Project name - Mining Minds

Service Curation Layer
Reasoning and
Prediction
KHU Member:
Rahman Ali
Soongsil Member:
H.K. Park
Introduction
/
Recommendations
Knowledge
Reasoning (definition)
An
application
whose
computational function is to
generate conclusions from
available knowledge using
logical techniques
Reasoning
Features
A
B
C
D
Human expert take care of all the data
Analyze whole data based on expertise
Features
Expert-based
Reasoning
Automatic
Reasoning
Extract some heuristic knowledge
Recommend Services
Limitations
A
B
C
2
A
B
C
D
E
F
Automatic learning algorithms take care of all the data
Analyze and learn whole data based on learning algorithms
Can analyze and learn hundred of attributes quickly
Create heuristic knowledge bases
Cost effective in terms of resources, time and efforts
Automatically recommend services
Data drastically increases with hundreds of attributes
Limitations
Human brain cannot process all attributes at the same time
Extremely expensive in terms of resources, time and efforts
Recommendations
A
B
Bottleneck: needs huge data
Difficult to implement in realistic environments
Taxonomy of automatic reasoning systems
/
3
Reasoning Systems
Constraint solvers
Theorem provers
Uses constraint
programming
Uses automatic
reasoning
techniques for
proofs of
mathematical
theorems
Used for optimal
scheduling, design
efficient integrated
circuits
Used for
verification of the
correctness of ICs,
software
programs,
engineering
designs
Logic programs
Uses generalpurpose logic
programming
language
Used for
application across
many disciplines
Rule engines
Uses discrete rules
Used in decisionmaking systems in
different domains
Deductive
classifier/
Machine learning
systems/
reasoning
inductive reasoning
Case-based
reasoning
systems
Procedural
reasoning
systems
Uses set of classes,
subclasses, and
relations among
the classes
(ontology)
Uses observed
data/training
examples to learn
hypothesis
Uses similarities to
other problems for
which known
solutions already
exist
Uses plans which
represent a course
of action for
achievement of a
given goal
Used for finding
relationships
between objects
Used for predicting
decision of new
example cases in
different domains
Used in industrial
manufacture,
agriculture,
medicine, law etc.
for making
decision
Used in control,
management,
monitoring and
fault detection
systems
Mining Minds
[Schalkoff, Robert 2011]
[Moses2003]
Reasoning Systems in Mining Minds
• Why we chose?
Mining Minds
Features Exploited in Mining Minds
Deals with a varity/heterogeneous
type of data from multiple data
sources
A large number of attributes from
multiple data sources are taken into
account for service creation
/
Implicit/explicit
relationship and context
& behavior modeling
Predictive modeling
Appropriate Solution
Ontological or
Deductive
Reasoning
ML-based
Reasoning
and CBR
Personalized services
& recommendations
Targets personalized healthcare
services where general healthcare
guidelines can be used
Rule-based
Reasoning
4
Reasoning systems in Mining Minds (Cont …)
/
• Characteristics and limitations
Ontological
reasoning
Machine-learning
reasoning
Rule-based
reasoning
Case-based
reasoning
Characteristics
- knowledge evolution
- easy to change
- explicit relationships
Characteristics
- predictive modeling
- minimum user involvement
- automatic rules creation
- reduce cost & users efforts
Characteristics
- validation and explanation
- easy to understand
- IF-THEN rules, similar to
human reasoning
Characteristics
- easier and fast development
- no need for explicit domain
model
- no need of explicit knowledge
Limitations
- Explicit model is needed in
advance
- need clear understanding of
the domain
Limitations
- need large training examples
- difficulty in understanding
- less preferred for critical
environments
Limitations
- need explicit domain model
- need experts
- slower development
- hard to maintenance
Limitations
- no validation and explanation
- understanding is difficult in
term of no explanation
5
Reasoning Levels in Mining Minds
• Reasoning Levels
• High-level
• Intermediate-level
• Low-level
Finding Service-level Personalized Recommendations
ML and Expert-based reasoning
Service Curation Team
Finding Context and User Behavior
Ontological reasoning
Information Curation Team
Finding Implicit/explicit Relationships
Scalable ontological reasoning over HDFS
SSU Team
/
6
Motivations
/
ML-based reasoning
Healthcare data increase drastically to gigabytes and
terabytes with hundreds of attributes
ML-based
reasoning &
prediction
Human brain cannot examine this much variables at a time
to understand and learn
Predictive modeling algorithms can automatically analyze
and learn this data, and perform reasoning and prediction
with reduced cost and time
7
Motivations
/
Ontological reasoning
SPARK framework: High-speed inmemory analytics over hadoop
Distributed ontological
reasoning
SSU
Quickly find implicit/explicit relationship
among the entities of personalized Big
data
Ontological
reasoning
Context modeling
Centralized ontological
reasoning
Info. Curation
Behavior modeling
8
/
ML-Based Reasoning
9
Related Work
Study/Project/Model/
Framework/System
Domain
(Zazzi2010)-Project (2006Treatment
08)
- HIV
-Study
(Husain2010)
- Model
(Yuan2014)
-Framework
(Drăgulescu2007)
- Tool as a proof of
research
[Pandey2012]
- Tool as a proof of
research
/
Service
Key Features
Methodology
Technique
Predicts the
• Use several
efficiency of possible
prediction engines
Anti-HIV drug
• combines their
combinations
results
Predicts suitable
• Uses CBR and RBR
Wellness
personalized
for the
Recommendation wellness therapy for
complementary and
individual
alternative medicine
• Uses sensor fusion
Wellness Therapy Predicts therapy to
in a smart home
Recommendation the elderly people
environment for
data
• Logical, Statistical
and ANN are used to
Diagnosis
Predicts hepatitis
first detect hepatitis
- hepatitis
infection
type
• Then treat type B
and C
• Uses Apriori
algorithm
Predicts diabetes
Diagnosis
• Finds type of
types
diabetes based on
the symptoms
Use
Classification
technique
Limitations
• Lack of proper integration between
multiple engines
• Lack of appropriate methodology for
Hybrid CBR, RBR
abstraction of data
• Lack of knowledge maintenance
CBR, RBR
• No feedback system
Logical,
Statistical and
ANN
• Weak integration of the hybrid techniques
• Only works in a sequential way
Apriori
algorithm
• Use only single learning algorithm
• Prediction accuracy is low
10
Limitations of Existing Work
/
• No support for reasoning guided by automatically selected reasoning
algorithm
• Usually, based on classical reasoning mechanism- single reasoning algorithm is
used
• No granularity of recommendations, either supports reasoning or prediction
• No appropriate mechanism for algorithm parameter tuning during learning
11
Limitations of Existing Work
/
• Characteristics, cause, effects, need and objective
Characteristics
Multiple Heterogeneous
data sources
Cause
Incomplete, noisy and
Inconsistent Service Request
Query parsing, validation
and verification
Effect
Need
Pattern &rules matching is difficult and
results in either misleading or no results
Objective
Fast and accurate
retrieval of rules
Ensemble and hybrid
Reasoning
Mining Minds
Diverse data sources
Varity of data
Classical reasoning may not produce
accurate recommendation
Improve
prediction
accuracy
Reasoning and prediction
Targets healthcare
services, uses guidelines
Guidelines and multiple
knowledge
Single-level reasoning is not enough to
get enhanced recommendations
Enhance
recommendations
12
Intra-core Communication View
/
Service Curation Layer
User satisfaction/
dissatisfaction
Filtered & explained
recommendations
Knowledge
Maintenance
Engine
Feedback Analysis
Recommendation
Manager
n Recommendations
Knowledge Base
Reasoning and Prediction
13
Inter-cores Communication View
/
Service Curation Layer
User satisfaction/
dissatisifaction
Filtered and explained final
recommendation
Knowledge
Maintenance
Engine
Feedback Analysis
Recommendation
Manager
UI/UX
Knowledge Base
Visualization
HDFS Data Access
Interface
Intermediate
Data
Reasoning and Prediction
Structured Data (behavior,
features)
n Recommendations
14
Progress
/
2nd MM Platform Workshop
1st MM Platform Workshop
Features and Uniqueness
•
•
•
•
Has support for query parser
Has support for hybrid and ensemble learning
Has support for predictive analysis as support of reasoning
Has clearly defined interfaces
15
Architecture (Abstract View)
/
16
UI/UX Authoring Tool
Reasoner and Predictor
Predictor
HDFS Data Access
Interface
Behavior and
Context Report
Intermediate
Data
Report
Analyzer
Predictions
Generator
Predictive
Modeler
Prediction Models
Request
Reasoner
Request
Validation &
Verification
Knowledge
Maintenance Engine
Predicted Patterns
+
Top n Reasoning Outcomes
Knowledge
(Rules,
Cases)
Service
Query
Parser
Knowledge
Base
Recommendation
Manager
Outcomes
Aggregator
Inference Engine
Inference
Generator
http://www.socscistatistics.com/tests/pearson/Default2.aspx
Architecture (Detailed View)
/
17
UI/UX Authoring Tool
Reasoner and Predictor
Predictor
HDFS Data Access
Interface
Behavior and
Context Report
Intermediate
Data
Report
Analyzer
Predictive
Modeler
Criterion Selector
Model Builder
Correlation
Extractor
Model Evaluator
Predictions
Prediction Models
Request
Predicted Patterns
+
Top n Reasoning Outcomes
Reasoner
Request
Validation &
Verification
Knowledge
Maintenance Engine
Predictions
Generator
Knowledge
(Rules,
Cases)
Service
Query
Parser
Inference Engine
Ensemble/hybrid Reasoner
Outcomes
Aggregator
n Results
Rules/Pattern Matcher
Knowledge
Base
Conflict Resolver
Inference
Generator
Recommendation
Manager
Architecture (Functional View)
/
Reasoner and Predictor
Behavior Analysis
Frequent pattern mining
Pattern analysis
Correlation coefficient
calculator
▶
▶
▶
Service Query Parser
▶
▶
▶
Predictor
Report
Analyzer
Predictive
Modeler
Criterion Selector
Model Builder
Correlation
Extractor
▶
▶
Predictions
Generator
Model Evaluator
Query Semantic Processing
Query Interpretation
Query Transformation
Prediction Generation
Statistical method
Machine learning
methods
Predictive Modeling
Prediction Models
▶
▶
▶
Linear Regression modeling
Naïve Bayes modeling
Logistic regression modeling
Reasoner
Ensemble Selector
Rules/Pattern Matcher
▶
▶
▶
Rules-based Matching
Attributes-value Matching
Facts Feed-forward Matching
Conflict Resolver
▶
Domain-based Prioritization
▶
Confidence Factor
▶
Recency Rule
Request
Validation &
Verification
Service
Query
Parser
Inference Engine
Ensemble/hybrid Reasoner
Outcomes
Aggregator
Rules/Pattern Matcher
Knowledge
Base
Conflict Resolver
Inference
Generator
▶
▶
▶
Semi-automatic approach
Service-based selection
Guideline-based selection
Combiner and Inference Generator
▶
▶
▶
Algebraic combination
Majority voting
Weighted majority voting
18
Architecture (Scenario)
5
M
Irregular Overeaten
No
5.1
178.9
6
M
Regular Overeaten
No
4.5
178.1
7
M
Irregular Overeaten
No
5.3
179.2
Yes
5.7
180.8
HDFS Data Access
M
Regular Balance
Interface
8
Reasoner and Predictor 6
Predictor
Criterion Selector
0
0
1
2
3
4 5
Week
6
7
8
9
0= Imbalance; 1= Balance; 2= Overeaten
8 Weeks Weight Trend
Y = 7.625x + 127.85
230
Predictions
Generator
19
1
5
Predictive
Predictor
Modeler
Report
Analyzer
Behavior and
Context Report
Intermediate
Data
UI/UX Authoring Tool
4
2
Weight in lbs
Weight
120.2
167.1
132.2
160.8
Y = 0.0833x + 0.5
3
Diet Bahaviour
1
Week Gender Exercise
Diet
Disease Height
1
M
Regular Imbalance No
5.9
2
M
Regular Balance
Yes
5.5
3
M Irregular Balance
No
5.3
4
M
Cycling Imbalance Yes
5.7
/
8 Weeks Diet Trend
Model Builder
Predictions
180
130
80
0 1 2 3 4 5 6 7 8 9
Week
Request
Correlation
Extractor
User: John
Exercise: Irregular
Food Intake: Overeaten
Health condition: Ok
Height: 175.5
Weight= 10Kg
Evaluate Goal
Knowledge
Maintenance
Engine
Knowledge
(Rules,
Cases)
1
1
Request
Validation &
Verification
Model Evaluator
User: John
Exercise: Irregular
Food Intake: Overeaten
Health condition: Ok
Height: 175.5
Weight= 10Kg
Evaluate Goal
Service
Query
Parser
Knowledge
Base
7
Prediction Models
Reasoner
Predicted Patterns
+
Top n Reasoning Outcomes
2
Recommendation
Manager
Inference Engine
3
Outcomes
Reasoner
Top
n
Outcomes
Aggregator
Ensemble/hybrid Reasoner
1.
Lose
Weight
Reasoner
2. Maintain Weight
Rules/Pattern Matcher 3. Option n
Inference
Generator
Conflict Resolver
http://www.socscistatistics.com/tests/pearson/Default2.aspx
Reasoning
Rule-based reasoning
/
Predictions
20
4
Inference
Recommendation
Generator
Manager
Reasoner
• Input:
• Input query
• Rules
User: John
Exercise: Irregular
Food Intake: Overeaten
Health condition: Ok
Height: 175.5
Weight= 10Kg
Evaluate Goal
+
Reasoner
• Methods:
•
•
•
•
•
•
validateVerify(query)
parseQuery(query)
selectReasoningMethod(query, service)
patternMatching(query, rules)
resolveConflict()
generateResults()
• Output:
• Reasoner Outcomes
Outcome 1: LoseWeight (0.75)
Outcome 2: GainWeight (0.45)
Outcome 3: MaintainWeight (0.55)
1
UI/UX
Service
Query
Parser
Request
Validation &
Verification
Ensemble/hybrid Reasoner
3
Outcomes
Aggregator
n Results
Rules/Pattern Matcher
Knowledge
Base
Knowledge
(Rules,
Cases)
Knowledge
Maintenance
Engine
Inference Engine
2
Conflict Resolver
Inference
Generator
Prediction : Historical Data Analysis
Prescriptive
What action
should be taken
Recommendation
manager
What Will happen?
Prediction
Predictive
Historical Data Analysis
Why did it happen?
Descriptive
21
/
What
happened?
Context and
Behavior analysis
Hindsight
Insight
Action
Prediction
22
Prediction
Y = 0.0833x + 0.5
3
Diet Bahaviour
Rules-based reasoning
/
8 Weeks Diet Trend
2
1
0
0
Predictor
1
2
3
4 5
Week
6
7
8
9
0= Imbalance; 1= Balance; 2= Overeaten
• Input:
• Reasoning outcomes
• Behaviors report
Outcome 1:
Lose Weight
+
M
M
M
M
Regular
Regular
Irregular
Cycling
Imbalance
Balance
Balance
Imbalance
No
Yes
No
Yes
5.9
5.5
5.3
5.7
120.2
167.1
132.2
160.8
5
M
Irregular Overeaten
No
5.1
178.9
6
M
Regular Overeaten
No
4.5
178.1
7
M
Irregular Overeaten
No
5.3
179.2
8
M
Regular
Yes
5.7
180.8
• Predictions
Balance
1
Weight
1
2
3
4
HDFS Data Access
Interface
8 Weeks Weight Trend
Weight
Exe:
Duration
Diet
Y = 7.625x + 127.85
230
Intermediate
Data
180
130
80
0
3
Predictor
criterionSelection (rOutcomes, bReport)
correlationAnalysis (bReport)
predictiveModeling()
evaluateModel ()
storeMOdel ()
redictPatterns (pModel, sGuidlines)
generatePredictions()
• Output:
Disease Height
Behavior Reports
• Methods:
•
•
•
•
•
•
•
Diet
Weight in lbs
Week Gender Exercise
Report
Analyzer
Predictive
Modeler
Criterion Selector
Model Builder
Correlation
Extractor
Model Evaluator
Predictions
Generator
Reasoner
3
4
5
6
7
8
9
4
Recommendation
Manager
Reasoner Top n Outcomes
1.
Lose Weight
2.
Maintain Weight
3.
Option n
1
2
Week
Prediction Models
Reasoner Top n Outcomes
1.
Lose Weight
2.
Maintain Weight
3.
Option n
1
Uniqueness and Contributions
• Accurate recommendations
• Hybrid & ensemble-based reasoning are used that increases the predictive accuracy
of recommendations
• Enhanced recommendations
• Both reasoning and prediction are combined together to get enhanced
recommendations
• The reasoner recommendations are supported by predictive analytics from the user
behaviors reports to make them more acceptable to users.
• Highly Integrated environment
• Closely coupled components for data abstraction, learning and reasoning
/
23
Conclusion and Future Work
• Conclusion
• Reasoner is designed to achieve the goal of highly accurate recommendations
• For accurate recommendation, multiple classical reasoning methods are
combined together either in a hybrid passion or ensemble way.
• The reasoner recommendations are supported by predictive analytics from
the users behavior reports to make them more acceptable to them.
• Future Work
•
•
•
•
Prediction detail study and utilization in Mining Minds Platform
Sub-component elaboration
Method and Data level with input/output specification
Real Scenario Finalization
/
24