Maverick: Discovering Exceptional Facts from Knowledge Graphs - PowerPoint PPT Presentation

Maverick: Discovering Exceptional Facts from Knowledge Graphs 12/03/19 Paper published in Proc. ACM SIGMOD International Conference on Management of Data, 2018. Presented by: Juan Carrillo Candidate for MASc. in Computer Software Department of Electrical & Computer Engineering University of Waterloo

Agenda 1. Introduction 2. Maverick core features 3. Experiments 4. Conclusions 5. Discussion Maverick: Discovering Exceptional Facts from PAGE 2 Knowledge Graphs

Introduction 1 Maverick: Discovering Exceptional Facts from PAGE 3 Knowledge Graphs

1. Introduction From knowledge graphs to exceptional facts Maverick: Discovering Exceptional Facts from PAGE 4 Knowledge Graphs

1. Introduction The problem, and the Maverick approach Knowledge graphs (Linked Data) Maverick approach Pattern generator Fact reporter Exceptionality Context evaluator evaluator Automated detection of exceptional facts Manually designed queries Maverick: Discovering Exceptional Facts from PAGE 5 Knowledge Graphs

1. Introduction Related background 2010 2016 2018 2018 Outlying aspect Maverick: Discovering Exceptional Facts Outlier detection from Knowledge Graphs mining J Gao - 2010 F Angiulli - 2016 SIGMOD’18 VLDB’18 On community Outlying property Comprehensive High-level description outliers and their detection with description and and demo efficient detection numerical math basis in information attributes networks Maverick: Discovering Exceptional Facts from PAGE 6 Knowledge Graphs

Maverick 2 core features Maverick: Discovering Exceptional Facts from PAGE 7 Knowledge Graphs

2. Maverick core features Entity, context, pattern Maverick: Discovering Exceptional Facts from PAGE 8 Knowledge Graphs

2. Maverick core features The overall framework Maverick: Discovering Exceptional Facts from PAGE 9 Knowledge Graphs

2. Maverick core features Main Algorithm Maverick: Discovering Exceptional Facts from PAGE 10 Knowledge Graphs

2. Maverick core features Description of components Exceptionality Context Evaluator Pattern generator Evaluator ● Takes the entity of ● Uses beam search to ● Uses a graph query interest and its look for promising system (Neo4j) contexts patterns ● Takes a pattern as input and returns the ● Looks for the k ● Implements domain subspaces with specific heuristics matches highest scores ● Beam width can be ● Agnostic to query ● Implements scoring tuned to requirements processing system functions Maverick: Discovering Exceptional Facts from PAGE 11 Knowledge Graphs

Experiments 3 Maverick: Discovering Exceptional Facts from PAGE 12 Knowledge Graphs

3. Experiments Experimental setup Single node: 16-core, 32GB RAM Datasets WCGoals Methods compared 49.078 nodes, 158.114 edges, 13 different Beam-Rdm ▪ edge labels, and 11 entity types. Beam-Opt ▪ OscarWinners Beam-Conv ▪ Breadth-First ▪ 42.148 nodes, 63.187 edges, 24 distinct edge labels, and 13 entity types. Maverick: Discovering Exceptional Facts from PAGE 13 Knowledge Graphs

3. Experiments Efficiency Maverick: Discovering Exceptional Facts from PAGE 14 Knowledge Graphs

3. Experiments Efficiency Maverick: Discovering Exceptional Facts from PAGE 15 Knowledge Graphs

3. Experiments Effectiveness Maverick: Discovering Exceptional Facts from PAGE 16 Knowledge Graphs

Conclusions 4 Maverick: Discovering Exceptional Facts from PAGE 17 Knowledge Graphs

4. Conclusions Takeaways and paper contributions ✓ The authors model an exceptional fact as a context-pattern pair on a knowledge graph ✓ Exponential complexity of search is handled using beam search ✓ The framework is adaptable to domain specific requirements Maverick: Discovering Exceptional Facts from PAGE 18 Knowledge Graphs

Thanks for your attention Maverick: Discovering Exceptional Facts from PAGE 19 Knowledge Graphs

Discussion 5 Maverick: Discovering Exceptional Facts from PAGE 20 Knowledge Graphs

5. Discussion Research 1. What other heuristics could be proposed in addition to the two presented in the paper? Design requirements for a third heuristic? 2. How Maverick would perform over a completely different dataset? Different proportions among nodes, edges, edge labels, and entity types. 3. What if we add attributes to the nodes and edges? Constraints 4. How to adapt Maverick to work over multiple/linked knowledge graphs? Industry 5. What is an example of an application over Google knowledge graph? Maverick: Discovering Exceptional Facts from PAGE 21 Knowledge Graphs