Machine Learning and Knowledge Discovery in Databases : European Conference, ECML PKDD 2022, Grenoble, France, September 19–23, 2022, Proceedings, Part I

The multi-volume set LNAI 13713 until 13718 constitutes the refereed proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD 2022, which took place in Grenoble, France, in September 2022. The 236 full papers presented in these proceedings were carefully reviewed and selected from a total of 1060 submissions. In addition, the proceedings include 17 Demo Track contributions. The volumes are organized in topical sections as follows: Part I: Clustering and dimensionality reduction; anomaly detection; interpretability and explainability; ranking and recommender systems; transfer and multitask learning; Part II: Networks and graphs; knowledge graphs; social network analysis; graph neural networks; natural language processing and text mining; conversational systems; Part III: Deep learning; robust and adversarial machine learning; generative models; computer vision; meta-learning, neural architecture search; Part IV: Reinforcement learning; multi-agent reinforcement learning; bandits and online learning; active and semi-supervised learning; private and federated learning; . Part V: Supervised learning; probabilistic inference; optimal transport; optimization; quantum, hardware; sustainability; Part VI: Time series; financial machine learning; applications; applications: transportation; demo track. Preface Organization Contents – Part I Clustering and Dimensionality Reduction Pass-Efficient Randomized SVD with Boosted Accuracy*-4pt 1 Introduction 2 Preliminaries 2.1 Basics of Truncated SVD 2.2 Randomized SVD Algorithm with Power Iteration 2.3 Tropp's Single-Pass SVD Algorithm 3 Pass-Efficient SVD with Shifted Power Iteration 3.1 Randomized SVD with Fewer Passes 3.2 The Idea of Shifted Power Iteration 3.3 Update Shift in Each Power Iteration 3.4 Analysis of Computational Cost 4 Experimental Results 4.1 Error Metrics 4.2 Comparison with Basic Randomized SVD Algorithm 4.3 Comparison with Single-Pass SVD Algorithm 5 Conclusion References CDPS: Constrained DTW-Preserving Shapelets*-4pt 1 Introduction 2 Related Work 2.1 Shapelets 2.2 Learning Shapelets 2.3 Unsupervised Shapelets 2.4 Constrained Clustering 3 Constrained DTW-Preserving Shapelets 3.1 Definitions and Notations 3.2 Objective Function 3.3 CDPS Algorithm 4 Evaluation 4.1 Experimental Setup 4.2 Results 5 Discussion 5.1 Model Selection 6 Conclusions References Structured Nonlinear Discriminant Analysis*-4pt 1 Introduction 2 Prerequisites 2.1 Circulant Matrices 2.2 (Circulant) Principal Component Analysis 2.3 Linear Discriminant Analysis 3 Structured Discriminant Analysis 3.1 Circulant Discriminant Analysis 3.2 Computational Aspects for Circulant Structures 3.3 Harmonic Solutions 3.4 Truncated -Circulants 3.5 Non-cyclic Structures 4 Examples and Interpretation 4.1 (Quasi-)Stationary Data 4.2 Non-stationary Data 5 Conclusion References LSCALE: Latent Space Clustering-Based Active Learning for Node Classification*-4pt 1 Introduction 2 Problem Definition 3 Methodology 3.1 Active Learning Latent Space 3.2 Clustering Module 4 Experiments 4.1 Experimental Setting 4.2 Performance Comparison (RQ1) 4.3 Efficiency Comparison (RQ2) 4.4 Ablation Study (RQ3) 5 Related Work 6 Conclusion References Powershap: A Power-Full Shapley Feature Selection Method 1 Introduction 2 Related Work 3 Powershap 3.1 Powershap Algorithm 3.2 Automatic Mode 4 Experiments 4.1 Feature Selection Methods 4.2 Simulation Dataset 4.3 Benchmark Datasets 5 Results 5.1 Simulation Dataset 5.2 Benchmark Datasets 6 Discussion 7 Conclusion References Automated Cancer Subtyping via Vector Quantization Mutual Information Maximization*-4pt 1 Introduction 2 Related Work 3 Method 3.1 Problem Setting 3.2 Proposed Model 4 Experiments 4.1 Ground Truth Comparison 4.2 Controversial Label Comparison 4.3 Ablation Study 5 Discussion and Conclusion References Wasserstein t-SNE 1 Introduction 2 Methods 2.1 t-SNE 2.2 Wasserstein Metric 2.3 Linear Programming 2.4 Data 3 Results 3.1 Wasserstein t-SNE on Simulated Data 3.2 German Parliamentary Election 2017 4 Discussion References Nonparametric Bayesian Deep Visualization*-4pt 1 Introduction 2 Infinite Warped Mixture Model 3 Proposed Methods 3.1 Neural Network Gaussian Processes 3.2 NN-iWMM 3.3 Nonparametric Bayesian Deep Visualization 4 Bayesian Training 5 Simulation Study 6 Experiments on Real-World Data 7 Conclusion References FastDEC: Clustering by Fast Dominance Estimation*-12pt 1 Introduction 2 Related Work 3 Preliminaries 4 Proposed Framework 4.1 Direct k-NN Dominator (DkD) Detection 4.2 DC Dominance Estimation 4.3 Complexity Analysis and Implementation 5 Evaluation 5.1 Comparison on Artificial and Real-World Datasets 5.2 Robustness Testing 6 Conclusion References SECLEDS: Sequence Clustering in Evolving Data Streams via Multiple Medoids and Medoid Voting 1 Introduction 2 Preliminaries 3 SECLEDS: Sequence Clustering in Evolving Streams 3.1 Stable Cluster Definition via Multiple Medoids 3.2 Center of Mass Estimation 4 The SECLEDS Algorithm 5 Experimental Setup 6 Empirical Results 6.1 Use Case: Intelligent Network Traffic Sampling via SECLEDS 7 Conclusions References Knowledge Integration in Deep Clustering 1 Introduction 2 Related Work 3 Expert Loss for Knowledge Integration 3.1 Expert Knowledge Representation 3.2 Constraint-Satisfaction Score 3.3 Constraint-Satisfaction Score Computed by a WMC Problem 3.4 Decomposition of the Problem 3.5 Expert Loss 4 Integrating Knowledge in Deep Clustering Frameworks 4.1 IDEC-LK 4.2 SCAN-LK 5 Experiments 5.1 Experiment Settings 5.2 Experiments and Analysis for Clustering Quality 5.3 Experiments and Analysis for Constraint Satisfaction 6 Conclusion References Anomaly Detection ARES: Locally Adaptive Reconstruction-Based Anomaly Scoring 1 Introduction 2 Background 2.1 Autoencoders 2.2 Local Outlier Factor 3 Related Work 4 Methodology 4.1 Problem Definition 4.2 Statistical Interpretation of Reconstruction-Based Anomaly Detection 4.3 Motivation and Empirical Results 4.4 Adaptive Reconstruction Error-Based Scoring 4.5 Local Reconstruction Score 4.6 Local Density Score 5 Experiments 5.1 Datasets and Experimental Setup 5.2 Baselines 5.3 RQ1 (Accuracy) 5.4 RQ2 (Ablation Study) 6 Conclusion References R2-AD2: Detecting Anomalies by Analysing the Raw Gradient 1 Introduction 1.1 Related Work 2 Prerequisites 2.1 Activation Anomaly Analysis 3 R2-AD2 4 Experimental Setup 5 Evaluation 5.1 Known Anomalies 5.2 Noise Resistance 5.3 Number of Known Anomalies 5.4 Unknown Anomalies 5.5 Ablation Study 6 Discussion and Future Work 7 Summary References Hop-Count Based Self-supervised Anomaly Detection on Attributed Networks 1 Introduction 1.1 Our HCM Approach in a Nutshell 1.2 Summary of the Contributions 2 Related Work 3 Problem Definition 4 Hop-Count Based Model 4.1 Model Framework 4.2 Model Training 4.3 Model Inference for Anomaly Detection 5 Experiments 5.1 Datasets 5.2 Experimental Settings 5.3 Experimental Results 5.4 Parameter Analysis 5.5 Ablation Study 6 Conclusion References Deep Learning Based Urban Anomaly Prediction from Spatiotemporal Data*-10pt 1 Introduction 2 Related Work 2.1 Deep Learning Based Methods 2.2 Hybrid Learning (Graph + Deep Learning) Based Methods 3 Preliminaries 3.1 Notation 3.2 Problem Statement 4 Framework: UrbanAnom 4.1 Semantic Spatial (SS) Module 4.2 Context Aware Temporal (CAT) Module 4.3 Global Attention Module 4.4 Multi-layer Perceptron Based Prediction Module 5 Evaluation 5.1 Dataset 5.2 Parameter Settings 5.3 Performance Validation 5.4 Parameter Sensitivity 5.5 Evaluation of Variants 6 Conclusion References Detecting Anomalies with Autoencoders on Data Streams 1 Introduction 2 Problem Statement 3 Related Work 3.1 Offline Anomaly Detection 3.2 Online Anomaly Detection 4 Streaming Anomaly Detection with Autoencoders 5 Experiments 5.1 Data Streams 5.2 Setup 6 Results 7 Conclusion and Future Work References Anomaly Detection via Few-Shot Learning on Normality*-12pt 1 Introduction 2 Related Work 2.1 Deep Anomaly Detection 2.2 Information Bottleneck 3 Motivating Example 4 Prototype Data Description 5 Empirical Results 5.1 Setup 5.2 Comparative Analysis 5.3 Ablation Study 5.4 Graphical Analysis 6 Conclusion References Interpretability and Explainability Interpretations of Predictive Models for Lifestyle-related Diseases at Multiple Time Intervalspg*-12pt 1 Introduction 2 Related Work 2.1 Prediction of Diabetes Stages Using Medical Records 2.2 Prediction of Chronic Kidney Diseases Stages Using Medical Records 2.3 Interpretable Prediction of Diseases 3 Dataset 3.1 Structure and Attributes 3.2 Ethical Considerations 4 Method 4.1 Target Attributes 4.2 Prediction Tasks 4.3 Preprocessing 4.4 Training and Interpretation 5 Evaluation 5.1 Prediction Accuracy 5.2 HbA1c Included as a Feature 5.3 HbA1c Not Included as a Feature 5.4 Creatinine Included as a Feature 5.5 Creatinine Not Included as a Feature 6 Conclusion References Fair and Efficient Alternatives to Shapley-based Attribution Methods*-10pt 1 Introduction 2 State of the Art 2.1 The Attribution Problem 2.2 Attribution Using Feature Coalisation Analysis 2.3 Attribution Based on Gradient Analysis 3 Fair-Efficient-Symmetric Perturbations-based AMs 3.1 The Equal Surplus Value 3.2 FESP 4 Experiments 4.1 Image Classification: Protocols and Results 4.2 Text Classification: Protocols and Results 4.3 Discussions 5 Conclusion References SMACE: A New Method for the Interpretability of Composite Decision Systems*-10pt 1 Introduction 2 Related Work 3 Challenges 4 SMACE 4.1 Setting 4.2 Assumptions 4.3 Overview 4.4 Explaining the Results of the Models 4.5 Explaining the Rule-Based Decision 4.6 Overall Explanations 5 Evaluation 5.1 Qualitative Analysis 5.2 Sanity Check 6 Conclusion and Future Work References Calibrate to Interpret*-10pt 1 Introduction 2 Related Works 3 Problem Statement and Other Related Works 3.1 Calibration 3.2 Interpretation Methods 4 Evaluation of Calibration's Impact on Interpretation 4.1 Objectives 4.2 Experimental Setup 4.3 Does Calibration Impact Interpretations? 4.4 Does Calibration Improve the Faithfulness of Interpretation Methods? 4.5 Are Saliency Maps with Calibration More Human-Friendly? 4.6 In Depth Analysis of Meaningful Perturbation 4.7 Discussions 5 Conclusions and Future Works References Knowledge-Driven Interpretation of Convolutional Neural Networks 1 Introduction 2 Related Works 3 Ontology-Driven Semantic Alignment 3.1 High-Level Concept Masks 3.2 Alignment Measure 3.3 Direction Learning 3.4 Neural Circuits 4 Results 4.1 Unit Semantic Alignment 4.2 Direction Learning 5 Conclusion References Neural Networks with Feature Attribution and Contrastive Explanations 1 Introduction 1.1 Contrastive vs. Counterfactual 2 Related Work 2.1 Contrastive Explanations 2.2 Counterfactual Explanations 2.3 Post-hoc Non-contrastive Explanations 3 Contrastive Explanation Generation 3.1 Neural Nets with Feature Attributions and Contrastive Explanations 3.2 Joint Objective 3.3 Explanations 4 Experiments 4.1 Setup 4.2 Explainability Metrics 4.3 Evaluating Why p? 4.4 Evaluating Why p and Not q? 4.5 Deep Learned Features 4.6 Discussion 5 Conclusion References Explaining Predictions by Characteristic Rules 1 Introduction 2 Related Work 3 From Local Explanations to General Characteristic Rules 3.1 Explanation Mining and Rules Selection 3.2 Discriminative vs. Characteristic Rules 4 Empirical Evaluation 4.1 Experimental Setup 4.2 Baseline Experiments 4.3 Comparing Discriminative and Characteristic Rules 4.4 Comparing Local Explanation Techniques 5 Concluding Remarks References Session-Based Recommendation Along with the Session Style of Explanation 1 Introduction 2 Related Work 3 Our Proposed Method 3.1 Meta Path-Based Similarity 3.2 Recommendation List Creation by Considering One Meta Path 3.3 Recommendation List Creation by Using Multiple Meta Paths 4 Recommendation Strategies and Single Explanations 5 Hybrid Meta Path-Based Explanation 6 Experimental Evaluation 6.1 Real-Life Datasets 6.2 Evaluation Protocol and Metrics 6.3 xPathSim Sensitivity Analysis 6.4 Comparison with Other Methods 6.5 User Study 7 Conclusion References ProtoMIL: Multiple Instance Learning with Prototypical Parts for Whole-Slide Image Classification 1 Introduction 2 Related Works 2.1 Multiple Instance Learning 2.2 Explainable Artificial Intelligence 3 ProtoMIL 4 Experiments 4.1 Bisque Breast Cancer and Colon Cancer Datasets 4.2 Camelyon16 Dataset 4.3 TCGA-NSCLC Dataset 4.4 TCGA-RCC Dataset 4.5 Pruning 4.6 Interpretability of MIL Methods 5 Discussion and Conclusions 5.1 Limitations 5.2 Negative Impact References VCNet: A Self-explaining Model for Realistic Counterfactual Generation 1 Introduction 2 Related Work 3 Backgrounds 3.1 Variational Autoencoder (VAE) 3.2 Conditional Variational Autoencoder (cVAE) 4 A Join Training Model 4.1 VCNet Architecture 4.2 Loss Function and Training Procedure 4.3 Counterfactual Generation 5 Experiments and Results 5.1 cVAE for Counterfactual Generation 5.2 Comparison Between VCNet and CounterNet 5.3 Impact of Join-Training on Counterfactual Quality 5.4 Qualitative Results on MNIST Dataset 6 Conclusion References Ranking and Recommender Systems A Recommendation System for CAD Assembly Modeling Based on Graph Neural Networks 1 Recommending Components in Assembly Modeling 2 Graph Neural Networks 3 Graph-Based Recommendations for Assemblies Using Pretrained Embeddings 3.1 Pretraining of Component Embeddings (comp2vec) 3.2 Generating Data Instances for Component Recommendation 3.3 Frequency-Based Baseline Model 4 Experiments 4.1 Experimental Setup 4.2 How Well Can GNNs Learn the Task at Hand? 4.3 Are Component Embeddings Better Than One-Hot Node Features? 4.4 Comparing GAT and GCN 5 Conclusion and Future Work References AD-AUG: Adversarial Data Augmentation for Counterfactual Recommendation 1 Introduction 2 Related Work 2.1 Autoencoder-Based CF 2.2 Counterfactual Data Augmentation 2.3 Adversarial Training 3 Preliminaries 3.1 Problem Definition 3.2 Autoencoder CF Framework 4 The Proposed Model 4.1 Model Overview 4.2 Data-Oriented Counterfactual Learning 4.3 Model-Oriented Counterfactual Learning 4.4 Implementation of Augmenter Model 4.5 Curriculum Adversarial Learning 5 Experiment 5.1 Experimental Settings 5.2 Experimental Result 6 Conclusion References Bi-directional Contrastive Distillation for Multi-behavior Recommendation 1 Introduction 2 Data Analysis 3 Our Proposed Model 3.1 Multi-behavior GCN 3.2 Bi-directional Contrastive Distillation 3.3 Prediction and Learning 4 Experiments 4.1 Datasets 4.2 Comparison Methods 4.3 Parameter Settings 4.4 Evaluation Metrics 4.5 Performance Comparison 4.6 Ablation Study 4.7 Parameter Analysis 5 Related Work 5.1 Multi-behavior Recommendation 5.2 Contrastive Distillation in Recommendations 6 Conclusions and Future Work References Improving Micro-video Recommendation by Controlling Position Bias 1 Introduction 2 Related Work 3 Our Model 3.1 Overview 3.2 Sequence Encoder 3.3 Contrastive Encoder 3.4 Prediction and Loss Function 3.5 Complexity Analysis 3.6 Discussion 4 Experimental Evaluation 4.1 Experimental Setup 4.2 Performance Comparison 4.3 Ablation Study 4.4 Impact of Contrastive Learning Strategies 5 Conclusion References Mitigating Confounding Bias for Recommendation via Counterfactual Inference 1 Introduction 2 Methodology 2.1 Preliminary 2.2 Causal Look in Recommendation 2.3 Deconfounded Analysis 2.4 Deconfounded Recommendation Model 3 Experiments 3.1 Experimental Settings 3.2 Performance Comparison (RQ1) 3.3 Case Study (RQ2) 3.4 Deconfounding Capability (RQ3) 4 Related Work 4.1 Debiasing in Recommender Systems 4.2 Deconfounded in Recommender Systems 4.3 Causal Recommendation 5 Conclusion and Future Work References Recommending Related Products Using Graph Neural Networks in Directed Graphs 1 Introduction 2 Related Work 3 Related Product Recommendation Problem 4 Proposed Framework 4.1 Product Graph Construction 4.2 DAEMON: Proposed GNN Model 5 Experiments 5.1 Experimental Setting 5.2 EQ1. Node Recommendation Task on Co-purchase Data 5.3 [EQ2, EQ3.] Link Prediction Tasks on Co-purchase Data 5.4 [EQ4, EQ5, EQ6.] Ablation Study on G1 Graph 5.5 Online Platform Performance 6 Conclusion and Future Work References A U-Shaped Hierarchical Recommender by Multi-resolution Collaborative Signal Modeling 1 Introduction 2 Related Work 3 Methodology 3.1 The Basic Collaborative Filtering Model 3.2 Modeling Implicit Hierarchies 3.3 UGCN Recommender 3.4 Theoretical Analysis for UGCN Recommender 4 Experiments 4.1 Experimental Settings 4.2 Prediction Accuracy Comparison (QR1) 4.3 Personalization Comparison (QR2) 4.4 Hyper Parameter Analysis (QR3) 5 Conclusion References Basket Booster for Prototype-based Contrastive Learning in Next Basket Recommendation 1 Introduction 2 Related Work 2.1 Next Basket Recommendation 2.2 Contrastive Learning 3 The Proposed Method 3.1 Problem Statement 3.2 BPCL 4 Experiments 4.1 Experiments Settings 4.2 Performance Comparison 4.3 Ablation Study 4.4 Hyper-Parameter Study 5 Conclusion References Graph Contrastive Learning with Adaptive Augmentation for Recommendation 1 Introduction 2 Preliminaries 3 Methodology 3.1 The Contrastive Learning Framework 3.2 Adaptive Augmentation 3.3 Contrastive Learning 3.4 Multi-task Training 4 Experiments 4.1 Experimental Setup 4.2 Performance Comparison (RQ1) 4.3 Further Study of GCARec 5 Related Work 5.1 Graph-based Recommendation 5.2 Self-supervised Learning in Recommender Systems 6 Conclusion and Future Work References Multi-interest Extraction Joint with Contrastive Learning for News Recommendation 1 Introduction 2 Related Work 2.1 Personalized News Recommendation 2.2 Contrastive Learning 3 Methodology 3.1 News Encoder 3.2 Multi-interest User Encoder 3.3 Multi-interest Graph-Enhanced Module 3.4 Multi-interest Contrastive Learning Module 3.5 Adaptive User Aggregator and Click Predictor 3.6 Model Training 4 Experiment 4.1 Dataset and Experimental Settings 4.2 Performance Evaluation 4.3 Ablation Study 4.4 Hyper-Parameters Analysis 4.5 Statistic Analysis 5 Conclusion References Transfer and Multitask Learning On the Relationship Between Disentanglement and Multi-task Learning 1 Introduction 2 Related Work 2.1 Disentanglement 2.2 Multi-task Learning 3 Methods 3.1 Dataset Creation 3.2 Models 3.3 Disentanglement Metrics 4 Results and Discussion 4.1 Does Hard Parameter Sharing Encourage Disentanglement? 4.2 What Are the Properties of the Learned Representations? 4.3 Does Disentanglement Help in Training Multi-task Models? 5 Conclusions References InCo: Intermediate Prototype Contrast for Unsupervised Domain Adaptation 1 Introduction 2 Related Work 2.1 Unsupervised Domain Adaptation 2.2 Contrastive Learning 2.3 Contrastive Learning for Domain Adaptation 3 Method 3.1 Problem Definition and Overall Idea 3.2 Revisit of Contrastive Learning 3.3 Intra-Domain Contrastive Learning 3.4 Inter-Domain Contrastive Learning 3.5 Other Losses 3.6 Overall 4 Experiments 4.1 Datasets 4.2 Setup 4.3 Baselines 4.4 Results 4.5 Insight Analysis 5 Conclusion References Fast and Accurate Importance Weighting for Correcting Sample Bias 1 Introduction 2 Problem Setting and Proposed Approach 2.1 Learning Scenario 2.2 MMD 2.3 Importance Weighting Network 3 Related Work 4 Experiments 4.1 IWN Settings 4.2 Competitors Settings 4.3 Synthetic Dataset 4.4 UCI Datasets 4.5 Impact of Network Architecture and Batch Size 5 Conclusion References Overcoming Catastrophic Forgetting via Direction-Constrained Optimization 1 Introduction 2 Related Work 3 Loss Landscape Properties 4 Algorithm 4.1 Loss Function 4.2 Reduced Linear Autoencoders 4.3 Compression of Autoencoders 4.4 Resulting Algorithm 5 Experiments 5.1 Data Sets and Architectures 5.2 Training Details 5.3 Hyperparameters 5.4 Metric and Results 6 Conclusion References Newer is Not Always Better: Rethinking Transferability Metrics, Their Peculiarities, Stability and Performance 1 Introduction 2 Transferability Setup 3 Related Work 4 Improved Estimation of H-score 4.1 Proposed Transferability Measure 4.2 Challenges of Comparing H(f) Across Source Models/Layers 4.3 Efficient Computation for Small Target Data 5 A Closer Look at NCE, LEEP and NLEEP Measures 6 Experiments 6.1 Case Study: Vision Models 6.2 Case Study: Graph Neural Networks 6.3 Timing Comparison Between LogME and H(f) 7 Conclusion References Learning to Teach Fairness-Aware Deep Multi-task Learning 1 Introduction 2 Related Work 3 Problem Setting and Basic Concepts 3.1 Fairness Definition and Metric 3.2 Vanilla Multi-task Learning (MTL) 3.3 Fairness-Aware Multi-task Learning (FMTL) 3.4 Deep Q-learning (DQN) and Multi-tasking DQN (MT-DQN) 4 Learning to Teach Fairness-Aware Multi-tasking 4.1 Dynamic Loss Selection Formulation 4.2 L2T-FMT Algorithm 4.3 Student Network 4.4 Teacher Network 5 Experiments 5.1 Experimental Setup 5.2 Overall Fairness-Accuracy Evaluation 5.3 Performance Distribution over the Tasks 5.4 Dynamic Loss Selection 6 Conclusion References Author Index