It's a Streaming World! Reasoning upon Rapidly Changing Information (Milano, 14-2-2014)

Stream Reasoning
http://streamreasoning.org/

It's a Streaming World!
Reasoning upon Rapidly
Changing Information
Emanuele Della Valle

emanuele.dellavalle@polimi.it

Agenda
Stream Reasoning in a nutshell
A sample of Stream Reasoning investigation: IMaRS
Conclusions

Emanuele Della Valle - http://streamreasoning.org/

3

Agenda
 It's a streaming world
 Requirements
 State-of-the-art solutions
 Stream Reasoning research questions
 Example of findings

Conclusions


4

It's a streaming World!

1/2

 Off-shore oil operations

 Smart Cities

 Global Contact Center

 Social networks

 Generate data streams!


5

It's a streaming World!

2/2

 What is the expected time to failure when that
turbine's barring starts to vibrate as
detected in the last 10 minutes?
 Is public transportation
where the people are?
 Who are the best available agents to
route all these unexpected contacts
about the tariff plan launched yesterday?
 Who is driving the discussion
about the top 10 emerging topics ?

E. Della Valle, S. Ceri, F. van Harmelen, D. Fensel It's a Streaming World! Reasoning
upon Rapidly Changing Information. IEEE Intelligent Systems 24(6): 83-89 (2009)


6

Requirements

1/8

A system able to answer those queries must be able to
 handle massive datasets

• A typical oil production platform is equipped
with about 400.000 sensors
• Telecom data is the most pervasive data
source in urban are, in Milano there are
1.8 million mobile users
• A global contact centre of a Telecom
operator counts 500 millions of clients
• Facebook alone has 1.1 billion
of active users


7

Requirements

2/8

 process data streams on the fly

• The sensors on typical oil production
platform generates 10,000 observations
per minute with peaks of 100,000 o/m
• The mobile users in Milano generates
20,000 call/sms/data connections
per minute with peaks of 80,000 c/m
• A global contact centre receives
10,000 contacts per minute with
peaks of 30,000 c/m
• Facebook, as of May 2013, observes
3 millions "I like" per minute


8

Requirements

3/8

 cope with heterogeneous dataset

• The sensors on typical oil production
have been deployed over 10 years
by 10s of different producers
• Tens of data sources are normally
needed to make sense of an urban
phenomena
• A global contact centre consists in 100s
of offices owned by different subsidiary
companies engaged yearly
• Each social network has its own
data model, APIs, …


9

Requirements

4/8

 cope with incomplete data

• 10s of sensors and networking links
broke down daily
• Coverage is incomplete

• Only standard cases are covered by
fully machine processable data records
100s of contacts per minute are
manage ad-hoc

• Conversations happen outside the
social networks, too!


10

Requirements

5/8

 cope with noisy data

• Sensor out-of-operating range

• Faulty sensors

• Agents misunderstand, get tired, …

•

Irony, sarcasm, …


11

Requirements

6/8

 provide reactive answers

• detection of dangerous situations
must occur within minutes
• recommendations to citizens must
be performed in few seconds
• routing a contact through each step of
the decision tree must take less than a
second
• Search autocompleting may need
to be updated every few minutes


12

Requirements

7/8

 support fine-grained information access
• Identify a turbine among thousands

• Locate a bus among thousands

• Contact an agent among thousands

• Identify an opinion maker among
thousands of influencers for a topic


13

Requirements

8/8

 integrate complex domain models of
• operational and control process

• various city aspects

• contact management, contract types,
agent skills, contactor profiles, …
• topics, user profiles, …


14

Requirements (wrap up)
 handle massive datasets
 process data streams on the fly
 cope with heterogeneous dataset
 cope with incomplete data
 cope with noisy data

 provide reactive answers
 support fine-grained information access
 integrate complex domain models


15

What are data streams anyway?
 Formally:

• Data streams are unbounded sequences of time-varying data
elements

time

 Less formally:

• an (almost) “continuous” flow of information

 Assumption

• recent information is more relevant as it describes the
current state of a dynamic system


16

The continuous nature of streams
 The nature of streams requires a
paradigmatic change*
• from persistent data

– to be stored and queried on demand
– a.k.a. one time semantics

• to transient data

– to be consumed on the fly by continuous queries
– a.k.a. continuous semantics

* This paradigmatic change first arose in DB community in the late '90s

17

Continuous Semantics
 Continuous queries registered over streams that
are observed trough windows
window

Dynamic
System

input streams

Registered
Continuous
Query


streams of answer

18

DSMS/CEP State of the Art
 Gianpaolo Cugola, Alessandro Margara: Processing
flows of information: From data stream to complex event
processing. ACM Comput. Surv. 44(3): 15 (2012)
 Content

• Type of models compared
–
–
–
–

Functional and processing
Deployment and interactions
Data, Time, and Rule
Language

• # of systems surveyed:
– Academic: 24
– Industrial: 9
– Total: 33

• To learn more:

– http://home.dei.polimi.it/margara/papers/survey.pdf


19

Existing solutions

Requirement
massive datasets
data streams
heterogeneous dataset
incomplete data
noisy data
reactive answers
fine-grained information access
complex domain models


DSMS
CEP

20

Ontology Based Data Access
 Given ontology O and query Q, use O to rewrite Q
as Q’ so that, for any set of ground facts A contained in
multiple databases:
• answer(Q,O,A) = answer(Q’,,A)

– The answer of the query Q using the ontology O for any set of
ground facts A is equal to answer of a query Q’ without
considering the ontology O

 Use mapping M to map Q’ to multiple SQL queries to the
various databases

O
Q

Rewrite

M
Q’

Map

SQL


A

answer
21

Existing solutions

Requirement

DSMS
CEP

OBDA

massive datasets
data streams
heterogeneous dataset
incomplete data
noisy data
reactive answers
fine-grained information access
complex domain models


22

Research Question
 is it possible to make sense in real time of
multiple, heterogeneous, gigantic and inevitably noisy
and incomplete data streams in order to support the
decision process of extremely large numbers of
concurrent user?


23

Stream Reasoning feasibility (intuition)
 Many relevant reasoning methods are not able to
deal with high frequency data streams
 However, trade-off exists between the complexity of the
reasoning method and the frequency of the data stream
1 Hz

NEXPTIME

Abstraction
PTIME

Selection
Interpretation

AC0

Complexity

DL

Reasoning

DL-Lite
Semantic Streams
Raw Stream Processing
Complexity vs. Dynamics

Querying
Re-writing
104 Hz

Change Frequency

Heiner Stuckenschmidt, Stefano Ceri, Emanuele Della Valle, Frank van Harmelen:
Towards Expressive Stream Reasoning. Proceedings of the Dagstuhl Seminar on
Semantic Aspects of Sensor Networks, 2010.

24

Sub-research questions
 Is it possible model at semantic level
heterogeneous data streams, continuous queries, and
continuous reasoning tasks?
 Does the ordered nature of data streams and the
possibility to forget old enough information allow to
optimize continuous querying and continuous
reasoning tasks so to provide reactive answers to
large number of concurrent users without forsaking
correctness or completeness?
 Can Semantic Web and Machine Learning technologies be
jointly employed to cope with the noisy and incomplete
nature of data streams?

 Are there practical cases where processing data stream
at semantic level is the best choice?

25

Findings

Model data stream at semantic level

1/2

 State of the art: RDF

• It allows to make statements about resources in the form
of subject-predicate-object expressions
– In RDF terminology triples
– E.g.
@BarakObama posts "Four more years"
subject

predicate

object

• A collection of RDF statements represents a labelled,
directed graph
– In RDF terminology a graph
– E.g., the tweet above by Barak Obama is connected to
-

800,000+ twitter user profiles via retweets
300,000+ twitter user profiles favorite
…

 Contribution: RDF stream

26

Findings

Model data stream at semantic level

2/2

 State of the art: RDF
 Contribution: RDF Stream

• Unbound sequence of time-varying triples
• each represented by a pair made of an RDF triple and its
timestamp
…
@BarakObama

2012

posts "Four more years",

8:16PM 6 Nov 2012

@Alice

posts "RT: Four more years",

8:17PM 6 Nov

subject

predicate

…
object

timestamp

D.F. Barbieri, D. Braga, S. Ceri, E. Della Valle, M. Grossniklaus: Querying RDF streams
with C-SPARQL. SIGMOD Record 39(1): 20-26 (2010)

27

Findings

Adding continuous semantics to SPARQL
Who are the opinion makers? i.e., the users who are
likely to influence the behavior their followers
REGISTER STREAM OpinionMakers COMPUTED EVERY 5m AS
CONSTRUCT { ?opinionMaker sd:about ?resource }

FROM STREAM <http://…>

[RANGE 30m STEP 5m]

WHERE {
?opinionMaker ?opinion ?resource .
?follower sioc:follows ?opinionMaker.

?follower ?opinion ?resource.
FILTER ( cs:timestamp(?follower) >
cs:timestamp(?opinionMaker) )
}

HAVING ( COUNT(DISTINCT ?follower) > 3 )

28

Findings

Adding continuous semantics to SPARQL
Who are the opinion makers? i.e., the users who are
Query registration
RDF Stream added
likely to influence the behavior their followers as
(for continuous execution)

new ouput format

REGISTER STREAM OpinionMakers COMPUTED EVERY 5m AS
CONSTRUCT { ?opinionMaker sd:about ?resource } clause
FROM STREAM

FROM STREAM <http://…>

[RANGE 30m STEP 5m]

WHERE {

WINDOW

?opinionMaker ?opinion ?resource .
?follower sioc:follows ?opinionMaker.Builtin to access

?follower ?opinion ?resource.

timestamps

FILTER ( cs:timestamp(?follower) >
cs:timestamp(?opinionMaker) )
}

HAVING ( COUNT(DISTINCT ?follower) > 3 )

29

Findings

Cope with the noisy and incomplete data
 Noise is reduced using DSMS techniques
 Deductive stream reasoning copes with
incompleteness deducing implicit facts

 Inductive stream reasoning copes "irrepairable"
incompleteness inducing missing facts

D.F. Barbieri, D. Braga, S. Ceri, E. Della Valle, Y. Huang, V. Tresp, A. Rettinger,
H. Wermser: Deductive and Inductive Stream Reasoning for Semantic Social Media
Analytics. IEEE Intelligent Systems 25(6): 32-41 (2010)

30

Findings

Practical cases
 10+ deployments in Sensor Networks,
Social media analytics and Cloud Computing, e.g.
BOTTARI

City Data Fusion

Winner of Semantic Web
Challenge 2011

Winner of IBM
faculty award 2013

M. Balduini, I. Celino, D. Dell’Aglio, E. Della Valle, Y. Huang, T. Lee, S.-H. Kim, V. Tresp:
BOTTARI: An augmented reality mobile application to deliver personalized and location-based
recommendations by continuous analysis of social media streams. J. Web Sem. 16: 33-41 (2012)

31

Findings

Optimize for reactive answers
 C-SPARQL time window-based selections outperform
SPARQL filter-based selections
 Incremental Reasoning on RDF streams (IMaRS):
new reasoning algorithm optimized for reactive query
answering

D.F. Barbieri, D. Braga, S.Ceri, E. Della Valle, M. Grossniklaus: Incremental Reasoning on
Streams and Rich Background Knowledge. ESWC (1) 2010: 1-15
D. Dell'Aglio, E. Della Valle: Incremental Reasoning on RDF Streams. In A.Harth, K.Hose,
R.Schenkel (Eds.) Linked Data Management, CRC Press 2014, ISBN 9781466582408

32

Agenda


It's a streaming world



Requirements



State-of-the-art solutions



Stream Reasoning research questions



Affirmative answers collected in 5 years of investigations



Problem statement



State-of-the art solutions



Proposed solution



Experimental results

Conclusions

33

The problem
 Reasoning upon rapidly changing information
 Example
 Context
Social Media Analytics: a set of methods and technologies for
collecting, monitoring, analyzing, summarizing, and visualizing
media generated and disseminated in a conversational, and
distributed mode among online communities



 Real query
Which have been the most discussed posts in the last hour?



 Simplified version
How many posts have discussed A in the last 40 min?



 Data
A

B

C

A, B, and C represent posts
A
B means B discusses A
A
C means C indirectly discusses A
NOTE: discusses is transitive property


34

The problem
Enter
Exit
Time window window

Explicitly in
window

10:00

AB

A

BC

A

B

10:20
10:30

C

A

C

C

A

C

C

A

C

E

AE
EC

A
A

10:40

AB

10:50

BC

11:00

AE

A

B
E
B
E
B
E

C

A

C

C

A

# Exp.

# Inf.

1

B

10:10

Inferred in
window

A

C

E
C


SUM

1

1

1

2

2

1

3

2

1

3

1

1

2

1

1

2

0

0

0
35

Is it an hard problem?
 In the database community it is known as the
incremental view maintenance problem
 The state of the art solution is the DRed algorithm

• Overestimation of deletion: Overestimates deletions by
computing all direct consequences of a deletion.
A

E
B

C

A

C

• Rederivation: Prunes those estimated deletions for which
alternative derivations (via some other facts in the program)
exist.
A

E
B

C

A

C

• Insertion: Adds the new derivations that are consequences of
insertions

 Deletion are twice as expensive as insertions

36

ms

rematerialize the view
after each update

Break-even point

1000
incremental
view maintenance
100
In a normal DB incremental
view maintenance is convenient
in most of the cases:
• Lot of inserts, few deletions
• Small % of deletions w.r.t.
the size of the DB

10

% of deletions w.r.t
the size of the DB

0
0%

2%

4%

6%

8%


10%

12%

14%
37

ms

after each update

Break-even point

1000
incremental
view maintenance
100
In a streaming setting
• The number of inserts are equals
to the number of deletions
• the data exiting the windows
causes large % of deletions w.r.t.
% of deletions w.r.t the
the size of the window
content of the window

10

0
0%

2%

4%

6%

8%


10%

12%

14%
38

ms

What is the target behaviour?
after each update

Break-even point

1000
incremental
view maintenance
100
Target behaviour

10
% of deletions w.r.t the
content of the window

0
0%

2%

4%

6%

8%


10%

12%

14%
39

Observation
 DRed works with random insertions and deletions
 In a streaming setting, when a triple enters the window,
given the size of the window, the reasoner knows already
when it will be deleted!
 E.g.,

• if the window is 40 minutes
long, and,
• it is 10:00, the triple(s)
entering now
• will exit on 10:40.

 Conclusion

• deletions are predictable

Enter
Exit
Time window window
10:00

AßB

A

10:10

BßC

A

B

10:20

AßE

10:30

EßC

Infer
win

B
C

A

C

A

C

A

C

A

C

A

E
A
A

10:40

AßB

10:50

BßC

11:00


Explicitly in
window

AßE

A

B
E
B
E
B
E

A
E

C

40

Optimized Stream Reasoning (IMaRS)
 Idea:

• add an expiration time to each triple and
• use an hash table to index triples by their expiration time

 The algorithm

1. deletes expired triples
2. Adds the new derivations that are consequences of insertions
annotating each inferred triple with an expiration time
(the min of those of the triple it is derived from), and
3. when multiple derivations occur, for each multiple
derivation, it keeps the max expiration time.


41

The problem
Enter
Exit
Time window window
10:00

AB

Explicitly in
window

Inferred in
window

10:40

A

BC

A

B

# Inf.

1

B

SUM

1

10:40

10:40 10:50

10:10

# Exp.

C

A

C

1

1

2

The min of the two
expiration times


42

The problem
Enter
Exit
Time window window
10:00

AB

Explicitly in
window

Inferred in
window

10:40

A

10:20

A

AE

11:00 E
10:40 10:50

B

SUM

1

10:40

BC

A

B

# Inf.

1

B
10:40 10:50

10:10

# Exp.

C

C

A

A

C
10:40

C

1

1

2

2

1

3

AC remains
unchanged


43

The problem
Enter
Exit
Time window window
10:00

AB

Explicitly in
window

Inferred in
window

10:40

A

B

# Inf.

1

B

SUM

1

10:40

10:40 10:50

BC

A

10:20

AE

11:00 E
10:40 10:50

10:30

EC

10:10

# Exp.

C

C
B
11:00 E 11:10
10:40 10:50
C
A
B
A

A

A
A

C

1

1

2

10:40

2

1

3

11:00

2

1

3

C
C

AC is inferred with a
longer lasting expiration
time, the max of the
expiration time is used

44

The problem
Enter
Exit
Time window window
10:00

Explicitly in
window

Inferred in
window

10:40

AB

A
A

10:20

AE

11:00 E
10:40 10:50

C

10:30

EC

C
B
11:00 E 11:10
10:40 10:50
C
A
B
11:00 E 11:10
10:50
A
B
C
A

AB

SUM

1

10:40

BC

10:40

B

# Inf.

1

B
10:40 10:50

10:10

# Exp.

A

A
A
A

C

1

1

2

10:40

2

1

3

11:00

2

1

3

11:00

2

1

3

C
C

C

This deletion causes no effect

45

The problem
Enter
Exit
Time window window
10:00

Explicitly in
window

Inferred in
window

10:40

AB

A

B

A

10:20

AE

11:00 E
10:40 10:50

C

10:30

EC

C
B
11:00 E 11:10
10:40 10:50
C
A
B
11:00 E 11:10
10:50
A
B
C
11:00 E 11:10
A
C
A

AB

10:50

BC

SUM

1

10:40

BC

10:40

# Inf.

1

B
10:40 10:50

10:10

# Exp.

A

A
A
A
A

C

1

1

2

10:40

2

1

3

11:00

2

1

3

11:00

2

1

3

11:00

2

1

3

C
C

C
C

This deletion causes no effect

46

The problem
Enter
Exit
Time window window
10:00

Explicitly in
window

Inferred in
window

10:40

AB

A

B

A

10:20

AE

11:00 E
10:40 10:50

C

10:30

EC

C
B
11:00 E 11:10
10:40 10:50
C
A
B
11:00 E 11:10
10:50
A
B
C
11:00 E 11:10
A
C
E
C
A

AB

10:50

BC

11:00

AE

SUM

1

10:40

BC

10:40

# Inf.

1

B
10:40 10:50

10:10

# Exp.


A

A
A
A
A

C

1

1

2

10:40

2

1

3

11:00

2

1

3

C
C

11:00

C

11:00

C

Deletion of inferred triples
3
is just 2 look up1
a

2

1

3

0

0

0
47



Re-materialize after each window slide



Use DRed



IMaRS

% of deletions w.r.t. the content of the window

48

Implication of the experimental results
 comparison of the average time needed to answer
a C-SPARQL query, when 2% of the content exits the
window each time it slides, using
• A backward reasoner on the window content
• DRed + standard SPARQL on the materialization
• IMaRS + standard SPARQL on the materialization

Backward

DRed +SPARQL


IMaRS + SPARQL

49

Agenda


It's a streaming world



Requirements



State-of-the-art solutions



Stream Reasoning research questions



Affirmative answers collected in 5 years of investigations



Problem statement



State-of-the art solutions



Proposed solution




Conclusions

50

Scholarly impact
BOTTARI
140

Inductive SR
120

# of citations

IMaRS
100

C-SPARQL

80
60
40

SR Vision
20
0
2009

2010

2011

year

2012


2013
51

Recent Work

Benchmarks are needed
 Stream Reasoning benchmarks are appearing

• SRBench Ying Zhang, Minh-Duc Pham, Óscar Corcho, JeanPaul Calbimonte
– http://www.w3.org/wiki/SRBench

• LSBench: Danh Le Phuoc, Minh Dao-Tran, Minh-Duc Pham,
Peter A. Boncz, Thomas Eiter, Michael Fink
– http://code.google.com/p/lsbench/

 What shall we evaluate?

• PeCK methodology: Sys Properties -> Challenges -> KPI

 How?

• Maximise throughput is OK, but do not forsake correctness!

T. Scharrenbach, J. Urbani, A. Margara, E. Della Valle, A. Bernstein: Seven Commandments
for Benchmarking Semantic Flow Processing Systems. ESWC 2013: 305-319
D. Dell'Aglio, J.-P. Calbimonte, M. Balduini, Ó. Corcho, E. Della Valle: On Correctness in RDF
Stream Processor Benchmarking. ISWC (2) 2013: 326-342

52

Recent Work

Supporting industrial take up
 W3C's RDF stream processing Community Group
• http://www.w3.org/community/rsp/
• Goals
–
–
–
–
–

collect use cases
cense approaches
define RDF stream models
define SPARQL extension for RDF stream processing
define RESTful interfaces

• 52 members from 24 organization (6 industries)
• Biweekly phone calls since September 2013


53

Next step

Generalizing from time to any ordering
 Observation: order reflects recency, relevance, …
Continuous top-k Q/A

Order-aware
reasoning

Top-k Q/A

Top-k Reasoning

Recency

DSMS/CEP

Stream reasoning

Indexes

Traditional solutions

Scalable reasoning

No

Yes

Types of orders

Combinations
Relevance
…

Reasoning

Emanuele Della Valle, Stefan Schlobach, Markus Krötzsch, Alessandro Bozzon, Stefano Ceri,
Ian Horrocks: Order matters! Harnessing a world of orderings for reasoning over
massive data. Semantic Web 4(2): 219-231 (2013)
Sara Magliacane, Alessandro Bozzon, Emanuele Della Valle: Efficient Execution of Top-K
SPARQL Queries. International Semantic Web Conference (1) 2012: 344-360

54

Credits
Politecnico di Milano’s colleagues
 Vision: Prof. S. Ceri
 C-SPARQL: M. Balduini, D. Barbieri, D. Braga, S. Ceri and M.
Grossniklaus
 Benchmarking: M. Balduini, D. Dell'Aglio, and A. Margara
 SPARQL-RANK: S. Magliacane and A. Bozzon
People I directly worked with on the topic
 CEFRIEL: I. Celino
 Saltlux: T. Lee and S.-H. Kim
 SIEMENS: prof. V. Tresp and Y. Huang
 U-Innsbruck: prof. D. Fensel
 U-Oxford: prof. I. Horrocks and M. Krötzsch
 UP-Madrid: Ó. Corcho and J.-P. Calbimonte
 U-Mannheim: prof. H. Stuckenschmidt
 U-Zurich: prof. A. Bernstein and T. Scharrenbach
 VA-Amsterdam: prof. F. van Harmelen, S. Schlobach and J. Urbani
The broader research community that showed interest in stream
reasoning

55

Stream Reasoning
http://streamreasoning.org/

Thank you for paying
attention!
Any question?
Emanuele Della Valle
emanuele.dellavalle@polimi.it

It's a Streaming World! Reasoning upon Rapidly Changing Information (Milano, 14-2-2014)

Recommandé

Recommandé

Contenu connexe

Similaire à It's a Streaming World! Reasoning upon Rapidly Changing Information (Milano, 14-2-2014)

Similaire à It's a Streaming World! Reasoning upon Rapidly Changing Information (Milano, 14-2-2014) (20)

Plus de Emanuele Della Valle

Plus de Emanuele Della Valle (20)

Dernier

Dernier (20)

It's a Streaming World! Reasoning upon Rapidly Changing Information (Milano, 14-2-2014)

Notes de l'éditeur