Machine Learning and Knowledge Discovery in Databases. Research Track European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part II

• Main Authors: Oliver, Nuria, Pérez-Cruz, Fernando, Kramer, Stefan, Read, Jesse, Lozano, Jose A
• Format: eBook
• Language: English
• Published: Cham Springer International Publishing AG 2021; Springer International Publishing
• Edition: 1
• Series: Lecture Notes in Computer Science
• Subjects: Artificial intelligence-Data processing
• ISBN: 3030865193; 9783030865191

The multi-volume set LNAI 12975 until 12979 constitutes the refereed proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD 2021, which was held during September 13-17, 2021. The conference was originally planned to take place in Bilbao, Spain, but...

Table of Contents:

• Intro -- Preface -- Organization -- Contents - Part II -- Generative Models -- Non-exhaustive Learning Using Gaussian Mixture Generative Adversarial Networks -- 1 Introduction -- 2 Related Work -- 3 Background -- 4 Methodology -- 5 Experiments -- 6 Conclusion -- References -- Unsupervised Learning of Joint Embeddings for Node Representation and Community Detection -- 1 Introduction -- 2 Related Work -- 2.1 Community Detection -- 2.2 Node Representation Learning -- 2.3 Joint Community Detection and Node Representation Learning -- 3 Methodology -- 3.1 Problem Formulation -- 3.2 Variational Model -- 3.3 Design Choices -- 3.4 Practical Aspects -- 3.5 Complexity -- 4 Experiments -- 4.1 Synthetic Example -- 4.2 Datasets -- 4.3 Baselines -- 4.4 Settings -- 4.5 Discussion of Results -- 4.6 Hyperparameter Sensitivity -- 4.7 Training Time -- 4.8 Visualization -- 5 Conclusion -- References -- GraphAnoGAN: Detecting Anomalous Snapshots from Attributed Graphs -- 1 Introduction -- 2 Related Work -- 3 Problem Definition -- 4 Proposed Algorithm -- 4.1 GAN Modeling -- 4.2 Architecture -- 4.3 Training Procedure -- 5 Datasets -- 6 Experiments -- 6.1 Baselines -- 6.2 Comparative Evaluation -- 6.3 Side-by-Side Diagnostics -- 7 Conclusion -- References -- The Bures Metric for Generative Adversarial Networks -- 1 Introduction -- 2 Method -- 3 Empirical Evaluation of Mode Collapse -- 3.1 Artificial Data -- 3.2 Real Images -- 4 High Quality Generation Using a ResNet Architecture -- 5 Conclusion -- References -- Generative Max-Mahalanobis Classifiers for Image Classification, Generation and More -- 1 Introduction -- 2 Background and Related Work -- 2.1 Energy-Based Models -- 2.2 Alternatives to the Softmax Classifier -- 3 Methodology -- 3.1 Approach 1: Discriminative Training -- 3.2 Approach 2: Generative Training -- 3.3 Approach 3: Joint Training
• 3.3 Learning Smoothing Distribution for Midpoints -- 3.4 Optimization -- 4 Experiments -- 4.1 Datasets, Baselines, and Settings -- 4.2 Predictive Accuracy -- 4.3 Ablation Studies -- 4.4 Uncertainty Label and Conservative Classification -- 4.5 Testing on Out-of-Distribution Data -- 5 Conclusion and Future Work -- References -- Learning Weakly Convex Sets in Metric Spaces -- 1 Introduction -- 2 Preliminaries -- 3 Weak Convexity in Metric Spaces -- 3.1 Some Basic Properties of Weakly Convex Sets -- 4 Learning in the Extensional Problem Setting -- 4.1 Application Scenario: Vertex Classification -- 5 The Intensional Problem Setting -- 5.1 Learning Weakly Convex Boolean Functions -- 5.2 Learning Weakly Convex Axis-Aligned Hyperrectangles -- 6 Concluding Remarks -- References -- Disparity Between Batches as a Signal for Early Stopping -- 1 Introduction -- 2 Related Work -- 3 Generalization Penalty -- 4 Gradient Disparity -- 5 Early Stopping Criterion -- 6 Discussion and Final Remarks -- References -- Learning from Noisy Similar and Dissimilar Data -- 1 Introduction -- 2 Problem Setup -- 3 Loss Correction Approach -- 4 Weighted Classification Approach -- 5 Experiments -- 6 Conclusion and Future Work -- References -- Knowledge Distillation with Distribution Mismatch -- 1 Introduction -- 2 Related Works -- 3 Framework -- 3.1 Problem Definition -- 3.2 Proposed Method KDDM -- 4 Experiments and Discussions -- 4.1 Datasets -- 4.2 Baselines -- 4.3 Results on MNIST -- 4.4 Results on CIFAR-10 -- 4.5 Results on CIFAR-100 -- 4.6 Distillation When Teacher-Data and Student-Data Are Identical -- 5 Conclusion -- References -- Certification of Model Robustness in Active Class Selection -- 1 Introduction -- 1.1 Active Class Selection Constitutes a Domain Gap -- 1.2 A Qualitative Intuition from Information Theory -- 2 A Quantitative Perspective from Learning Theory
• 2.1 Quantification of the Domain Gap -- 2.2 Certification of Domain Robustness for Binary Predictors -- 3 Experiments -- 3.1 Binary (, ) Certificates Are Tight -- 3.2 Binary (, ) Certificates in Astro-Particle Physics -- 4 Related Work -- 5 Conclusion -- References -- Graphs and Networks -- Inter-domain Multi-relational Link Prediction -- 1 Introduction -- 2 Preliminary -- 2.1 RESCAL -- 2.2 Optimal Transport -- 2.3 Maximum Mean Discrepancy -- 3 Problem Setting and Proposed Method -- 3.1 Problem Setting -- 3.2 Proposed Objective Function -- 4 Experiments -- 4.1 Datasets -- 4.2 Evaluation Methods and Baselines -- 4.3 Implementation Details -- 4.4 Experimental Results -- 5 Related Work -- 6 Conclusion and Future Work -- References -- GraphSVX: Shapley Value Explanations for Graph Neural Networks -- 1 Introduction -- 2 Related Work -- 3 Preliminary Concepts and Background -- 3.1 Graph Neural Networks -- 3.2 The Shapley Value -- 4 A Unified Framework for GNN Explainers -- 5 Proposed Method -- 5.1 Mask and Graph Generators -- 5.2 Explanation Generator -- 5.3 Decomposition Model -- 5.4 Efficient Approximation Specific to GNNs -- 5.5 Desirable Properties of Explanations -- 6 Experimental Evaluation -- 6.1 Synthetic and Real Datasets with Ground Truth -- 6.2 Real-World Datasets Without Ground Truth -- 7 Conclusion -- References -- Multi-view Self-supervised Heterogeneous Graph Embedding -- 1 Introduction -- 2 Related Work -- 2.1 Self-supervised Learning on Graphs -- 2.2 Heterogeneous Graph Embedding -- 3 The Proposed Model -- 3.1 Model Framework -- 3.2 Heterogeneous Context Encoding -- 3.3 Multi-view Contrastive Learning -- 4 Experiment -- 4.1 Experimental Setup -- 4.2 Node Classification -- 4.3 Link Prediction -- 4.4 Ablation Study -- 4.5 Visualization -- 5 Conclusion -- References
• 3.4 GMMC for Inference -- 4 Experiments -- 4.1 Hybrid Modeling -- 4.2 Calibration -- 4.3 Out-Of-Distribution Detection -- 4.4 Robustness -- 4.5 Training Stability -- 4.6 Joint Training -- 5 Conclusion and Future Work -- References -- Gaussian Process Encoders: VAEs with Reliable Latent-Space Uncertainty -- 1 Introduction -- 1.1 Contributions -- 2 Background -- 2.1 Variational Autoencoder -- 2.2 Latent Variance Estimates of NN -- 2.3 Mismatch Between the Prior and Approximate Posterior -- 3 Methodology -- 3.1 Gaussian Process Encoder -- 3.2 The Implications of a Gaussian Process Encoder -- 3.3 Out-of-Distribution Detection -- 4 Experiments -- 4.1 Log Likelihood -- 4.2 Uncertainty in the Latent Space -- 4.3 Benchmarking OOD Detection -- 4.4 OOD Polution of the Training Data -- 4.5 Synthesizing Variants of Input Data -- 4.6 Interpretable Kernels -- 5 Related Work -- 6 Conclusion -- References -- Variational Hyper-encoding Networks -- 1 Introduction -- 2 Variational Autoencoder (VAE) -- 3 Variational Hyper-encoding Networks -- 3.1 Hyper-auto-encoding Problem -- 3.2 Hyper-encoding Problem -- 3.3 Minimum Description Length -- 3.4 Compact Hyper-decoder Architecture -- 3.5 Applications -- 4 Experiments -- 4.1 Data Sets -- 4.2 Model Settings -- 4.3 Model Behavior -- 4.4 Robust Outlier Detection -- 4.5 Novelty Discovery -- 5 Related Work -- 6 Conclusion -- References -- Principled Interpolation in Normalizing Flows -- 1 Introduction -- 2 An Intuitive Solution -- 3 Normalizing Flows -- 4 Base Distributions on p-Norm Spheres -- 4.1 The Case p = 1 -- 4.2 The Case p = 2 -- 5 Experiments -- 5.1 Performance Metrics and Setup -- 5.2 Data -- 5.3 Architecture -- 5.4 Quantitative Results -- 5.5 Qualitative Results -- 6 Related Work -- 7 Conclusion -- References -- CycleGAN Through the Lens of (Dynamical) Optimal Transport -- 1 Introduction
• Semantic-Specific Hierarchical Alignment Network for Heterogeneous Graph Adaptation
• 2 Desiderata for UDT and Analysis of CycleGAN -- 2.1 What Should Be the Properties of a UDT Solution? -- 2.2 CycleGAN Is Biased Towards Low Energy Transformations -- 3 UDT as Optimal Transport -- 3.1 A (Dynamical) OT Model for UDT -- 3.2 Regularity of OT Maps -- 3.3 Computing the Inverse -- 4 A Residual Instantiation from Dynamical OT -- 4.1 Linking the Dynamical Formulation with CycleGAN -- 4.2 A Typical UDT Task -- 4.3 Imbalanced CelebA task -- 5 Related Work -- 6 Discussion and Conclusion -- References -- Decoupling Sparsity and Smoothness in Dirichlet Belief Networks -- 1 Introduction -- 2 Preliminary Knowledge -- 3 Sparse and Smooth Dirichlet Belief Networks -- 3.1 Generative Process -- 3.2 Necessity of Fixing biK(l)=1 -- 4 Related Work -- 5 ssDirBN for Relational Modelling -- 5.1 Inference -- 5.2 Experimental Results -- 6 Conclusion -- References -- Algorithms and Learning Theory -- Self-bounding Majority Vote Learning Algorithms by the Direct Minimization of a Tight PAC-Bayesian C-Bound -- 1 Introduction -- 2 Majority Vote Learning -- 2.1 Notations and Setting -- 2.2 Gibbs Risk, Joint Error and C-Bound -- 2.3 Related Works -- 3 PAC-Bayesian C-Bounds -- 3.1 An Intuitive Bound-McAllester's View -- 3.2 A Tighter Bound-Seeger's View -- 3.3 Another Tighter Bound-Lacasse's View -- 4 Self-bounding Algorithms for PAC-Bayesian C-Bounds -- 4.1 Algorithm Based on McAllester's View -- 4.2 Algorithm Based on Seeger's View -- 4.3 Algorithm Based on Lacasse's View -- 5 Experimental Evaluation -- 5.1 Empirical Setting -- 5.2 Analysis of the Results -- 6 Conclusion and Future Work -- References -- Midpoint Regularization: From High Uncertainty Training Labels to Conservative Classification Decisions -- 1 Introduction -- 2 Related Work -- 3 Label Smoothing over Midpoint Samples -- 3.1 Preliminaries -- 3.2 Midpoint Generation