Exploiting Latent Information in Databases
                                  via Database Embedding:
                              technology, applications, ethics
                                        (Invited Talk)
                                                                        Oded Shmueli
                                                       Technion – Israel Institute of Technology
                                                                     Haifa, Israel
                                                              oshmu@cs.technion.ac.il
We are witnessing the emergence of AI-powered database sys-                           Limiting information disclosure is also an important consider-
tems, embedding AI-ideas and techniques in query processors,                       ation. Especially within an organization, there is a need to share
concurrency controllers, and more. We aim at improving rela-                       information. However, it would be desirable that this sharing
tional querying, as well as other functionalities, by introducing                  enable productive work while hiding information that is not
another layer of data, word vectors, into traditional database                     essential for that work. To this end, we introduce degrees of
systems. Word vectors originate in Natural Language Processing                     disclosure. Here, some information in the database is encrypted,
(NLP) where they are used to represent words in a language. In                     some is simply not supplied, while additional information is in-
NLP, there are a number of methods for obtaining word vectors                      tentionally supplied in the form of a model.
from text, we use a variation of one of these methods, word2vec.

   The idea in a nutshell is as follows: we produce text from a
relation (or a view thereof) and then use this text to generate a
model, i.e., a set of vectors, for all terms in the database. Once the
model is available, we can formulate Cognitive Intelligence (CI)
queries. These queries may be realized by SQL queries, enhanced
by User Defined Functions (UDFs) that take advantage of the
model to formulate conditions that were previously practically
not expressible in SQL.

   The process of vector construction is different than in NLP. It
reflects the characteristics of relations, with integrity constraints
and named columns which contain various data types, strings,
dates, numeric values, images and more. We call this process
db2vec. There are a number of options for model generation:
based on the textification of a single or multiple relations, incor-
porating external text sources (e.g., Wikipedia), incorporating
externally produced models, standalone, or as building material
for constructing a local model.

   There are many application areas that may benefit from our
approach: Commerce, Finance, HR, Science, and more. Whereas
there are generic UDFs, some application areas require develop-
ing specialized UDFs. One example is a food database application
in which a record has a list of ingredients, in decreasing order of
importance.

    A model reflects the textual and vector sources used to produce
it. As decisions may be based on queries using the model, the
production of models brings to the forefront issues of fairness
and ethics. An important issue is the specifics of the data and text
sources, their weighting in producing the model, and whether
they are biased in some way.



© 2020 Copyright held by the owner/author(s). Published in the Workshop Pro-
ceedings of the EDBT/ICDT 2020 Joint Conference, March 30-April 2, 2020 on
CEUR-WS.org. Distribution of this paper is permitted under the terms of the Cre-
ative Commons license CC BY 4.0.