Alon Halevy - Microsoft Research

advertisement
Lowell 2003 Challenges
Alon Y. Halevy
University of Washington
The KR/DB Impedance Mismatch
We often feel that knowledge and inference would be
helpful.


E.g., semantic query optimization, web search, data
integration, meta-data management, wrapper construction.
To date, most KR/DB work has been quite theoretical (data
models, query containment, integrity constraint reasoning).
Why haven’t database systems benefited more from
KR technology?



One explanation: there is an impedance mismatch:
Classical KR systems provide pure logical reasoning.
DB applications often need knowledge to rank multiple
plausible answers/plans/transformations…
Going Forward
The good news from KR:



Probabilistic reasoning answers questions such as “what
is the probability of X?”,
With machine learning it is possible to create models from
data, rather than brittle hand-crafting.
The DB community needs to define the needs (API)
Example: Self-tuning systems



Too many knobs to set. Took several years to understand a
small subset (index selection).
We need a giant rule set, but the rules need to handle
uncertainty.
Need to learn this rule set from data.
Other applications: web querying (focused
crawling), user interfaces, schema matching, and
lifelong personal data management.
Lifelong Personal Data
Management
Save all the stuff I ever care(d) about
(contacts, grades, boy scout assignments,
stock portfolio, files, talks, restaurant reviews)
Challenges:



Our schema evolves (radically) over time.
Data management systems change constantly.
Our focus of attention changes.


Find data in your information attic: find the photo of the
girl next door from 8th grade.
Need to combine text and structured data, and
make it accessible to everyone.
Download