Machine Learning and Knowledge Discovery in Databases. Research Track : European Conference, ECML PKDD 2024, Vilnius, Lithuania, September 9–13, 2024, Proceedings, Part IV / edited by Albert Bifet, Jesse Davis, Tomas Krilavičius, Meelis Kull, Eirini Ntoutsi, Indrė Žliobaitė.

Format:

Book

Contributor:

Bifet, Albert, editor.

Series:

Lecture Notes in Artificial Intelligence, 2945-9141 ; 14944

Language:

English

Subjects (All):

Artificial intelligence.

Computer engineering.

Computer networks.

Computers.

Image processing--Digital techniques.

Image processing.

Computer vision.

Software engineering.

Artificial Intelligence.

Computer Engineering and Networks.

Computing Milieux.

Computer Imaging, Vision, Pattern Recognition and Graphics.

Software Engineering.

Local Subjects:

Artificial Intelligence.

Computer Engineering and Networks.

Computing Milieux.

Computer Imaging, Vision, Pattern Recognition and Graphics.

Software Engineering.

Physical Description:

1 online resource (507 pages)

Edition:

1st ed. 2024.

Place of Publication:

Cham : Springer Nature Switzerland : Imprint: Springer, 2024.

Summary:

This multi-volume set, LNAI 14941 to LNAI 14950, constitutes the refereed proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD 2024, held in Vilnius, Lithuania, in September 2024. The papers presented in these proceedings are from the following three conference tracks: - Research Track: The 202 full papers presented here, from this track, were carefully reviewed and selected from 826 submissions. These papers are present in the following volumes: Part I, II, III, IV, V, VI, VII, VIII. Demo Track: The 14 papers presented here, from this track, were selected from 30 submissions. These papers are present in the following volume: Part VIII. Applied Data Science Track: The 56 full papers presented here, from this track, were carefully reviewed and selected from 224 submissions. These papers are present in the following volumes: Part IX and Part X.

Contents:

Intro

Preface

Organization

Invited Talks Abstracts

The Dynamics of Memorization and Unlearning

The Emerging Science of Benchmarks

Enhancing User Experience with AI-Powered Search and Recommendations at Spotify

How to Utilize (and Generate) Player Tracking Data in Sport

Resource-Aware Machine Learning-A User-Oriented Approach

Contents - Part IV

Research Track

Model Fusion via Neuron Transplantation

1 Introduction

2 Related Work

3 Neuron Transplantation

3.1 Analysis of NT

4 Experiments

4.1 Experimental Settings

4.2 Main Properties and Ablation Studies

4.3 Comparisons with Other Fusion Methods

5 Discussion and Conclusion

References

Compressed Federated Reinforcement Learning with a Generative Model

2.1 Single-Agent RL Algorithms

2.2 Distributed and Federated RL Algorithms

2.3 Communication-Efficient Learning Algorithms

2.4 Notation

3 Preliminaries and Background

3.1 Discounted Infinite-Horizon MDP

3.2 Policy, and Q-Function

3.3 Optimal Policy and Bellman Operator

3.4 RL with a Generative Model

4 CompFedRL

4.1 Compression Options for CompFedRL

5 Convergence Analysis

5.1 CompFedRL with UnbiasedComp

5.2 CompFedRLwith BiasedComp

6 Experiments

6.1 Setup

6.2 Communication Efficiency of CompFedRL

6.3 Convergence Speedup

6.4 Impact of Federated Parameter and Learning Rate

7 Conclusion

Walking Noise: On Layer-Specific Robustness of Neural Architectures Against Noisy Computations and Associated Characteristic Learning Dynamics

3 Methods

3.1 Global and Walking Noise

3.2 Data Sets, Models and Experimental Setup

4 Analysis of Global Noise

4.1 Quantifying Robustness

4.2 Robustness Results for Global Noise

5 Additive Noise.

5.1 Impact of Batch Normalization

5.2 Impact of Weight Magnitude

6 Multiplicative Noise

6.1 Self-binarization of Model Activations

6.2 The Impact of Batch Normalization

7 Mixed Noise

8 Making Use of Walking Noise Results

9 Summary

KAFÈ: Kernel Aggregation for FEderated

2.1 KDE in Federated Learning

2.2 Statistical Heterogeneity in Federated Learning

3 Methodology

3.1 Preliminaries

3.2 The Proposed Framework: KAFÈ

3.3 Convergence Analysis

4 Experiment

4.1 Experimental Setting

4.2 Empirical Results

5 Conclusion

On Suppressing Range of Adaptive Stepsizes of Adam to Improve Generalisation Performance

2 Algorithmic Design of SET-Adam

2.1 Motivation of Layerwise Down-Scaling Operation

2.2 Design of Layerwise Down-Scaling Operation

2.3 -Embedding for Suppressing Range of Adaptive Stepsizes

2.4 Down-Translating for Avoiding Extremely Small Adaptive Stepsizes

2.5 Convergence Analysis

3 Experiments

3.1 On Training a Transformer

3.2 On Training LSTMs

3.3 On Training VGG11 and ResNet34 over CIFAR10 and CIFAR100

3.4 On Training WGAN-GP over CIFAR10

3.5 On Training ResNet18 over ImageNet

4 Conclusions

Graph Attention Network with Relational Dynamic Factual Fusion for Knowledge Graph Completion

2.1 Classical Embedding Models

2.2 Graph Convolutional Network Models

2.3 Graph Attention Network Models

3 Preliminary

3.1 Point Mutual Information

3.2 Related Impact Factor

4 Our Proposal

4.1 Relational Diversified Information Extraction

4.2 Relational Dynamic Factual Fusion

4.3 Relational Diversified Information Sharing

4.4 Entity Updating

4.5 Training Objective

4.6 Decoder

5 Experiments.

5.1 Experimental Setup

5.2 Results

6 Conclusion

Low-Hanging Fruit: Knowledge Distillation from Noisy Teachers for Open Domain Spoken Language Understanding

3 Problem Setting

4 Noise Teacher and Consistently Guiding Student Paradigm

4.1 Incremental Progress Prompting Scheme for Intent and Slot Filling Distillation

4.2 Positively Fine-Tuned Paradigm

5 Experiment

6 Discussion

The Price of Labelling: A Two-Phase Federated Self-learning Approach

3 Overview and Fundamentals

4 The 2-Phase Federated Self-learning Framework

4.1 Local Data Augmentation

4.2 2PFL Training Phases

5 Experimental Evaluation

5.1 Experimental Set-up

5.2 Experimental Results

6 Conclusions

Disentangled Representations for Continual Learning: Overcoming Forgetting and Facilitating Knowledge Transfer

3.1 Disentangled Representation Encoders

3.2 Prevent Forgetting in Shared Encoder

3.3 Facilitate Knowledge Transfer Among Task Encoders

4.1 Experimental Setup

4.2 Results and Analysis

4.3 Knowledge Transfer Results

4.4 Ablation Study

4.5 Hyperparameter Sensitivity Analysis

On the Calibration of Epistemic Uncertainty: Principles, Paradoxes and Conflictual Loss

3 Principles of Epistemic Uncertainty

3.1 Data-Related Principle of Epistemic Uncertainty

3.2 Model-Related Principle of Epistemic Uncertainty

4 Conflictual Deep Ensembles

5 Empirical Analysis

Improving the Evaluation and Actionability of Explanation Methods for Multivariate Time Series Classification

2 Related Work.

2.1 Multivariate Time Series Classification

2.2 Attribution Methods for Multivariate Time Series Classification

2.3 Quantitative Evaluation of Attribution Methods for MTSC

2.4 Actionability of Explanation Methods

3 Background

3.1 Classifiers

3.2 XAI Methods

3.3 InterpretTime

3.4 Datasets

4 Methodology

4.1 MTS Chunking

4.2 Assessment Based on Ground Truth

4.3 Improving InterpretTime

4.4 Actionable XAI: Channel Selection for MTSC Using Attribution

5 Experiments

5.1 Validating InterpretTime Results

5.2 Improved XAI Evaluation Methodology

5.3 Real World Data

5.4 Actionability

Novel Node Category Detection Under Subpopulation Shift

2.1 PU Learning and Novel Category Detection

2.2 Subpopulation Shift and PU Learning

2.3 PU Learning on Graphs

2.4 Anomaly Detection and OOD Detection on Graphs

3 Problem Formulation

4 Recall-Constrained Optimization with Selective Link Prediction (RECO-SLIP)

4.1 Recall-Constrained Optimization

4.2 Selective Link Prediction

5.1 Experimental Setup

5.2 Results and Discussions

5.3 Auxiliary Experiments

SynODC: Utilizing the Syntactic Structure for Outlier Detection in Categorical Attributes

3 Syntactic Patterns

3.2 Expressive and Generalized Patterns

4 The System Architecture

4.1 Data Profiling and Numerical Attributes Pruning

4.2 Pattern Generation

4.3 Modified Min-Edit Distance (MMED)

4.4 Dominant Pattern Identification

4.5 Eliminating False Predictions

4.6 Time Complexity Analysis

5 Evaluation

5.1 Datasets

5.2 Baseline Methods

5.3 Experiment Setup

5.4 Results and Analysis

6 Conclusion and Future Work Directions

References.

FELIX: Automatic and Interpretable Feature Engineering Using LLMs

3 FELIX

3.1 Feature Generation

3.2 Feature Selection

3.3 Value Assignment

4.1 Datasets

4.2 Baselines

4.3 Experiment A: Feature Relevance

4.4 Experiment B: Sample Efficiency

4.5 Experiment C: Generalization Capabilities

4.6 Interpretability

Harnessing Superclasses for Learning from Hierarchical Databases

2 Related Works

3 Hierarchical Loss

3.1 Notation

3.2 Weighting Scheme

3.3 Hierarchical Loss

3.4 Comparison with Hierarchical Cross-Entropy

4 Experimental Protocol

4.1 Hierarchical Measures of Accuracy

4.2 Architectures, Baselines and Datasets

4.3 Training Protocol

5 Results and Analysis

5.1 Variable Training Set Size

5.2 Coarsening Accuracy Curves

5.3 Sensitivity to Hyper-parameters

A Proofs

Approximation Error of Sobolev Regular Functions with Tanh Neural Networks: Theoretical Impact on PINNs

2 Preliminary Background and Notations

2.1 PINNs and Functional Analysis

2.2 Notations

3 Approximation Bound for Sobolev Functions

3.1 General Bound for Tanh Networks

3.2 Instantiation for Our Proposed Methods for Choosing

3.3 Theoretical Impact on PINNs: the Case of Navier-Stokes PDEs

4 Differentiable Smoothing Windows

4.1 General Step Function and Differentiable Windows Phi_i

4.2 Tanh-Based Step Function

5 Novel Tanh Derivatives Analysis

A Theoretically Grounded Extension of Universal Attacks from the Attacker's Viewpoint

2 Preliminaries and Related Works

3 Generalization Guarantees: We Can Attack New Examples

3.1 From Universal Perturbation.

3.2 …to Generalized Universal Perturbations.

Notes:

Includes bibliographical references and index.

ISBN:

3-031-70359-6