Univ.) 10 / 54

14. Challenges
Possible Advantages of Distributed Data
Analytics (Cont'd)
Work
ow not interrupted
If data are already distributedly stored, it's not
convenient to reduce some to one machine for
analysis
Chih-Jen Lin (National Taiwan Univ.) 11 / 54

15. Challenges
Possible Disadvantages of Distributed
Data Analytics
More complicated (of course)
Communication and synchronization
Everybody says moving computation to data, but
this isn't that easy
Chih-Jen Lin (National Taiwan Univ.) 12 / 54

16. Challenges
Going Distributed or Not Isn't Easy to
Decide
Quote from Yann LeCun (KDnuggets News 14:n05)
I have seen people insisting on using Hadoop for
datasets that could easily

17. t on a
ash drive and
could easily be processed on a laptop.
Now disk and RAM are large. You may load several
TB of data once and conveniently conduct all
analysis
The decision is application dependent
We will discuss this issue again later
Chih-Jen Lin (National Taiwan Univ.) 13 / 54

18. Challenges
Distributed Environments
Many easy tasks on one computer become dicult
in a distributed environment
For example, subsampling is easy on one machine,
but may not be in a distributed system
Usually we attribute the problem to slow
communication between machines
Chih-Jen Lin (National Taiwan Univ.) 14 / 54

19. Challenges
Challenges
Big data, small analysis
versus
Big data, big analysis
If you need a single record from a huge set, it's
reasonably easy
For example, accessing your high-speed rail
reservation is fast
However, if you want to analyze the whole set by
accessing data several time, it can be much harder
Chih-Jen Lin (National Taiwan Univ.) 15 / 54

20. Challenges
Challenges (Cont'd)
Most existing data mining/machine learning
methods were designed without considering data
access and communication of intermediate results
They iteratively use data by assuming they are
readily available
Example: doing least-square regression isn't easy in
a distributed environment
Chih-Jen Lin (National Taiwan Univ.) 16 / 54

21. Challenges
Challenges (Cont'd)
So we are facing many challenges
methods not ready
no convenient tools
rapid change on the system side
and many others
What should we do?
Chih-Jen Lin (National Taiwan Univ.) 17 / 54

22. Opportunities
Outline
1 From data mining to big data
2 Challenges
3 Opportunities
4 Discussion and conclusions
Chih-Jen Lin (National Taiwan Univ.) 18 / 54

23. Opportunities
Opportunities
Looks like we are in the early stage of a research
topic
But what is our chance?
Chih-Jen Lin (National Taiwan Univ.) 19 / 54

24. Opportunities Lessons from past developments in one machine
Outline
3 Opportunities
Lessons from past developments in one machine
Successful examples?
Design of big-data algorithms
Chih-Jen Lin (National Taiwan Univ.) 20 / 54

25. Opportunities Lessons from past developments in one machine
Algorithms for Distributed Data Analytics
This is an on-going research topic.
Roughly there are two types of approaches
1 Parallelize existing (single-machine) algorithms
2 Design new algorithms particularly for distributed
settings
Of course there are things in between
Chih-Jen Lin (National Taiwan Univ.) 21 / 54

26. Opportunities Lessons from past developments in one machine
Algorithms for Distributed Data Analytics
(Cont'd)
Given the complicated distributed setting, we
wonder if easy-to-use big-data analytics tools can
ever be available?
I don't know either. Let's try to think about the
situation on one computer

27. rst
Indeed those easy-to-use analytics tools on one
computer were not there at the

28. rst day
Chih-Jen Lin (National Taiwan Univ.) 22 / 54

29. Opportunities Lessons from past developments in one machine
Past Development on One Computer
The problem now is we take many things for
granted on one computer
On one computer, have you ever worried about
calculating the average of some numbers?
Probably not. You can use Excel, statistical
software (e.g., R and SAS), and many things else
We seldom care internally how these tools work
Can we go back to see the early development on
one computer and learn some lessons/experiences?
Chih-Jen Lin (National Taiwan Univ.) 23 / 54

30. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product
Consider the example of matrix-matrix products
C = A B; A 2 Rnd ; B 2 Rdm
where
Cij =
Xd
k=1
AikBkj
This is a simple operation. You can easily write your
own code
Chih-Jen Lin (National Taiwan Univ.) 24 / 54

31. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
A segment of C code (assume n = m here)
for (i=0;in;i++)
for (j=0;jn;j++)
{
c[i][j]=0;
for (k=0;kn;k++)
c[i][j] += a[i][k]*b[k][j];
}
For 3; 000 3; 000 matrices
$ gcc -O3 mat.c
$ time ./a.out
3m24.843s
Chih-Jen Lin (National Taiwan Univ.) 25 / 54

32. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
But on Matlab (single-thread mode)
$ matlab -singleCompThread
tic; c = a*b; toc
Elapsed time is 4.095059 seconds.
Chih-Jen Lin (National Taiwan Univ.) 26 / 54

33. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
How can Matlab be much faster than ours?
The fast implementation comes from some deep
research and development
Matlab calls optimized BLAS (Basic Linear Algebra
Subroutines) that was developed in 80's-90's
Our implementation is slow because data are not
available for computation
Chih-Jen Lin (National Taiwan Univ.) 27 / 54

34. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
CPU
#
Registers
#
Cache
#
Main Memory
#
Secondary storage (Disk)
: increasing in speed
#: increasing in
capacity
Optimized BLAS: try
to make data available
in a higher level of
memory
You don't waste time
to frequently move
data
Chih-Jen Lin (National Taiwan Univ.) 28 / 54

35. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
Optimized BLAS uses block algorithms
A B =
2
4
A11 A14
...
A41 A44
3
5
2
4
B11 B14
...
B41 B44
3
5
=
A11B11 + + A14B41
...
. . .
If we compare the number of page faults (cache
misses)
Ours: much larger
Block: much smaller
Chih-Jen Lin (National Taiwan Univ.) 29 / 54

36. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
I like this example because it involves both
mathematical operations (matrix products),
and
computer architecture (memory hierarchy)
Only if knowing both, you can make breakthroughs
Chih-Jen Lin (National Taiwan Univ.) 30 / 54

37. Opportunities Lessons from past developments in one machine
Example: Matrix-matrix Product (Cont'd)
For big-data analytics, we are in a similar situation
We want to run mathematical algorithms
(classi

38. cation and clustering) in a complicated
architecture (distributed system)
But we are like at the time point before optimized
BLAS was developed
Chih-Jen Lin (National Taiwan Univ.) 31 / 54

39. Opportunities Lessons from past developments in one machine
Algorithms and Systems
To have technical breakthroughs for big-data
analytics, we should know both algorithms and
systems well, and consider them together
Indeed, if you are an expert on both topics,
everybody wants you now
Many machine learning Ph.D. students don't know
much about systems. But this isn't the case in the
early days of computer science
Chih-Jen Lin (National Taiwan Univ.) 32 / 54

40. Opportunities Lessons from past developments in one machine
Algorithms and Systems (Cont'd)
At that time, every numerical analyst knows
computer architecture well.
That's how they successfully developed
oating-point systems and IEEE 754/854 standard
Chih-Jen Lin (National Taiwan Univ.) 33 / 54

41. Opportunities Lessons from past developments in one machine
Example: Machine Learning Using Spark
Recently we developed a classi

42. er on Spark
Spark is an in-memory cluster-computing platform
Beyond algorithms we must take details of
Spark
Scala
into account
For example, you want to know
the dierence between mapPartitions and
map in Spark, and
the slower for loop than while loop in Scala
Chih-Jen Lin (National Taiwan Univ.) 34 / 54

43. Opportunities Lessons from past developments in one machine
Example: Machine Learning Using Spark
(Cont'd)
During our development, Spark was signi

44. cantly
upgraded from version 0.9 to 1.0. We must learn
their changes
It's like when you write a code on a computer, but
the compiler or OS is actively changed. We are in a
stage just like that.
Chih-Jen Lin (National Taiwan Univ.) 35 / 54

45. Opportunities Successful examples?
Outline
3 Opportunities
Lessons from past developments in one machine
Successful examples?
Design of big-data algorithms
Chih-Jen Lin (National Taiwan Univ.) 36 / 54

46. Opportunities Successful examples?
Example of Distributed Machine Learning
I don't think we have many successful examples yet
Here I will show one: CTR (Click Through Rate)
prediction for computational advertising
Many companies now run distributed classi

47. cation
for CTR problems
Chih-Jen Lin (National Taiwan Univ.) 37 / 54

48. Opportunities Successful examples?
Example: CTR Prediction
De

49. nition of CTR:
CTR =
# clicks
# impressions
.
A sequence of events
Not clicked Features of user
Clicked Features of user
Not clicked Features of user
A binary classi

50. cation problem.
Chih-Jen Lin (National Taiwan Univ.) 38 / 54

51. Opportunities Successful examples?
Example: CTR Prediction (Cont'd)
Chih-Jen Lin (National Taiwan Univ.) 39 / 54

52. Opportunities Design of big-data algorithms
Outline
3 Opportunities
Lessons from past developments in one machine
Successful examples?
Design of big-data algorithms
Chih-Jen Lin (National Taiwan Univ.) 40 / 54

53. Opportunities Design of big-data algorithms
Design Considerations
Generally you want to minimize the data access and
communication in a distributed environment
It's possible that
method A better than B on one computer
but
method A worse than B in distributed environments
Chih-Jen Lin (National Taiwan Univ.) 41 / 54

54. Opportunities Design of big-data algorithms
Design Considerations (Cont'd)
Example: on one computer, often we do batch
rather than online learning
Online and streaming learning may be more useful
for big-data applications
Example: very often we design synchronous parallel
algorithms
Maybe asynchronous ones are better for big data?
Chih-Jen Lin (National Taiwan Univ.) 42 / 54

55. Opportunities Design of big-data algorithms
Work
ow Issues
Data analytics is often only part of the work
ow of
a big-data application
By work
ow, I mean things from raw data to

56. nal
use of the results
Other steps may be more complicated than the
analytics step
In one-computer situation, the focus is often on the
analytics step
Chih-Jen Lin (National Taiwan Univ.) 43 / 54

57. Opportunities Design of big-data algorithms
How to Get Started?
In my opinion, we should start from applications
Applications ! programming frameworks and
algorithms ! general tools
Now almost every big-data application requires
special settings of algorithms, but I believe general
tools will be possible
Chih-Jen Lin (National Taiwan Univ.) 44 / 54

58. Discussion and conclusions
Outline
1 From data mining to big data
2 Challenges
3 Opportunities
4 Discussion and conclusions
Chih-Jen Lin (National Taiwan Univ.) 45 / 54

59. Discussion and conclusions
Risk of This Topic
It's unclear how successful we can be
Two problems:
Technology limits
Applicability limits
Chih-Jen Lin (National Taiwan Univ.) 46 / 54

60. Discussion and conclusions
Risk: Technology limits
It's possible that we cannot get satisfactory results
because of the distributed con

61. guration
Recall that parallel programming or HPC (high
performance computing) wasn't very successful in
early 90's. But there are two dierences this time
1 We are using commodity machines
2 Data become the focus
Well, every area has its limitation. The degree of
success varies
Chih-Jen Lin (National Taiwan Univ.) 47 / 54

62. Discussion and conclusions
Risk: Technology Limits (Cont'd)
Let's compare two matrix products:
Dense matrix products: very successful as the

63. nal
outcome (optimized BLAS) is much better than
what ordinary users wrote
Sparse matrix products: not as successful. My code
is about as good as those provided by Matlab
For big data analytics, it's too early to tell
We never know until we try
Chih-Jen Lin (National Taiwan Univ.) 48 / 54

64. Discussion and conclusions
Risk: Applicability Limits
What's the percentage of applications that need
big-data analytics?
Not clear. Indeed some think the percentage is
small (so they think big-data analytics is a hype)
One main reason is that you can always analyze a
random subest on one machine
But you may say this is a chicken and egg problem {
because of no available tools, so no applications??
Chih-Jen Lin (National Taiwan Univ.) 49 / 54

65. Discussion and conclusions
Risk: Applicability Limits (Cont'd)
Another problem is the mis-understanding
Until recently, few universities or companies can
access data center environments. They therefore
think those big ones (e.g., Google) are doing
big-data analytics for everything
In fact, the situation isn't like that
Chih-Jen Lin (National Taiwan Univ.) 50 / 54

66. Discussion and conclusions
Risk: Applicability Limits (Cont'd)
A quote from Dan Ariely, Big data is like teenage
sex: everyone talks about it, nobody really knows
how to do it, everyone thinks everyone else is doing
it, so everyone claims they are doing it ...
In my recent visit to a large company, their people
did say that most analytics works are still done on
one machine
Chih-Jen Lin (National Taiwan Univ.) 51 / 54

67. Discussion and conclusions
Risk: Applicability Limits (Cont'd)
A quote from Dan Ariely, Big data is like teenage
sex: everyone talks about it, nobody really knows
how to do it, everyone thinks everyone else is doing
it, so everyone claims they are doing it ...
In my recent visit to a large company, their people
did say that most analytics works are still done on
one machine
Chih-Jen Lin (National Taiwan Univ.) 51 / 54

68. Discussion and conclusions
Risk: Applicability Limits (Cont'd)
A quote from Dan Ariely, Big data is like teenage
sex: everyone talks about it, nobody really knows
how to do it, everyone thinks everyone else is doing
it, so everyone claims they are doing it ...
In my recent visit to a large company, their people
did say that most analytics works are still done on
one machine
Chih-Jen Lin (National Taiwan Univ.) 51 / 54

69. Discussion and conclusions
Risk: Applicability Limits (Cont'd)
A quote from Dan Ariely, Big data is like teenage
sex: everyone talks about it, nobody really knows
how to do it, everyone thinks everyone else is doing
it, so everyone claims they are doing it ...
In my recent visit to a large company, their people
did say that most analytics works are still done on
one machine
Chih-Jen Lin (National Taiwan Univ.) 51 / 54

70. Discussion and conclusions
Open-source Developments
Open-source developments are very important for
big data analytics
How it works:
The company must do an application X. They
consider an open-source tool Y. But Y is not
enough for X. Then their engineers improve Y and
submit pull requests
Through this process, core developers of a project
are formed. They are from various companies
Chih-Jen Lin (National Taiwan Univ.) 52 / 54

71. Discussion and conclusions
Open-source Developments (Cont'd)
For Taiwanese data-science companies, I think we
should actively participate in such developments
Indeed industry rather than schools are in a better
position to do this
Chih-Jen Lin (National Taiwan Univ.) 53 / 54

72. Discussion and conclusions
Conclusions
Big-data analytics is in its infancy
It's challenging to development algorithms and tools
in a distributed environment
To start, we should take both algorithms and
systems into consideration
Hopefully we will get some breakthroughs in the
near future
Chih-Jen Lin (National Taiwan Univ.) 54 / 54