Download Hidden Concept Detection in Graph

HIDDEN CONCEPT DETECTION IN GRAPHBASED RANKING ALGORITHM FOR
PERSONALIZED RECOMMENDATION
Nan Li
Computer Science Department
Carnegie Mellon University
INTRODUCTION
Previous work:
 Represents past user behavior through a
relational graph.
Fail to represent individual differences among
items of a same type.
 Our work:
 Detect hidden concepts embedded in the
original graph
 Build a two-level type hierarchy for explicit
representation of item characteristics.

RELATIONAL RETRIEVAL
1.
2.
Entity-Relation Graph G=(E, T, R):
•
Entity set E={e} Entity types set T={T} Entity
relations R={R}
•
Each entity e in E has its type e.T .
•
Each relation R has two entity types R.T1 and
R.T2. If two entities has relation R, then R(e1,
e2) = 1, o/w 0.
Relational Retrieval Task: Query q = (Eq , Tq)
•
Given Eq = {e’}, predict the relevance of
each entity e of the target type Tq.
PATH RANKING ALGORITHM
Relational Path:

P = (R1, R2, …, Rn) R1.T1=T0 and
Ri.T2=Ri+1.T1.
2. Relational Path Probability Distribution:

The probability corresponds to the
probability of a path random walker reaching
that entity from a query entity.
1.
PRA MODEL

(G, l, θ)
•
The feature matrix A has its each column to be
the distribution hp(e).
•
The scoring function:
TRAINING PRA MODEL
1.
2.
3.
Training data: D = {(q(m),y(m))}, ye(m)=1 if e is
relevant to the query q(m)
Parameter: The weight of path θ
Objective function:
HIDDEN CONCEPT DETECTOR (HCD)

Two-Layer PRA
Find hidden subtype of
relations
autho
r
title
pape
r
gene
journ
al
yea
r
autho
r
title
pape
r
gene
journ
al
yea
r
BOTTOM-UP HCD
Bottom-Up merging algorithm:
 For each relation type Ri


Step1: Divide every starting node of relation Ri as a
subrelation Rij.
author

paper
author
paper
Step2: HAC: Each time merge two subrelations Rim
and Rin to maximize the gain of objective functions until
no positive gain:
author
paper
author
paper
APPROXIMATE THE GAIN OF OBJECTIVE
FUNCTION
1.
Calculate the maximum gain of two relations: gm
and gn
2.
Use taylor series to approximate:
EXPERIMENTAL RESULTS
1.
Data Sets:

Saccharomyces Genome Database, a publication data set
about the yeast organism Saccharomyces cerevisiae
Three measurements:
2.



Mean Reciprocal Rank (MRR): inverse of the rank of the
first correct answer
Mean Average Precision (MAP): the area under the
Precision-Recall curve
p@K: precision at K, where K is the actually number of
relevant entities.
NORMALIZED CUT

Training data:


Number of clusters ↑
Recommendation quality↑
Test data:

NCut outperforms random
HCD
•
Training data:
•
•
HCD outperforms PRA
in all three
measurements
Test data:
•
Two systems perform
equally well
FUTURE WORK
Bottom-Up vs Top Down
 Improve Efficiency
 Type Recovery in Non-Labeled Graph

Building an intelligent agent
that simulates human-level
learning using machine
learning techniques
A COMPUTATIONAL MODEL OF
ACCELERATED FUTURE
LEARNING THROUGH FEATURE
RECOGNITION
Nan Li
Computer Science Department
Carnegie Mellon University
ACCELERATED FUTURE LEARNING

Accelerated Future Learning
Learning more effectively because of prior learning
 Has been observed a lot
 How?


Expert vs Novice
Expert  Deep functional feature (e.g. -3x  -3)
 Novice  Shallow perceptual feature (e.g. -3x  3)

A COMPUTATIONAL MODEL
Model Accelerated Future Learning
 Use Machine Learning Techniques
 Acquire Deep Feature
 Integrated into a Machine-Learning Agent

AN EXAMPLE IN ALGEBRA
FEATURE RECOGNITION AS
PCFG INDUCTION
Under lying structure in the problem 
Grammar
 Feature  Intermediate symbol in a grammar
rule
 Feature learning task  Grammar induction
 Error  Incorrect parsing

PROBLEM STATEMENT

Input is a set of feature recognition records
consisting of
An original problem (e.g. -3x)
 The feature to be recognized (e.g. -3 in -3x)


Output
A PCFG
 An intermediate symbol in a grammar rule

ACCELERATED FUTURE LEARNING
THROUGH FEATURE RECOGNITION
Extended a PCFG Learning Algorithm (Li et al.,
2009)
 Feature Learning
 Stronger Prior Knowledge:



Transfer Learning Using Prior Knowledge
Better Learning Strategy:

Effective Learning Using Bracketing Constraint
A TWO-STEP ALGORITHM
•
Greedy Structure
Hypothesizer:

•
Hypothesizes the schema
structure
Viterbi Training Phase:


Refines schema
probabilities
Removes redundant
schemas
Generalizes Inside-Outside Algorithm (Lary & Young,
1990)
GREEDY STRUCTURE HYPOTHESIZER
Structure learning
 Bottom-up
 Prefer recursive to non-recursive

EM PHASE

Step One:
Plan parse tree
computation
 Most probable parse
tree


Step Two:
Selection probabilities
update
 s: ai  p, aj ak

FEATURE LEARNING

Build Most Probable
Parse Trees


For all observation
sequences
Select an
Intermediate Symbol
that

Matches the most
training records as the
target feature
TRANSFER LEARNING USING PRIOR
KNOWLEDGE

GSH Phase:
Build parse trees
based on previously
acquired grammar
 Then call the original
GSH


Viterbi Training:

Add rule frequency in
previous task to the
current task
0.5
0.3
0.5
0.6
3
6
EFFECTIVE LEARNING USING
BRACKETING CONSTRAINT

Force to generate a
feature symbol
Learn a subgrammar
for feature
 Learn a grammar for
whole trace
 Combine two
grammars

EXPERIMENT DESIGN IN ALGEBRA
EXPERIMENT RESULT IN ALGEBRA
Fig.2. Curriculum one
Fig.3. Curriculum two
Fig.4. Curriculum three
 Both stronger prior knowledge and a better learning strategy can yield
accelerated future learning
 Strong prior knowledge produces faster learning outcomes
 L00 generated human-like errors
LEARNING SPEED IN
SYNTHETIC DOMAINS
 Both stronger prior knowledge and a better learning strategy yield faster
learning
 Strong prior knowledge produces faster learning outcomes with small amount
of training data, but not with large amount of data
 Learning with subtask transfer shows larger difference, 1) training process; 2)
SCORE WITH INCREASING DOMAIN SIZES
 The base learner, L00, shows the fastest drop
 Average time spent per training record
 Less than 1 millisecond except for L10 (266 milliseconds)
 L10: Need to maintain previous knowledge, does not separate trace into
small traces
INTEGRATING ACCELERATED FUTURE
LEARNING IN SIMSTUDENT
•
Prepare Lucky for Quiz Level 3 !
A machine-learning
agent that
•
Curriculum Browser
x+5
Level 1:
[+] One-Step Linear Equation
•
Level 2:
[+] Two-Step Linear Equation
Level 3:
[-] Equation with Similar Terms
Overview
In this unit, you will solve
equations with integer or decimal
coefficients, as well as equations
involving more than one variable.
More…
•
Integrate the acquired
grammar into
production rules

Lucky
Tutor Lucky
Next Problem

Quiz Lucky
Acquires production
rules from
Examples and
problem solving
experience
Requires weak
operators (non-domain
specific knowledge)
Less number of
operators
CONCLUDING REMARKS
Presented a computational model of human
learning that yields accelerated future learning.
 Showed

Both stronger prior knowledge and a better learning
strategy improve learning efficiency.
 Stronger prior knowledge produced faster learning
outcomes than a better learning strategy.
 Some model generated human-like errors, while
others did no make any mistake.
