Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval,
Journal Year:
2022,
Volume and Issue:
unknown, P. 176 - 186
Published: July 6, 2022
Conversational
search
is
a
crucial
and
promising
branch
in
information
retrieval.
In
this
paper,
we
reveal
that
not
all
historical
conversational
turns
are
necessary
for
understanding
the
intent
of
current
query.
The
redundant
noisy
context
largely
hinder
improvement
performance.
However,
enhancing
denoising
ability
quite
challenging
due
to
data
scarcity
steep
difficulty
simultaneously
learning
query
encoding
denoising.
To
address
these
issues,
present
novel
Curriculum
cOntrastive
conTExt
Denoising
framework,
COTED,
towards
few-shot
dense
Under
curriculum
training
order,
progressively
endow
model
with
capability
via
contrastive
between
noised
samples
denoised
generated
by
new
conversation
augmentation
strategy.
Three
curriculums
tailored
exploited
our
framework.
Extensive
experiments
on
two
datasets,
i.e.,
CAsT-19
CAsT-20,
validate
effectiveness
superiority
method
compared
state-of-the-art
baselines.
Next
point-of-interest
(POI)
recommendation
is
a
hot
research
field
where
recent
emerging
scenario,
next
POI
to
search
recommendation,
has
been
deployed
in
many
online
map
services
such
as
Baidu
Maps.
One
of
the
key
issues
this
scenario
providing
satisfactory
for
cold-start
cities
with
limited
number
user-POI
interactions,
which
requires
transferring
knowledge
hidden
rich
data
from
other
these
cities.
Existing
literature
either
does
not
consider
city-transfer
issue
or
cannot
simultaneously
tackle
sparsity
and
pattern
diversity
among
various
users
multiple
To
address
issues,
we
explore
that
transfers
scarce
data.
We
propose
novel
Curriculum
Hardness
Aware
Meta-Learning
(CHAML)
framework,
incorporates
hard
sample
mining
curriculum
learning
into
meta-learning
paradigm.
Concretely,
CHAML
framework
considers
both
city-level
user-level
hardness
enhance
conditional
sampling
during
meta
training,
uses
an
easy-to-hard
city-sampling
pool
help
meta-learner
converge
better
state.
Extensive
experiments
on
two
real-world
datasets
Maps
demonstrate
superiority
framework.
Mathematics,
Journal Year:
2023,
Volume and Issue:
11(5), P. 1098 - 1098
Published: Feb. 22, 2023
The
explosive
increase
in
educational
data
and
information
systems
has
led
to
new
teaching
practices,
challenges,
learning
processes.
To
effectively
manage
analyze
this
information,
it
is
crucial
adopt
innovative
methodologies
techniques.
Recommender
(RSs)
offer
a
solution
for
advising
students
guiding
their
journeys
by
utilizing
statistical
methods
such
as
machine
(ML)
graph
analysis
program
student
data.
This
paper
introduces
an
RS
advisors
that
analyzes
records
develop
personalized
study
plans
over
multiple
semesters.
proposed
system
integrates
ideas
from
theory,
performance
modeling,
ML,
explainable
recommendations,
intuitive
user
interface.
implicitly
implements
many
academic
rules
through
network
analysis.
Accordingly,
systematic
comprehensive
review
of
different
students’
was
possible
using
metrics
developed
the
mathematical
theory.
systematically
assesses
measures
relevance
particular
student’s
plan.
Experiments
on
datasets
collected
at
University
Dubai
show
model
presented
outperforms
similar
ML-based
solutions
terms
metrics.
Typically,
up
86%
accuracy
recall
have
been
achieved.
Additionally,
lowest
mean
square
regression
(MSR)
rate
0.14
attained
compared
other
state-of-the-art
regressors.
IEEE Transactions on Neural Networks and Learning Systems,
Journal Year:
2023,
Volume and Issue:
35(8), P. 10894 - 10908
Published: Feb. 23, 2023
Multiple
unmanned
aerial
vehicles
(UAVs)
are
able
to
efficiently
accomplish
a
variety
of
tasks
in
complex
scenarios.
However,
developing
collision-avoiding
flocking
policy
for
multiple
fixed-wing
UAVs
is
still
challenging,
especially
obstacle-cluttered
environments.
In
this
article,
we
propose
novel
curriculum-based
multiagent
deep
reinforcement
learning
(MADRL)
approach
called
task-specific
MADRL
(TSCAL)
learn
the
decentralized
with
obstacle
avoidance
UAVs.
The
core
idea
decompose
task
into
subtasks
and
progressively
increase
number
be
solved
staged
manner.
Meanwhile,
TSCAL
iteratively
alternates
between
procedures
online
offline
transfer.
For
learning,
hierarchical
recurrent
attention
actor-critic
(HRAMA)
algorithm
policies
corresponding
subtask(s)
each
stage.
transfer,
develop
two
transfer
mechanisms,
i.e.,
model
reload
buffer
reuse,
knowledge
neighboring
stages.
A
series
numerical
simulations
demonstrate
significant
advantages
terms
optimality,
sample
efficiency,
stability.
Finally,
high-fidelity
hardware-in-the-loop
(HITL)
simulation
conducted
verify
adaptability
TSCAL.
video
about
HITL
available
at
https://youtu.be/R9yLJNYRIqY.
IEEE Transactions on Pattern Analysis and Machine Intelligence,
Journal Year:
2024,
Volume and Issue:
46(6), P. 4188 - 4205
Published: Jan. 16, 2024
Existing
studies
on
knowledge
distillation
typically
focus
teacher-centered
methods,
in
which
the
teacher
network
is
trained
according
to
its
own
standards
before
transferring
learned
a
student
one.
However,
due
differences
structure
between
and
student,
by
former
may
not
be
desired
latter.
Inspired
human
educational
wisdom,
this
paper
proposes
Student-Centered
Distillation
(SCD)
method
that
enables
adjust
transfer
network's
needs.
We
implemented
SCD
based
various
e.g.,
identified
validation
set,
then
transferred
it
latter
through
training
set.
To
address
problems
of
current
deficiency
knowledge,
hard
sample
learning
forgetting
faced
process,
we
introduce
improve
Proportional-Integral-Derivative
(PID)
algorithms
from
automation
fields
make
them
effective
identifying
required
network.
Furthermore,
propose
curriculum
learning-based
fuzzy
strategy
apply
proposed
PID
control
algorithm,
such
can
actively
pay
attention
challenging
samples
after
with
certain
knowledge.
The
overall
performance
verified
multiple
tasks
comparing
state-of-the-art
ones.
Experimental
results
show
our
student-centered
outperforms
existing
