EDBT-ICDT 2018

整理筆記時，翻到兩年前去研討會的筆記
想說放著也不會增值，就整理出來了
雖然大部分的內容的印象都已經有點模糊了
不過就加減把當初的筆記湊起來

當時我覺得最有趣的論文是 Interactive Rule Refinement for Fraud Detection.
不過竟然沒有做到太多筆記

Day 1 - Keynote

• In theoretical CS
  • Polynomial time → easy/fast
    • However, that's not always the case
    • e.g., \(O(n^{100})\)
    • When n grows, even \(O(n^2)\) is not efficient
• We're stuck on many problems even just in \(O(n^2)\)
• No \(N^{2-\epsilon}\) time algorithms known for
  • String matching
  • computational geometry
  • graph problem in sparse graphs
  • many problems from database
  • many other problems
• Why are we stuck?
  • The traditional hardness in complexity tells us little about runtime
  • fine-grained hardness idea
    1. identify key hard problem
    2. ......

Large Scale Machine Learning: Where Do Relational Systems Fit In? (by Chris Jermaine)

Currently, ML community cares about new models instead of theory and fundamental ML design

ML vs AI

• ML is one approach to AI
• Classic AI: a programmer/expert imparting knowledge to a system
• ML is fundamentally statistical

Intro to ML

• Distributed ML
  • Most ML systems use a "parameter server" model
    • Essentially a distributed key-value pair
  • Negatives
    • Parameter server compute model very limiting
• Data Parallel ML
  • Each compute server runs same computation on different data
  • Global state updated via aggregation

Want to scale out to speed up learning?

• scale out ineffective in data parallel param server
  • no easy way to add machines and have a graph execute faster
• Only easy way to scale out is to add compute servers

Take-Home Point

• Current ML systems are easily applicable only to
  • Relatively small model problems
  • That is run on a single machine

Detecting Database File Tampering through Page Carving

• Attack Vector: File Tampering
  • Occurs at the OS level → outside DBMS control
    • Bypass DBMS control
• Page Deconstruction
  • Page Header
    • Checksum
    • PageID
    • Row Count
• DBStorageAuditor
  • Goal: find inconsistency in storage
    • which is created by direct file manipulation

Extracting Statistical Graph Features for Accurate and Efficient Time Series Classification

• Time series: Any data that is ordered
• Time Series Classification
  • similarity-based kNN (e.g., kNN-ED, kNN-DTW)
    • similarity can be unreliable
  • Shaplets
    • high computation complexity
• Why multiscale
  • sometimes global features are more important while sometimes local features are more important
  • in this research, both global and local are considered
• Visibility Graphs
• Multiscale Visibility Graphs