Recently, the Artificial Intelligence (AI) field has been experiencing a
resurgence. AI broadly covers a wide swath of techniques which include
logic-based approaches, probabilistic graphical models, and machine
learning/deep learning approaches. Advances in hardware capabilities, such as Graphics
Processing Units (GPUs), software components (e.g., accelerated
libraries, programming frameworks), and systems infrastructures (e.g.,
GPU-enabled cloud providers) has led to a wide-spread adaptation of AI
techniques to a variety of domains. Examples of such domains
include image classification, autonomous driving, automatic speech
recognition (ASR) and conversational systems (chatbots). AI solutions
not only support multiple datatypes (e.g., free text, images, or
speech), but are also available in various configurations, from personal
devices to large-scale distributed systems.
In spite of the wide ranging applications of AI techniques, its
interactions with the data management systems remains in infancy. At
present, a majority of database management systems (DBMS) are being
used primarily as a repository for feeding input data and storing
results. Recently, there has been some activity in using AI techniques
in data management systems, e.g., enabling natural language interfaces to relational
databases and applying machine learning techniques for query
optimizations. However, a lot more needs to done to fully exploit the
power of AI for data management workloads.
We propose to organize a one-day workshop that will bring together
people from academia and industry to discuss various ways of
integrating AI techniques with data management systems. The primary
goal of the proposed workshop is to explore opportunities for AI
techniques for enhancing different components of the data management
systems, e.g., user interfaces, tooling, performance optimizations,
new query types, and workloads. Special emphasis would be given to
transparent exploitation of AI techniques using existing data
management for enterprise class workloads. We hope this workshop
will identify important areas of research and spur new efforts in this
emerging field.
The goal of the workshop is to take a holistic view of various AI technologies and
investigate how they can be applied to different component of an end-to-end data management
pipeline. Special emphasis would be given to how AI techniques could be used for enhancing
user experience by reducing complexity in tools, or providing newer insights, or providing
better user interfaces. Topics of interest include, but are not restricted to:
- Characterizing different AI approaches: Logic-based, Probabilistic Graphical models, and machine learning/deep learning approaches
- Evaluation of different learning approaches: unsupervised
learning, supervised or reinforced learning, transfer learning,
zero-shot learning, adversarial networks, and deep probabilistic models
- New AI-enabled business intelligence (BI) queries for relational databases
- Natural language queries and chatbot interfaces
- Natural language result summarization
- Impact of the lack of model interpretability
- Evaluating quality of approximate results from AI-enabled queries
- Supporting multiple datatypes (e.g., images or time-series data)
- Supporting semi-structured, streaming, and graph databases
- Reasoning over knowledge bases
- Data exploration and visualization
- Integrating structured and unstructured data sources
- AI-enabled data integration strategies
- Re-inforcement Learning for Database Tuning
- Impact of AI on tooling, e.g., ETL or data cleaning
- Performance implications of AI-enabled queries
- Case studies of AI-accelerated workloads
- Social Implications of AI-enabled database
Learning Data Systems Components
Tim Kraska, MIT
Abstract:
Database systems have a long history of carefully selecting efficient algorithms, e.g., a merge vs hash-join, based on data statistics. Yet, existing databases remain general purpose systems and are not built on a case-by-case basis for the specific needs and data characteristics of a user. For example, if the goal would be to build a highly tuned system to store and query fixed-length records with continuous integers keys (e.g., the keys 1 to 100M), one would not use a conventional design. In this setting, a traditional B-Tree to index the key would not make much sense, since they key itself can be used as an offset, making it a constant $O(1)$ operation to look-up any key. Similar for sorting; if the keys come from a dense integer domain, we can simplify the sorting of incoming records based on the primary key, as the key can be again used as an offset to put the data directly into a sorted array, reducing the complexity of sorting from $O(N \log N)$ to $O(N)$. Joins over dense integer keys also simplify to lookups, requiring only a direct lookup using the key as an offset. Maybe surprisingly, the same optimizations are still be possible for other data patterns. For example, if the data contains only even numbers or if the data contains duplicates following a deterministic bell-curved shape, etc., we can still do similar optimizations. In other words, knowing the exact data distributions allows to highly optimize almost any algorithm and data structure the database system uses and potentially even change the complexity class of some algorithms.
Of course, in most real-world use cases the data does not perfectly follow a known pattern, and it is usually not worthwhile to engineer a specialized system for every use case. However, if it were be possible to learn the data pattern, correlations, etc. of the data, it might be possible to automatically synthesize index structures, sort and join algorithms, and even query optimizers algorithms that leverage these patterns for significant performance gains. Ideally, if the pattern can be efficiently learned, it might allow to change the complexity class of algorithms as done in the example above.
In this talk, I will start from this premise, present initial results of synthesizing index structures and other algorithms as well as outline directions for future work.
Bio:
Tim Kraska is an Associate Professor of Electrical Engineering and Computer Science in MIT's Computer Science and Artificial Intelligence Laboratory. Currently, his research focuses on building systems for machine learning, and using machine learning for systems. Before joining MIT, Tim was an Assistant Professor at Brown, spent time at Google Research, and was a PostDoc in the AMPLab at UC Berkeley after he got his PhD from ETH Zurich. Tim is a 2017 Alfred P. Sloan Research Fellow in computer science, received the 2017 VMware Systems Research Award, an NSF CAREER Award, an Air Force Young Investigator award, two Very Large Data Bases (VLDB) conference best-demo awards, and a best-paper award from the IEEE International Conference on Data Engineering (ICDE).
Bringing together Data Integration, Semantic Web and Machine Learning for Drug Safety and Efficacy Predictions
Achille
Fokoue, IBM Research AI
Abstract: Accurate predictions of potential
interactions between drugs, proteins, genes, and diseases are a major requirement both in drug
development and in patient care. Fortunately, the increasingly large
amount of drug-related (open) data published on the Web, describing
various properties of drugs and their relationships to other drugs,
proteins, genes, diseases, related concepts and entities, provides the
raw material needed to enable such accurate predictions. In this talk,
I will present an end-to-end system that takes in various Web data
sources as input and provides as output drug-drug interaction (DDI)
and drug-protein interaction (DPI) predictions along with an
explanation of why two entities (drug-drug or drug-protein) may
interact. The system first creates a knowledge graph out of the input
data sources through large-scale semantic integration, and then
performs link prediction among drug/protein entities in the graph
through large-scale similarity analysis and deep representation
learning. Finally, I will discuss some of the key lessons learned from
building such a system that might be relevant to the data management
community.
Bio:
Achille Fokoue is a Research Staff Member and Master Inventor at IBM Research AI. His research has focused on knowledge representation and reasoning, semantic web technologies, and machine learning. He has developed theories, algorithms and systems for scaling reasoning over large and expressive description logics knowledge bases that tolerate inconsistencies and uncertainties. He has also successfully applied semantic web technologies and machine learning in a variety of domains ranging from healthcare & life sciences to text analysis. In particular, in the past three years, he has been exploring the problem of predicting and elucidating the mechanisms of drug-drug interactions and drug-protein interactions using similarity based and deep learning based approaches on knowledge-graphs curated and linked from a variety of sources on the web. He is author or coauthor of 16 patents, 80+ scientific reports and manuscripts that have been cited, in aggregate, more than 2,000 times.
Learning to Make Usable Database Systems
H. V. Jagadish, University of Michigan in Ann Arbor
Abstract:
Database systems have traditionally expected users to be experts both in the programming interface/language and in the specifics of the schema. Over the past decade or so, the importance of usability has gradually been realized, and several worthwhile efforts have taken important steps towards making database systems usable. The main challenge is getting the system to understand and act upon poorly and partially specified user intent. But this is a type of problem that artificial intelligence is often used to solve. So AI should be a major tool in improving database usability. In this talk, I will describe some efforts we have made in this direction. Based on this experience, I will speculate on what AI is likely to be good for, and what it will not be as good for.
Bio:
H. V. Jagadish is Bernard A Galler Collegiate Professor of Electrical Engineering and Computer Science, and Distinguished Scientist at the Institute for Data Science, at the University of Michigan in Ann Arbor. Prior to 1999, he was Head of the Database Research Department at AT&T Labs, Florham Park, NJ.
Professor Jagadish is well known for his broad-ranging research on information management, and has approximately 200 major papers and 37 patents. He is a fellow of the ACM, "The First Society in Computing," (since 2003) and of AAAS (since 2018). He served on the board of the Computing Research Association (2009-2018). He has been an Associate Editor for the ACM Transactions on Database Systems (1992-1995), Program Chair of the ACM SIGMOD annual conference (1996), Program Chair of the ISMB conference (2005), a trustee of the VLDB (Very Large DataBase) foundation (2004-2009), Founding Editor-in-Chief of the Proceedings of the VLDB Endowment (2008-2014), and Program Chair of the VLDB Conference (2014). Among his many awards, he won the ACM SIGMOD Contributions Award in 2013 and the David E Liddle Research Excellence Award (at the University of Michigan) in 2008.
He has recently launched an online course on Data Ethics. See https://www.edx.org/course/data-science-ethics-michiganx-ds101x
Session 1 (8 - 10.30 am)
-
Contextual Intelligence for
Unified Data Governance,
Edward
Seabolt, Eser Kandogan and
Mary Roth, IBM Almaden
Research Center
-
StreamScope: Automatic Pattern
Discovery over Data Streams,
Kouki
Kawabata, Yasuko Matsubara and
Yasushi Sakurai, Kumamoto
University
-
(Keynote Presentation, 9-10.30 am) Bringing together Data
Integration, Semantic Web and
Machine Learning for Drug
Safety and Efficacy
Predictions, Achille
Fokoue, IBM Research AI
Coffee Break (10.30 - 11 am)
Session 2 (11 am - 12.30 pm)
-
GridFormation: Towards
Self-Driven Online Data
Partitioning using
Reinforcement Learning,
Gabriel Campero Durand,
Marcus Pinnecke, Rufat
Piriyev, Mahmoud Mohsen,
David Broneske, Gunter
Saake, Maya Sekeran, Fabián
Matias Rodriguez and Laxmi
Balami, University of Magdeburg
-
Deep Reinforcement Learning
for Join Order Enumeration,
Ryan
Marcus and Olga Papaemmanouil,
Brandeis University
-
Deep Reinforcement-Learning
Framework for Exploratory Data
Analysis, Tova Milo
and Amit Somech, Tel Aviv University
Lunch Break (12.30 - 2 pm)
Session 3 (2 - 3.30
pm)
-
(Keynote Presentation)
Learning Data Systems Components Tim
Kraska, MIT
Poster
Session (3.30 - 4.30
pm)
Session 4 (4.30 - 6.30 pm)
-
(Keynote Presentation 4.30-6 pm)
Learning to Make Usable Database Systems
H. V. Jagadish, University
of Michigan in Ann Arbor
-
AI and Data Management:
Open Issues and Future
Directions Open
Discussion
Workshop Co-Chairs
For questions regarding the
workshop please send email to bordaw AT us DOT ibm DOT com.
Program Committee
- Sandeep Agrawal, Oracle
- Subhabrata Mukherjee, Amazon
- Tin Kam Ho, IBM Watson
- Chid Apte, IBM Research
- Vijil Chenthamarakshan, IBM Research
- Tim Kraska, Brown University
- Jens Dittrich, Saarland University
- H. V. Jagadish, Univ. of Michigan
- Tim Oates, Univ. of Maryland, Baltimore County
- Sharad Mehrotra, University of California, Irvine
- Rajesh Parekh, Facebook
- Daisy Zhe Wang, University of Florida
- Fei Chiang, McMaster University
- Ken Pu, University of Ontario Institute of Technology
- Yongjoo Park, Univ. of Michigan
- Paper Submission: Friday, 16th March 2018
- Notification of Acceptance: Monday, 16th April, 2018
- Camera-ready Submission: Monday, 30th April, 2018
- Workshop Date: Monday, 10th June, 2018
Submission Site
All submissions will be handled electronically via EasyChair.
Formatting Guidelines
We will use the same document templates as the SIGMOD/PODS'18
conferences (using the 2017
ACM format). It is the authors' responsibility to ensure that
their submissions adhere
strictly to the 2017 ACM
format . In particular, it is not allowed to modify the format with the objective of squeezing in more material. Submissions that do not comply with the formatting detailed here will be rejected without review.
The paper length for a full paper is limited upto 8
pages. However, shorter papers
(4 pages)
are encouraged as
well.
All accepted papers will be
indexed via the ACM digital
library and available for
download from the workshop
webpage in the digital
library.
|
