Conference and journal publications, organized by year.
Showing all 00 publications
Filter by topic
Filter by type
2027
Energy-based Reference Distribution Learning
for Graph-level Open-World
Classification
Heng Lian, Ziyu Wang, Lusi Li, Xingquan
Zhu, and
Yi
He
ACM
SIGKDD Conference on
Knowledge Discovery and
Data Mining
(KDD)
· 2027
OGRD compares each graph’s node
embedding distribution with
class-specific
references and uses energy scores to
assign known class labels or flag graphs
from unseen classes.
U-Net-Guided Detection and Measurement of Growth Bands in
Atlantic Surfclam Shell Images
Sophia Sultan, Heng Lian, Roger Mann, and
Yi He
IEEE
International Conference on Big Data
(IEEE Big Data) · High School
Symposium ·
2026
Uses U-Net-guided image analysis to
detect and measure growth bands in Atlantic surfclam
shells, supporting the quantitative analysis of shell
imagery.
U-NetShell ImagingGrowth Bands
Augmentation with Reinforcement Transformation for
Low-textured Scientific Sketch Image Retrieval
Zhiqiang Wang, Sheri L. Barton, Jennifer
H. Johnson, Yi He,
and Xingquan Zhu
IEEE
International Conference on Big Data
(IEEE Big Data) ·
2026
Studies augmentation with reinforcement
transformation for retrieving low-textured scientific
sketch images.
Extends learning from label proportions
to graphs by generating topology-aware soft
pseudo-labels through optimal transport and training
graph neural networks from aggregate supervision.
Label
ProportionsWeak
SupervisionOptimal Transport
Benchmarks cold-start drug–drug
interaction prediction with mechanism-stratified splits
and evidence interventions to distinguish access to
biomedical knowledge from models’ actual use of it.
Combines ghost-node proxies for
unobserved boundary forcing with local and global
physics refiners to stabilize autoregressive forecasting
on open-boundary hydrologic graphs.
Develops online decision trees for varying feature
spaces through
gradient-based surrogate modeling, addressing prediction
when the
attributes available in a data stream change.
Combines online Gaussian-copula
alignment, uncertainty-calibrated geometric
pseudo-labels, and dual-track learning to adapt to
heterogeneous, incomplete, drifting data streams when
ground-truth feedback arrives late.
Relates context engineering to
fine-tuning and uses curriculum-guided,
confidence-adaptive prompts to align multimodal entities
progressively from simpler to more complex evidence.
Improves unsupervised multimodal entity
alignment by refining pseudo-seed precision and
balancing their graph coverage through multimodal
information, clustering, and resampling.
Recovers missing sensor signals through
federated latent factor learning that exchanges
gradients rather than raw readings and regularizes
representations with spatial and temporal correlations.
Federated
LearningSignal
RecoverySpatiotemporal Data
Generates SQL from natural-language
questions using a trained schema linker, schema-aware
retrieval of examples, and Pareto-based selection among
candidate queries.
Uses state-specific and state-invariant
reinforcement-learning agents, a DAG generation policy,
and factored actions to incrementally discover causal
graph structures from observational data.
Constructs query-specific evidence
graphs by recovering missing relations from source text
and selecting relevant facts jointly with existing
knowledge-graph edges for question answering.
Models unobserved external forcing with
ghost-node proxies and coupled boundary–interior
fixed-point inference to make graph neural networks more
consistent for fluid-network flux prediction.
A spatiotemporally regularized tensor
factorization model represents changing sensing-network
dynamics by combining latent tensor representations with
spatial and temporal structural constraints.
NGHF uses discrete differential
evolution to optimize binary latent factors directly,
learning representations of high-dimensional incomplete
data without the quantization step used by continuous
relaxations.
This survey organizes computational
pathology foundation models by pretraining datasets,
adaptation strategies, and evaluation tasks,
highlighting multimodal integration, domain variability,
and gaps in standardized assessment.
SADD-RFCO partitions features into two
views and co-trains random forests with iterative
pseudo-labeling and soft voting to detect anomalies in
partially labeled, contaminated data.
MMA combines four autoencoders defined
by different loss and regularization norms, using
adaptive weighting to represent heterogeneous
high-dimensional data with missing entries.
A controlled study evaluates GPT-4 and a
scholarly search plugin for generating WIC policy
research questions, comparing literature-informed
prompting with prompts without literature context
through expert assessment.
Large Language
ModelsResearch
QuestionsPublic HealthWIC
PhysicsNFP combines directional
difference matrices with physical constraints on graph
message passing to preserve high-frequency topological
signals when predicting flows on directed networks.
Outlines a dissertation on learning from streams with
evolving features,
variable input quality, and heterogeneous modalities,
using shared
representations to handle changes in feature
dimensionality.
Open Feature SpacesData StreamsMultimodal Learning
Investigates link completion in dynamic
knowledge graphs whose entities drift over time,
focusing on maintaining graph relationships as the
represented entities evolve.
LHGEL ensembles graph learners trained
on sampled batch views, combining residual attention and
diversity regularization to learn scalable
representations of large heterogeneous graphs.
HGEN ensembles heterogeneous-graph
neural networks built from meta-paths and random
dropping, using residual attention and correlation
regularization to balance learner accuracy and
representational diversity.
TCR-QF restores source-text context
around knowledge-graph triples and iteratively adds
query-relevant missing knowledge, addressing information
loss in graph-augmented question answering.
SAD combines sparse online feature
selection with CUR-based active instance selection,
using reconstruction residuals to score anomalies in
evolving streams under computational and memory
constraints.
This framework constructs query-specific
multimodal knowledge graphs through semantic graph
patterns, multi-path retrieval, and on-demand
information completion to support cross-modal question
answering.
MacFRL reorders continual-learning tasks
so a classifier learns from confident tasks before
tackling larger distribution shifts, aiming to sustain
group fairness with unlabeled data.
SOOFS combines mixed-norm row sparsity
with online CUR decomposition to select informative
instances within sliding windows while learning from
data streams with changing feature spaces.
Introduces a federated latent factor
learning model that combines projection operations and
Hamming hashing to learn compact representations from
distributed data sources.
Introduces a benchmark of complex
scientific tables with human-verified questions that
test language models on cell selection and arithmetic
reasoning across irregular table layouts.
Combines table structure recognition,
optical character recognition, and conformal prediction
to flag uncertain extractions and prioritize human
verification of complex scientific tables.
Uses deep neural networks to localize
distinctive scars and other markers in expert-drawn
sketches, enabling interpretable matching and
identification of individual manatees.
Surveys computational pathology
foundation models that combine histology images with
language, knowledge graphs, or gene expression,
organizing their datasets, downstream tasks, training
methods, and evaluation challenges.
Uses a copula-based latent
representation, adaptive drift-detection windows, and
geometric label propagation to learn from mixed-type
streaming features when distributions shift and labels
are incomplete.
Recovers random and consecutive missing
energy measurements through matrix reshaping, two-stage
temporal reconstruction, and multi-norm optimization
solved with an alternating-direction method.
Combines image deraining and dehazing
with an attention-enhanced YOLOv7 detector to identify
insulator defects in aerial photographs captured under
adverse weather.
Combines low-rank latent feature
modeling, spatiotemporal regularization, and an
outlier-resistant loss to recover missing sensor
measurements from noisy wireless sensing networks.
Uses mixed-norm online CUR decomposition
to jointly select informative instances and stable
features for sparse active learning in streams with
evolving feature spaces.
Online Active
LearningFeature
SelectionCUR
DecompositionMixed-Norm Sparsity
Combines meta-path topology
homogenization with a topology-and-content-aware graph
transformer to learn label distributions for nodes in
heterogeneous graphs while preserving structural and
semantic consistency.
Heterogeneous
GraphsLabel
Distribution LearningGraph
TransformersMeta-Paths
Learns a universal representation for
evolving soft-sensing features and balances prediction
risk against abstention costs when previously unknown
classes appear in streaming observations.
Open-World
LearningSoft
SensingSelective
PredictionEvolving Features
Builds causal graphs around prediction
labels and protected attributes to select streaming
features while balancing predictive accuracy and
fairness as features arrive over time.
Learns runtime models from dynamic
type-check overheads to predict how type-annotation
configurations affect gradual programs and support
performance-aware migration between typing
configurations.
Co-trains graph convolutional networks
over instance and label relationships to learn node
label distributions while jointly modeling dependencies
among instances, labels, and their cross-relations.
Label Distribution
LearningGraph
ConvolutionLabel Correlations
MKG-FENN predicts drug–drug interaction
events by extracting features from four complementary
molecular and biomedical knowledge graphs and fusing
them through an end-to-end neural network.
RSOL uses a mixed-norm regularizer and
closed-form proximal updates to maintain sparse online
predictors when features emerge or disappear from a data
stream.
MLF-RFDC combines density-based noise
detection, latent-factor data recovery, and
random-forest forecasting to integrate multiple
electrical load types while accounting for distinct
consumption patterns.
OLD3S learns shared latent
representations and adapts model capacity online,
transferring knowledge across evolving feature spaces
while retaining information learned from previously
observed features.
Feature
EvolutionVariational
ModelsKnowledge
TransferAdaptive Model
Capacity
MMLF adaptively ensembles six
latent-factor models spanning different vector spaces
and loss norms to represent heterogeneous,
high-dimensional matrices with missing observations.
Multi-Metric
LearningMatrix
CompletionAdaptive
EnsemblesIncomplete Data
AutoRec++ combines preprocessing and
training bias corrections with mixed loss norms,
addressing multiple sources of bias and outliers in
autoencoder-based recommendation from user behavior
data.
PMLF builds a multilayer latent-factor
architecture that progressively supplements sparse
matrices with synthetic entries, using prediction
sampling and nonlinear activations to constrain the
data-generation process.
Matrix
CompletionMultilayer
ModelsPrediction
SamplingSparse Data
IKGN integrates visit histories,
trajectories, and location categories into an
intention-aware knowledge graph, combining
attention-based propagation with sequential modeling for
point-of-interest recommendation.
POI
RecommendationUser
IntentionsKnowledge
GraphsSequential Modeling
MMA combines four autoencoders with
different loss and regularization norms through adaptive
weighting, learning representations of heterogeneous
high-dimensional data with incomplete observations.
AutoencodersMulti-Metric
LearningAdaptive
EnsemblesIncomplete Data
This survey organizes online learning in
open feature spaces into a unified taxonomy, compares
representative algorithms, and identifies application
scenarios, limitations, and research directions.
Open Feature
SpacesSurveyAlgorithm
TaxonomyBenchmarking
OSLMF uses copula-based feature
alignment and online neighborhood label propagation to
learn from mixed-type, evolving feature spaces when only
a subset of arriving instances is labeled.
ORF3V maintains an interpretable
ensemble of feature-specific decision forests, updating
and pruning compressed feature representations as
variables appear or disappear in online data streams.
GLFA extracts hidden high-order
interactions from sparse matrices as a graph and
incorporates them into a recurrent latent-factor
architecture for missing-entry prediction.
LDSA-AE classifies energy-internet data
by activity and adapts replication and erasure coding to
balance distributed storage overhead, access latency,
repair speed, and reliability.
Distributed
StorageAdaptive
EncodingErasure CodesEnergy Internet
BCTC-KSM combines verifiable threshold
secret sharing, blockchain records, and attribute-based
access controls to manage encryption keys for
cloud-stored power-IoT data shared among authorized
participants.
D2E-LF ensembles inner-product and
distance-based latent-factor predictors with mixed loss
norms, capturing complementary characteristics of sparse
web-service quality-of-service observations for
missing-value prediction.
Combines Bayesian risk modeling,
epidemic-style propagation, and optimal load shedding to
assess how cyberattacks and device dependencies affect
the security of distribution cyber-physical systems.
Investigates multi-granular neural
architecture search in federated settings, combining
model selection across distributed datasets with the
requirement to preserve the privacy of client data.
Learns a shared latent representation
across evolving feature spaces while adapting neural
network depth and parameters online to accommodate
changes in the complexity of streaming data.
Estimates missing streaming features
with latent factor analysis, then combines fuzzy logic
and neighborhood rough sets to address uncertainty
between imputed feature values and class labels.
Combines preprocessing and training
debiasing strategies within an autoencoder-based
recommender to address biases in user behavior data
without changing the underlying neural network
architecture.
Uses a Cauchy-loss autoencoder to model
user–item interactions while reducing the influence of
outlier ratings and potentially malicious user behavior
on the resulting recommendations.
Generates camouflaged fake nodes and
learned connections to poison graph neural networks for
global or targeted misclassification without directly
modifying the features of original graph nodes.
Studies robust latent factor analysis
for recovering incomplete wireless sensor network data,
using latent representations to estimate missing
measurements from the available sensor observations.
Integrates differential evolution into
recommender training to learn layer-specific learning
rates and regularization coefficients automatically,
reducing dependence on manually configured searches for
these hyperparameters.
Combines a K-means and least-squares
forecasting-error model with gene expression programming
to incorporate uncertain load-data errors into
quantitative prediction of electricity demand from
observed data.
Uses latent factor analysis to estimate
missing values before selecting features online,
enabling feature selection as incomplete feature columns
arrive for a fixed set of instances.
Uses density-peaks clustering over
latent user and service representations to identify
neighborhoods and noise, adapting quality-of-service
prediction to the characteristics of sparse observations
from web services.
Builds user and service neighborhoods
from initially learned latent factors, then uses those
posterior neighborhoods to regularize quality-of-service
predictions without requiring additional geographical
information.
Studies an ensemble of latent factor
models for predicting missing web-service
quality-of-service values, combining multiple predictors
to model sparse observations of interactions between
users and services.
Models dependencies among mixed-type
streaming features with a Gaussian copula, allowing an
online learner to transfer information across feature
spaces as features appear and disappear.
Uses a Bayesian-regularized generative
adversarial network to synthesize minority-class
examples with interpretable feature dependencies
specified through human input, supporting classification
when labeled data are imbalanced.
Initializes lifelong learning with one
labeled task, then orders unlabeled tasks by their
distance to accumulated knowledge to reduce negative
transfer.
Combines a universal feature space with
sparse manifold regularization to update online
classifiers when feature sets vary and supervision is
available only intermittently.
Toward Wiser Funding Allocation with Bayesian Networks
Jing Chen, Yi
He, and Vijay Raghavan
Machine Learning
for Consumers and Markets (MLCM) · KDD
Workshop ·
2021
Examines the use of Bayesian networks to
inform funding-allocation decisions, representing
uncertain dependencies relevant to distributing limited
resources among competing alternatives.
Bayesian
NetworksFunding
AllocationDecision Support
Uses a generative graph to reconstruct
related features in a universal space, allowing online
learning from streams whose features can appear or
disappear without predefined rules.
Builds a deep recommender by
sequentially connecting latent factor models through
nonlinear activations, learning from high-dimensional
sparse rating matrices with computation that scales with
depth.
Combines node attributes and network
topology through a multi-granular graph autoencoder to
select informative instances for labeling, with a query
strategy designed to tolerate noisy labels.
Learns a shared latent feature space
across related tasks and uses a generative graph to
recover missing representations, enabling predictions
from only partially collected task examples.
Trains traveling classifiers online from
consecutive neural-network layers to identify redundant
computation after deployment and produce earlier
predictions as incoming data reveal application-specific
patterns.
Uses cross-outlook feature relationships
to build a shared repository that reconstructs missing
representations, enabling prediction when instances are
observed through only some heterogeneous views.
Learns nonlinear combinations of
heterogeneous vertex features through deep matrix
tri-factorization, extracting higher-level
representations to model interactions between vertices
in multi-space attributed graphs.
Attributed
GraphsMatrix
Tri-FactorizationHeterogeneous
Features
Completes missing entries in newly
arriving features with latent factor analysis, allowing
existing online feature-selection algorithms to process
capricious streams containing incomplete feature
observations.
Combines source-code analysis with
transfer learning and project-specific features to infer
C error-return specifications, requiring no manual
labels from users of previously unseen projects.
Program
AnalysisZero-Shot
LearningError
Specifications
Constructs a universal feature space
with a generative graphical model so online learners can
transfer previously learned patterns to new features in
arbitrarily changing data streams.
Integrates density-peak clustering into
a latent factor model to identify service neighborhoods
and unreliable quality-of-service measurements,
predicting missing values without requiring additional
geographic information.
QoS PredictionDensity-Peak
ClusteringUnreliable Data
Studies a supervised data-synthesis and
evolution framework for traffic-crash severity
classification, connecting generated training examples
with predictive models for real-world transportation
records.
Applies three layers of
conditional-independence filtering to approximate a
Markov blanket, removing irrelevant and redundant
features as candidate variables arrive in a stream.