;; -*- mode: outline -*-


	   CRAM - a Contextual Relational Augmented Memory for
			 Personal Data Mining

		     KDL Challenge Grant Proposal

			   Matthew Cornell
		    Knowledge Discovery Laboratory
		     Computer Science Department
		University of Massachusetts Amherst
			      2004-06-14


1. ROADMAP

The MAIN GOAL of this work is to create self-structuring storage that
integrates disparate information sources (e.g., web pages visited,
email sent/received, RSS feeds subscribed to) in order to infer
relationships between entities, then utilize context to focus the
user's attention on input that is important, novel, or interesting.

End users would obtain TANGIBLE BENEFITS from self-organizing data
storage because it would directly address the worsening problem of
information overload. The techniques developed would allow users to
save time by focusing attention where most needed, and to efficiently
find previously seen items that are currently lost in the flood of
incoming data. Areas of application would include widely-used
information sources such as web pages visited, email sent/received,
RSS feeds subscribed to, news groups read, and instant messages
sent/received.

The CRITICAL TECHNICAL BARRIERS that have stood in the way are due
primarily to both the failure of previous approaches to consider the
interconnectedness of information, and their failure to utilize
relatively new techniques emerging from the field of Relational
Knowledge Discovery.

The MAIN ELEMENTS of our approach involve research in three related
aspects: 1) Identifying relevant entities from disparate information
sources, relying on research from Information Extraction, 2)
inferring the interrelationships between entities, relying on work in
Link Detection, and 3) learning and recognizing user context in order
to focus attention on desired input. Additionally we will need to
continue our work in relational data storage and access that
facilitates this type of knowledge discovery.

The RATIONALE for our work in this direction is based on our lead in
research into the techniques of Relational Knowledge Discovery,
esp. in the creation of models that leverage relational data without
succumbing to newly-discovered biases inherent in such networks of
data. Additionally, we will take advantage of our work in creating a
high-performance data store supporting: a) heterogeneous types of
items and their relationships, and b) the new kinds of operations
required for learning and applying these models.

Our EXPECTED RESULTS include the development of a class of algorithmic
techniques to support creating attention focusing models, and the
creation of end-user tools to integrate these models with a small
number of diverse information streams.

The RISKS ENTAILED by not doing this work are that our vision of an
integrated relational knowledge store won't come to pass, and that the
techniques and lessons of Relational Knowledge Discovery won't be
considered for this important problem area. This could result in a
loss of a significant amount of time by many of us who continue needing
to manage an increasingly large influx of daily information.

We plan to EVALUATE PROGRESS and capabilities on two levels: a)
studying their application to small standardized test datasets (which
we will have to develop) for scientific correctness and accuracy, and
2) through qualitative user analysis and user studies of the
effectiveness of tools using these models.


2. DESCRIPTION

Our proposal seeks to address the increasingly important problem of
information overload, which often results in 'digital deja vu' -
trying to recall the location of an important piece of information
that was seen before but is now lost. As many have pointed out, at
least three emerging trends are dramatically exacerbating the problem:
massive (essentially unlimited) personal digital storage, increasing
numbers of digital information sources, and diverging quality of those
sources. From [2]:

  ...it's not difficult to imagine saving, on your PC's disks - all
  interlinked at least in terms of who, what, where, when, and why; and
  subsets of which will be replicated across many devices,

    o all of your email, forever
    o all of your personal and business calendars, forever
    o all of your class and meeting notes, and scribblings and
      annotations on everything - including audio and maybe even other
      media, forever
    o all of your documents and presentations, forever, as well as the
      transcripts of meetings that you attend
    o all of your contacts and address books
    o all of your pictures - hundreds of thousands of digital
      snapshots at at least a megabyte a pop, as sensors become ever
      more dense, and as every phone becomes a camera phone
    o innumerable voice messages, as cell phones begin to have "record
      this & send it to me" buttons on the side
    o of course, huge collections of audio, as well as video - some
      personally created, some licensed and cached

As a result, tools to help manage human attention will be
critical. From [1]:

    ... human attention has become the limiting resource of the
    information age. Informational complexity has exploded, making
    attention -- the resource needed to make sense of information --
    exceedingly valuable. Consequently, an organization's ability to
    focus attention effectively has become a key determinant of its
    success.

Our goal is to create self-structuring storage that accepts input from
diverse information sources, structures it by finding relationships
between (possibly inferred) entities, and supports trained 'attention
focus' models that prioritize information for the user. Put another
way, we seek to develop techniques that support a free-form database
with automated analytic tools for personal data mining of the kinds of
information listed above.

Most current approaches to solving this problem use either manual
hierarchical organization (e.g., outliner programs), manual keyword
tagging (e.g., browser bookmarks with keywords), manual network
organization (hypertext and personal wiki programs), or automated text
indexing ('Googling your data') via Information Retrieval methods
(e.g., Google's Gmail, Zoe). We believe using these alone or in
combination will fail because they either a) aren't amenable to
automation, b) do not take advantage of the inherent relational
characteristics of the data, or c) don't allow mixed-initiative
(manual) manual intervention for the very high quality
advice/organization people excel at.

Our proposed approach, summarized here and discussed below in detail,
can be broken into three parts: incremental data representation and
storage, relationship and entity inference, and context-based
attention focus modeling. Underlying all three parts is a
network-based data representation that explicitly models relationships
between data items. In addition to being the richest representation,
it closely mimics relationships that people reason about. Upon the
data representation (which must scale and support incremental updates)
we propose applying Machine Learning techniques from the field of
Relational Knowledge Discovery to help users focus attention.

From the user's perspective, the system would work like a black box
that is blindly fed data from existing information manipulation tools
(such as email and web browsing) and produces recommendations for
attention, using training information obtained from the user. It would
also support other powerful analysis tools such as network-based ones
(e.g., centrality measures, hubs and authorities), and graphical
browsing of the data.


Next we discuss the proposed problem's three parts in detail.


  Data Representation and Storage

The first part of our proposal is a flexible network representation
that is general enough to handle disparate information sources such as
web pages browsed, email sent and received, and RSS feeds subscribed
to. We plan to use an 'overlay' approach to integration with existing
data-specific tools (e.g., web browsers, email clients) in which
frequent incremental updates import data saved in those applications
into the store, easing adoption. The storage will support millions of
data items and relationships, and will allow for very fast graphical
querying.


  Relationship and Entity Inference

Because we will accept data from disparate information sources and
will import data (rather than trying to force users to adopt a
'kitchen sink' tool like [3]), we must use existing data
formats. These formats are very different, and most do not support
explicit representations of data items and relationships. Thus, an
important and challenging step of organizing the store is to detect
entities and the relationships between them. We will adopt approaches
from Information Extraction for this purpose.


  Attention Focus Modeling

The final part of our proposal is the most challenging and the most
important - the ability to learn where a user's attention should be
focused. The central idea is that the user's work context is crucial
to determining what information is relevant to her. For example, when
a user is investigating an idea, she will perform a number of
activities using information tools, such as searching for relevant web
pages and sending email. Taken collectively during her investigation,
these information sources are related only by the user's goal in
accessing them. The proposed solution should be able to learn these
contexts, so that later when the user wants to re-activate one, the
analysis tools should be able to indicate new (and old) relevant
information. Note that it is possible that users will encounter
similar resources (from the same web site or related to the same email
user) when performing different activities, but those should be
excluded from their attention. Another way to think of such a model is
as a 'digital fovea' [4] - an trainable artifact that is sensitive to
a particular need. Once trained, it will activate (call for attention)
when it sees something of interest.


3. FUTURE APPLICATIONS

A natural extension to this proposal is that of adding personal
information management (PIM) data, including inputting todo items,
calendar entries, and personal notes/stickies (which would link to and
mark up any other kind of data entity).

Additionally, many other data input sources could be integrated,
including various forms of media (photos, phone calls, video), as well
as location/GPS information.

Finally, an intriguing application of the proposal to collaboration
would be to share trained attention models. The thought would be to
give a trained model to a colleague who would then allow the model to
see her incoming data, but without sharing the data itself. If the
shared model indicated the owner's interest, the colleague could
(possibly automatically) send the data (or a reference to it) to the
model's owner.


4. PROPOSAL STRENGTHS AND CHALLENGES

Scope: We envision the proposal will see results in 2-3 years.

Collaboration opportunities: As mentioned above, this problem would
require the integration of areas of research in Information Retrieval,
Information Extraction, and Machine Learning (possibly including
various kinds of models including Reinforcement Learning).

Ease of evaluation: Evaluation is a challenge for this proposal
because initial ideas tend to involve human judgement to determine
the quality of attention recommendations. We suspect more thinking
would provide alternative evaluation techniques.

Readily available data: Because we would be using personal information
that each individual would control and not share, data would be both
readily available (say from email archives and browser histories), and
its privacy would be maintained.


5. REFERENCES

[1] Staying on Point The Science and Art of Managing Attention, Andrew
    Norman, http://www.imakenews.com/techyvent/e_article000214657.cfm?x=a2sSp5g,aLBqYML

[2] '640KB ought to be enough for anyone', Ray Ozzie's Weblog,
    http://www.ozzie.net/blog/stories/2003/11/14/640kbOughtToBeEnoughForAnyone.html

[3] Universal information client projects: Haystack
    (http://haystack.lcs.mit.edu/), OmniaMea
    (http://www.jetbrains.com/omea/index.html), and Chandler
    (http://www.osafoundation.org/)

[4] http://www.stlukeseye.com/anatomy/Fovea.asp