Data Interpretation and Domain Knowledge

1
Data Interpretation and Domain Knowledge

• Yaji Sripada

2
In this lecture you learn

• Need and role of domain knowledge in
• Data analysis and
• Data interpretation
• Issues involved in using domain knowledge

3
Introduction

• So far we have studied data analysis methods and
  visualization techniques that are domain
  independent
• They are useful for extracting information from
  input data
• However, the information extracted by these
  methods is factual
• On the other hand users want interpreted
  information
• Domain knowledge is required to transform factual
  information into interpreted information

4
Example

• We learnt how to use bottom-up segmentation
  method to analyse scuba dive profile data
• Consider the dive profile shown above and its
  segmented representation shown below
• User is interested in knowing safety related
  information extracted from the dive profile

5
Fact Vs Interpreted

• The segmented representation presents the factual
  information such as
• The diver ascended rapidly for a while and
• The diver stopped twice on the ascent
• What is missing is the domain knowledge
• The rate of ascent is more than the recommended
  rate and
• One of the stops is unnecessary
• Without such domain knowledge it is impossible to
  communicate useful and interesting information
  (interpreted information) to the user

6
More uses of Domain knowledge

• Domain knowledge is also required for controlling
  data analysis so that
• the results of data analysis contain the required
  level of detail and
• the data analysis process is sensitive to certain
  data items or patterns that are significant (in
  some way) in the domain

7
Example

• Consider the dive profile shown in the earlier
  example
• Bottom-up segmentation produces many different
  segmented representations from this dive profile

8
Multiple Segmented representations

• Three possibilities are shown here, although many
  more are possible

9
Controlling Segmentation

• Which one of the segmented representations is an
  interesting and useful one?
• Which one of them has the right level of detail
  for the user?
• Because bottom-up segmentation follows the
  process of iterative merging of adjacent segments
  achieving the right level of detail means
  controlling the merging process
• You have studied the merging process in bottom-up
  segmentation in practical 2

10
Domain significant data values and patterns

• Continuing with our dive profile example
• If the depth of this dive is deeper than the
  recommended depth for safe dives, the
  segmentation process does not generate any extra
  information about the safety of the dive.
• This means that in this case the two segment
  representation on the right is just as good as
  the detailed segmented representation on the left
• Similarly if a dive profile is reverse (shallow
  first and deeper later) then once again
  segmentation details do not matter

11
Domain significant data values and patterns (2)

• In many other domains we have observed such
  significant values or patterns that affect data
  analysis
• Meteorology wind speeds above or below certain
  ranges are useless for oilrig people because such
  ranges of wind speed either completely prohibit
  operations or do not cast influence on their
  operations at all
• Medical domain the blood pressure changes are
  unimportant after it crosses certain threshold
  values

12
Domain Knowledge

• Domain knowledge is required for
• Data interpretation and
• Controlling data analysis
• In fact, exploiting domain knowledge in computer
  systems has been a generic goal in Computing
  Science
• Database theory which models data based on domain
  independent relational theory uses business
  rules to bring in domain knowledge
• AI and Expert System communities have always
  worked with domain experts

13
Strategies to include domain knowledge

• Domain knowledge can be included
• directly to control the data analysis process
• Because data analysis procedures essentially
  involve searching for patterns and rules domain
  knowledge will reduce the amount of search
• or
• In a post-processing stage where the results of
  domain independent data analysis are inspected to
  eliminate uninteresting information
• or
• To do both

14
Domain Knowledge for Controlling Data Analysis

• A simple way of controlling data analysis methods
  is by introducing domain thresholds as parameters
  in the data analysis algorithm
• E.g. bottom-up segmentation algorithm uses an
  error parameter (ETV) to control its runtime
  behaviour and we can relate some domain threshold
  value to determine values for the error threshold
• However, it is always not possible to find a
  simple mapping between domain thresholds and
  parameters of data analysis algorithms
• Studies in the SumTime project with wind speed
  threshold values controlling segmentation of wind
  speed data have not been very encouraging

15
Domain Knowledge for interpretation

• Interpretation involves inferring interesting
  information from factual information
• Based on the amount of domain knowledge available
  interpretation can happen to several levels of
  inferences
• Therefore knowledge of user tasks should be known
  to determine the desired level and the required
  domain knowledge
• In other words, in order to carry out
  interpretation, we need to know a level
  determined by user task to which inferences are
  required
• We also need to know the required domain
  knowledge to perform the inference - mapping of
  factual information to interpreted information
• In general, user models and context models are
  important for performing data interpretation

16
Expert Systems

• Expert systems are software systems that exploit
  domain and problem solving knowledge acquired
  from experts
• Expert systems offer one solution to the problem
  of data interpretation
• However, we need to face the following
  challenges
• It is difficult to acquire all the required
  knowledge from any one source
• It is difficult to merge (consolidate) knowledge
  from multiple sources

17
Resume

• Resume is a knowledge based expert system for
  deriving temporal abstractions from medical time
  series
• Focuses on interpretation of time series based on
  contextual knowledge
• Input Time stamped clinical data and relevant
  events
• Output abstracted temporal intervals
• Identifies past and present trends and states
• Supports decisions based on temporal patterns,
  such as
• modify therapy if the patient has a second
  episode of Grade II bone-marrow toxicity lasting
  more than 3 weeks

18
Cost of expert systems

• Resume uses a tool called Protégé for experts
  specifying knowledge
• Protégé allows experts define an ontology and
  then specify domain knowledge using that ontology
• It is expensive to build Resume like tools and
  therefore data interpretation is an expensive
  step in building our data interpretation and
  communication technology
• As a cheaper alternative, simple levels of
  interpretation can be achieved by using domain
  knowledge in the form of thresholds on values
• E.g, we can infer that a blood pressure value of
  150 is high by knowing the threshold for normal
  blood pressures

19
ScubaText

• ScubaText is a research project in our department
• Task produces reports on interpretations of dive
  data
• Input Dive Computer Data similar to the data
  you work with in the practicals
• Output Reports contain textual presentations
  combined with annotated graphical presentations
  of dive data
• Text presents an interpretation of the data/graph
• Annotations on the graph provide interpretations
  of some of the phrases in the text

20
ScubaText Prototype

• Prototype architecture is a four stage pipeline

Data Analysis analysing raw data for required
features/patterns Data Interpretation mapping
the data features/patterns to the actual dive
features e.g. saw tooth pattern (data pattern)
mapped to poor buoyancy control rating the dive
based on the dive features e.g. long bottom
times receive low rating
21
TextAnnotated Graphics
Risky dive with some minor problems. Because your
bottom time of 12.0min exceeds no-stop limit by
4.0min this dive is risky. But you performed the
ascent well. Your buoyancy control in the bottom
zone was poor as indicated by saw tooth
patterns marked A on the depth-time profile.
22
Dive Data Interpretation

• Features and patterns (abstractions) extracted
  using data analysis methods cannot be
  communicated to the user directly
• No use in saying you have a saw tooth pattern on
  your dive profile
• If the user is ignorant of the implications of
  that pattern
• Abstractions need to be mapped to domain features
• Useful to say Your buoyancy control in the
  bottom zone was poor as indicated by saw tooth
  patterns
• A given abstraction may have more than one
  interpretation
• Saw tooth pattern could be either due to poor
  buoyancy control or the result of the diver
  moving up and down inspecting a ship wreck
• Knowledge of the dive context and user modelling
  can help in disambiguating data interpretation

23
Summary

• Domain knowledge is required for
• Controlling data analysis and
• Performing data interpretation
• User modelling and contextual knowledge play a
  major role in data interpretation
• However, the cost of incorporating domain
  knowledge needs to be justified by the human
  effort needed to interpret factual information
• If humans can interpret factual information
  easily, our technology should not attempt to
  build expensive interpretation modules
• However because users vary in their
  interpretation capacity, for some classes of
  users interpretation might always be required