2024 Invited Speakers


 

Invited Speech I by Assoc. Prof. Jing Zhang, Southeast University, China


Title: Machine Learning with Crowdsourced Labeled Data

 

Abstract: In today's big data era, a large amount of raw data is continuously generated, but data annotation for machine learning tasks is expensive and time-consuming. The emergence of crowdsourcing systems provides a feasible solution to this problem. Requesters can fast recruit workers from crowdsourcing platforms to carry out data annotation at a low cost. However, the crowdsourced workers are not experts so noises inevitably exist in the annotations. How to use these noisy labeled data for machine learning has become one of the hotspots in current machine learning research. This talk focuses on improving label quality and building high-quality learning models in a crowdsourcing environment. The content covers classic algorithms of crowdsourcing truth inference and model learning, error correction for crowdsourced annotations, active crowdsourcing learning, and so on.

     
 

Invited Speech II by Dr. Xiucai Ye, University of Tsukuba, Japan


Title: Multiview Network Embedding for Drug-target Interaction Prediction

 

Abstract: Drug-target interactions (DTI) prediction aims to identify new targets for existing drugs, which can significantly reduce the time and cost required for drug repositioning and drug discovery. We develop a computational framework based on multi-view network embedding for DTI prediction. Our framework uses a network embedding model to learn the feature representations of drugs and targets, and then adopts matrix completion scheme to predict the potential DTIs based on drug and target features, respectively. Experimental results on two public datasets show its effectiveness for DTI prediction.

     
 

Invited Speech III by Prof. Dong Huang, South China Agricultural University, China


Title: Ultra-Scalable Spectral Clustering: From Single-View to Multi-View

 

Abstract: Traditional spectral clustering suffers from cubic time complexity and quadratic space complexity, which significantly restricts its application in large-scale scenarios. In this talk, we focus on the robustness and scalability of spectral clustering for extremely large datasets. Specifically, we first introduce two large-scale clustering algorithms, namely, ultra-scalable spectral clustering (U-SPEC) and ultra-scalable ensemble clustering (U-SENC), which enjoy the time complexity of O(Np^0.5) and O(Nmp^0.5), respectively, and are capable of robustly and efficiently partitioning ten-million-level datasets on a personal computer with 16GB of memory. Then we extend the single-view framework to multiple views, and introduce the fast multi-view clustering via ensembles (FastMICE) algorithm, which is capable of partitioning very large-scale multi-view datasets in linear time. Future directions will also be discussed.

     

Copyright © 2024 16th International Conference on Machine Learning and Computing