Download Semantic Associations in Recommender Systems: Expanding Personalization and more Study notes Computer Science in PDF only on Docsity! ICAI Review
James Michaelis
CSCI 6965 — Adv. Semantic Technologies
(Class 7)
3/3/2009
Semantic Reasoning: A Path To New
Possibilities of Personalization.
Strategy
• Two
key
parts:
– Seman:c
associa:ons
• Trace
seman:c
bonds
between
the
user’s
preferences
and
the
items
available
in
the
recommender
system
• Formalize
these
bonds
in
a
domain
ontology,
along
with
their
seman:c
annota:ons
– Spreading
Ac:va:on
techniques
• Explore
these
seman:c
rela:onships
and
discover
new
knowledge
related
to
the
users’
interests.
• Domain
for
demonstra:on:
Movie/TV
Reviewing
Seman:c
Associa:ons
• Domain
knowledge
is
encoded
in
ontology
• In
turn,
users
can
specify
interest
in
certain
movies,
using
parameter
DegreesOfInterest
(con:nuous
variable
ranging
from
1
to
‐1)
• With
user
preferences
entered,
traversal
through
domain
ontology
takes
place
• Relevance
Index
computed
through
following
metrics:
–
ρ‐path
associa:on:
Applies
for
two
programs
linked
by
a
chain
or
sequence
of
proper:es
in
the
ontology.
–
ρ‐join
associa:on:
Two
programs
with
respec:ve
akributes
belonging
to
the
same
class
in
the
domain
ontology.
–
ρ‐cp
associa:on:
Two
programs
sharing
a
common
ancestor
in
the
genre
hierarchy
defined
in
the
ontology.
Semantic Associations
_ ee a anCaeecr Semantic Associations
—_ — a _ p-path (Jerry Maguire, Born on 4th July)
cae ae on, p-join (Welcome to Tokyo, The Last Samurai)
we RL f/orN p-join (Learn about WW I, Born on 4th July)
Come comme — <7 sain (camer) Cm p-ep (Vanilla Sky, The Last Samurai)
, “ C Poe p-cp (Danny the Dog, Game of Death)
SS ) cc. a> ce. ae p-path (Danny the Dog, Million Dollar Baby)
wwii, : \\atsouire ) + Seam Sy) Samara i
— Ne para) (Woteoma
ihe Dog Dollar Baby 4th of July ‘Shut of Glery te Tokyo
Toyo Kyoto er = Kung Fu Kamte
a 1D, Karate 10,
atic ag MasTepie 1D, aes
‘Dollar Baby. Has Actor Raney Teties Bace Lee
Morgan he Dog >
Game of
Freeman / Actorin Hastopic_( Come Hasactor
Kung Fu
Wor Nicole
Aaem seat —"\_ Var, }-—— Drecormn — HasActressy Kidman
WW ner
rope nt otGiony —e ‘Stantey >, Eyes wide os
Saaieeee nis 1D, ) Masdeector Haaren "Otsd
fo Tearyo, Tetyo on Aetoan Tom 1D,
rata —_—_—
rabid Ath of duly * Cruise. HasActor
if a HasTopc ae SY L nals
0 Jey
— — Maguire | vaspwecter 54
Kyoto ee
1D, 1. )*rarsa Ip, ) "ie (Cameran ‘4 HasDirector
Evalua:on
• Authors
implemented
prototype
of
recommender
system
for
tes:ng
purposes
including:
– An
OWL
ontology
covering
the
domain
of
TV
shows
• Similar
to
the
movie
ontology
the
technology
was
developed
around
– User
modeling
technique
based
on
ontology‐profiles
• Goal
of
study
was
to:
– To
evaluate
the
accuracy
of
our
reasoning‐based
recommenda:ons
– To
compare
this
approach
with
two
exis:ng
machine‐ learning
techniques
that
are
devoid
of
seman:c
inference
capabili:es
used
by
authors.
Evalua:on
• 400
Students
asked
to
rate
400
TV
programs
on
con:nuous
scale
of
[‐1,1]
• 40
%
of
student
evalua:ons
used
as
training
data
for
the
two
machine
learning
techniques
• For
remaining
60%
of
students,
the
following
took
place:
– 10
good,
and
10
bad
TV
shows
randomly
selected
from
each
student
evalua:on
to
build
profile
– In
turn,
these
profiles
were
fed
into
the
authors
recommenda:ons
system,
as
well
as
the
other
two
systems
90
80
70
60
50
40
30
20
10
Results
C0 Asso-SA (threshold = 0.65)
O Rules
[| i sem-cF
lt 10
8
6 —
4
2
o 2 2
Precision
Review
of
Paper
• Significance
of
Results:
8/10
– Interes:ng
comparison
done
between
the
authors
system
and
two
alternate
systems
– However,
it
is
not
par:cularly
clear
why
these
two
were
selected
(a
likle
more
background
on
selec:on
ra:onale
would
be
helpful)
– In
addi:on,
expanding
the
pool
of
par:cipants
and
or
domain
data
in
a
follow‐up
evalua:on
would
be
helpful
(an
inten:on
the
authors
expressed
as
future
work).
Review
of
Paper
• Technical
Soundness:
8/10
– Specifics
of
some
techniques
(such
as
those
used
to
compute
significance
of
proper:es
related
to
movies
rated
by
users)
omiked
from
paper
due
to
length
constraints
• However,
a
thorough
technical
review
of
all
the
techniques
used
in
this
system
given
in
Blanco
Y.
et
al.
– A
Flexible
seman:c
inference
methodology
to
reason
about
user
preferences
in
knowledge
based
recommender
systems.
Knowledge‐based
Systems,
In
press.
Review
of
Paper
• Quality
of
Evalua:on:
6/10
– Two
metrics
used
for
ra:ng
effec:veness
of
three
techniques
• Recall:
Defined
as
the
percent
of
returned
programs
that
the
user
rated
interes:ng
• Precision:
Defined
as
percent
of
programs
returned
with
posi:ve
user
ra:ng
on
[‐1,1]
scale
– Not
very
well
defined
what
dis:nguishes
these
metrics
in
the
results
– Expanded
sta:s:cal
analysis
would
be
helpful