Crockett, KA ORCID logoORCID: https://orcid.org/0000-0003-1941-6201,
O’Shea, James ORCID logoORCID: https://orcid.org/0000-0001-5645-2370,
Khan, Wasiq ORCID logoORCID: https://orcid.org/0000-0002-7511-3873
and Bandar, Zuhair (2018) A hybrid model combining neural networks and
decision tree for comprehension detection. In: 2018 International Joint Con-
ference on Neural Networks (IJCNN), 08 July 2018 - 13 July 2018, Rio de
Janeiro, Brazil.
Downloaded from:
https://e-space.mmu.ac.uk/624526/
Publisher: IEEE
DOI: https://doi.org/10.1109/IJCNN.2018.8489621
Please cite the published version
https://e-space.mmu.ac.uk
A hybrid model combining neural networks and
decision tree for comprehension detection.
James O’Shea
1
, Keeley Crockett
1
, Wasiq Khan
1
, Zuhair Bandar
2
1
School of Computing, Mathematics and Digital Technology
Manchester Metropolitan University
Chester Street, Manchester, M1 5GD, UK
2
Silent Talker Ltd, Manchester, UK
J.D.OShea@mmu.ac.uk
Abstract— The Artificial Neural Network is generally
considered to be an effective classifier, but also a “Black Box”
component whose internal behavior cannot be understood by
human users. This lack of transparency forms a barrier to
acceptance in high-stakes applications by the general public. This
paper investigates the use of a hybrid model comprising multiple
artificial neural networks with a final C4.5 decision tree classifier
to investigate the potential of explaining the classification
decision through production rules. Two large datasets collected
from comprehension studies are used to investigate the value of
the C4.5 decision tree as the overall comprehension classifier in
terms of accuracy and decision transparency. Empirical trials
show that higher accuracies are achieved through using a
decision tree classifier, but the significant tree size questions the
rule transparency to a human.
Keywords—knowledge rule extraction, artificial neural
networks, decision trees, backpropagation, comprehension,
FATHOM, Silent Talker, non-verbal behavior.
I. INTRODUCTION
Non-Verbal Behaviour (NVB) was first studied systematically
by Charles Darwin and it has become a well-established part of
sciences such as biology and psychology. NVB consists of all
of the signs and signals - visual, audio, tactile and chemical
used by human beings to express themselves apart from speech
and manual sign language. It has been postulated that NVB
features are indicators of internal mental states, in particular
that they can be used to detect deception during interviews [1].
The first system to classify deceptive behaviour automatically,
Silent Talker (ST), used Artificial Neural Networks [2].
The Silent Talker architecture is highly flexible, and has been
adapted to monitor human comprehension in clinical trials
using non-verbal behaviour, employing ANN classifiers [3].
This version is known as FATHOM and currently work is
underway to incorporate FATHOM in an intelligent tutoring
system to provide round-the-clock support in the form of
learner-adaptive online teaching and learning tutorials. For
both FATHOM and ST there has been great interest in how the
system works i.e. which non-verbal indicators are actually
conveying the information to perform the classification. This is
particularly true for Lie Detection where interrogators are
looking for techniques they can apply during questioning and
suspects are looking for countermeasures they can use to avoid
being detected, for example, the well-known myth that looking
up and to the right indicates lying. Unfortunately, although
ANNs are powerful and versatile components in the AI
toolbox, they are also black boxes with no ready explanations
of how they achieve their ends and this has been a concern for
decades [4].
There are many other fields than education in which ANNs
may make high-stakes decisions and some progress has been
made in extracting rules from ANNs, although the degree to
which solutions to reasonably complex problems could be
understood by a non-AI specialist remains debatable. These
include classifying incipient faults in a power transformer [5],
hydrological modelling [6], Credit-Risk Evaluation [7] and
software cost estimation [8]. Some progress has been made in
extracting rules from recurrent neural networks by
transforming them to finite state machines [9], and [10] has
attempted to unify various neuro-fuzzy rule approaches for
ruled generation from recurrent and feedforward neural
networks in a single soft computing framework. Nevertheless,
in analysing a study of using neural networks to predict
academic performance of college students one year in advance,
Schneider et al. [11] observed that the basic problems of
communicating how they reach their conclusions in
meaningful terms has yet to be solved. They highlighted the
problem of explaining how a combination of currently high
subject performances could lead to an anticipated decrease in
the student’s achievement.
Decision trees [12] are highly effective for classification
tasks. They are also considered inherently transparent in
explaining how they reach their conclusions and may be
expressed in the form of production rules, which are generated,
by learning and reasoning from feature-based examples. Many
studies have been conducted to compare decision trees with
neural networks – a more recent study of multiple classifiers
can be found in Delgado et al. [13]. In general, ANNs take
longer to train than decision trees due to the large number of
iterations required to ensure training reaches its full potential
[14]. Classification accuracy is largely dependent on the
dataset, but the transparent nature of decision trees gives
insight into the relationships between features [14]. In such a
domain as the analysis of NVB for comprehension detection,
decisions trees would provide an insight into key behaviours
and their interactions.
In the FATHOM architecture to date, classification of
comprehension / non-comprehension has been performed by a
single, final back propagation artificial neural network
(BPANN), preceded by layers of BPANNs that process
individual features. It is this final stage in which an
intervention should be possible to explain how these features
indicate comprehension / non-comprehension. Therefore, the
research questions addressed in the work presented in this
paper are:
1. Can the final ANN classifier be replaced by a decision tree
without loss of performance?
2. Can the decision tree be converted into comprehensible
production rules?
For a comprehensible rule set be possible, there must be a
limited number of rules for the human user to interpret and
these are proportional to the number of nodes in the tree.
Consequently, the primary interest in answering question 2 is
whether or not the tree has a manageable number of nodes.
To answer these research questions, two datasets collected
from FATHOM studies have been used. The experimental
study known as “Termites”, reported in [15] was used to
identify whether high and low human comprehension
associated multi-channels of non-verbal behaviour reside
within a video-recorded British (UK-based/English speaking)
sample of participants in a class room environment. The
Termites exploratory study builds upon lessons learned in prior
work [3] where evidence was found that comprehension / non-
comprehension could be detected in an African female
population sample using a BPANN. This second study is
known as HIV Informed Consent.
This paper continues as follows: Section II reviews related
work in non-verbal behaviour and comprehension, and then
describes the FATHOM comprehension monitoring system
that uses BPANNs. Section III describes the comprehension
scenarios from which the two datasets used in this study were
obtained. Section IV and V describe the experimental
methodology and results. Conclusions and recommendations
for future work are presented in section VI.
II. RELATED WORK
A. Non-verbal Behaviour and Comprehnsion
Non-verbal behaviour comprises all of the signals or cues,
which human beings use to communicate, including visual,
audio, tactile and chemical components [16, 17]. During a
spoken dialogue, humans will often transmit non-verbal cues
before the verbal component [17], which can be used to detect
the sender’s state. It has been recognised that the face is a
source of rich information in terms of exhibiting meaningful
non-verbal behaviour. Little work has been done in the
automatic detection of classification of non-verbal behaviour.
Traditional methods employed human judges to code each
channel [18, 19]. However, each judge needs to be trained and
will provide a subjective opinion on the behaviour being
delivered by a particular channel. The process is time
consuming and an impossible task for a human to monitor
more than a limited number of channels accurately.
Two recent research strategies for acquiring non-verbal
behavioural cues have attracted attention in the literature; these
are Facial Microexpressions [1] and using the Microsoft Kinect
computer vision algorithm [20]. Micro-expressions are said to
be a small “universal” set of expressions of extreme emotion:
disgust, anger, fear, sadness, happiness, surprise, and contempt,
and a formalised method of encoding them was defined by
Ekman. The weaknesses of this technique are: its results are
largely based on highly artificial “posed” images using actors
or students provided with highly specific instructions [21,22]
or even training in how to produce facial actions [23,24], low
numbers of detectable Ekman micro-expressions in
spontaneous interviews [25] and a low Classification
Accuracy (CA) for those micro-expressions actually found
[26].
The Microsoft Kinect is primarily aimed at observing
whole body gestures in commercial video game applications.
However, there has been some interest in adapting it for NVB
research. For example, facial expressions have been
investigated as indicators of happiness, anger, sadness and
surprise that are integrated with the head pose changing
information to conceive the human interaction with 3D sensing
technology [27]. Although it should be noted that the
experimental results show emotional and head position change
instead of discrete level accuracy in terms of emotional
classification for the four aforementioned emotions. Likewise,
it was tested on a limited participants (i.e. 20) as well as
insufficient facial channels (i.e. 12). Typically, psychological
experiments do not provide any methodology for applying
these population sample differences to classify particular
individuals.
FATHOM (described in Section II, C) is distinguished
from these two techniques in three respects. Firstly, it uses
large numbers of features at a much finer level of granularity
than body gestures or facial expressions. Secondly, the domain
it operates in, human comprehension has not been a previous
subject of AI research. Thirdly, it does classify an individual
person’s state of comprehension / non-comprehension based
the non-verbal behaviour. Fathom does not rely on high frame-
rate cameras or constrained recording environments that
facilitate the setup of the technology, nor does it depend on
specialised hardware whose future availability may be
dependent on market forces (such as the game-oriented Kinect)
– making it suited for everyday classroom use.
B. Non-Comprehension
Non-comprehension is regarded as “a state of knowledge that
ranges from uncertainty to complete lack of understanding of
the materials under discussion” [28], i.e. an absence of
comprehension. The vast majority of research on
comprehension concerns reading and the understanding of
written text, initially by identifying the main ideas in the text
[28, 29]. A further elaboration is the view that successful
comprehension depends on the construction of a coherent
representation of text in memory [30]. Despite the traditional
bias towards reading texts, there has been interest for some
time in comprehending audio and video materials in language
teaching [31] and the informed consent process [32]. At a
more abstract level, the comprehension of metaphors, requires
thinking beyond the literal meaning in order to understand the
figurative meaning of the sentence [33] – yet metaphors and
similes are frequently used by good teachers to convey
complex ideas. In the completely independent field of
advertising, a controlled degree of cognitive complexity is
considered desirable, where confronting an audience with a
cognitive challenge generates an appreciative payoff if they
can solve the challenge [34]. So the non-comprehension state
may be characterized as an inability to extract and characterize
the salient elements of information received, an inability to
model such information in a more abstract form or an inability
to generalize from a specific meaning to more abstract
thoughts about such a communication.
C. FATHOM
FATHOM utilises a bank of BPANN’s to capture, monitor
and detect multiple channels of human non-verbal behaviour
continuously. FATHOM has been successfully shown to detect
non-verbal behaviour associated with comprehension in two
studies [3] [15].
Input to FATHOM is currently offline through recorded
videos, which are streamed into FATHOM where a series of
BPANN facial object locators, identify the location in a video
frame of key visual features such as the eyes. For each non-
verbal behavioural feature identified from a specific visual
feature, the BPANN facial object pattern detectors identify its
state i.e. the left eye is half-open. The NVBs identified are then
coded into individual channels and group channels i.e. all
channels associated with eye behaviour.
States are typically collated over a time interval, e.g. 3
seconds grouped into one vector for Silent Talker– but this can
be varied depending on the problem domain and FATHOM
uses a 1-second interval. Classification features (patterns) are
extracted from aggregated video-streamed frames over the time
interval and compiled to form a vector. Each vector is passed
to the final BPANN Comprehension classifier which outputs a
value between +1 and -1, indicating whether the person
exhibits high comprehension (+1) or low comprehension (-1)
during that period of time. If there were insufficient
information in the vector during a specific time slot, FATHOM
would recognise this and categorise the timeslot as
unclassifiable. At the end of a session i.e. a tutorial, the overall
comprehension/non-comprehension classification level is
displayed.
FATHOM simultaneously monitors 40 non-verbal
behavioural channels that include 20 channels capturing facial
features such as blushing and 16 channels capturing eye
movement such as right eye looking left. An overview of the
FATHOM architecture can be seen in Figure 1.
The work presented in this paper investigates the
consequences of replacing the BPANN comprehension
classifier in the FATHOM system by a C4.5 decision tree [12],
to answer questions about their relative performance and
transparency.
Fig.1 FATHOM Architecture
III. COMPREHENSION SCENARIOS
This section outlines the two comprehension scenarios
used to collect the data.
A. Study 1:HIV Informed Consent
The first comprehension study was undertaken in Tanzania in
Africa by FHI-360 [35] in collaboration with the National
Institute for Medical Research (NIMR) [36]. NIMR enlisted
sexually active women aged 18-35, who were native Kiswahili-
speakers. 292 participants took part in the study. Two different
experimental conditions (tasks) were used for data collection:
condition A was designed to be familiar and easy-to-
comprehend (condom use) and condition B was designed to be
unfamiliar and intentionally hard-to-comprehend (the effects of
HIV viral mutation on antiretroviral treatment). Each
participant listened to a short learning task script and then
received the associated ten closed and open-ended questions
with randomisation applied. Task order was also randomised so
that half of the participants completed task A followed by task
B and vice versa.
B. Study 2: Termites
Prior to the study a short learning topic was selected, which
was a factual digital video on Termites with a total duration of
8 minutes 40 seconds. The Termite video was targeted at the
general public with no age restriction and covered: functional
architectural aspects of the termite mounds, roles within the
social structure of a termite colony and locations where termite
colonies thrive. Two experts (Academic Professors in the field)
on the subject area were recruited to develop ten difficult
(hard) questions and ten easy questions related to the video
content. The experts agreed both the question difficulty levels
and the contents of the answer that the participants should
provide. The experts were required to devise five open
questions and closed questions within each set of hard and easy
questions. At the same time, the experts noted down the correct
answer(s) for each question, which were later incorporated into
a scoring scheme.
Forty participants were selected to participate in the study,
from academic and technical staff at the Manchester
Metropolitan University (MMU) in the UK. The sample was
composed of 20 males and 20 females. The males had a mean
age of 41 years old (SD = 14 years) and the females had a
mean age of 39 years old (SD = 14 years). Each participant was
invited to engage individually in a short learning task, which
was comprised of watching a short video on Termites and then
answering a small set of associated assessment questions whilst
being video recorded.
IV. EXPERIMENTAL METHODLOGY
The experimental methodology was to take the pair of
datasets outlined in Section III and use them to train and
evaluate C4.5 decision trees to replace the final stage BPANN
classifier.
A. Using a Back Propagation ANN as the final classifier
For each study, FATHOM’s object locators and pattern
detectors were used to extract and collate the non-verbal
vector-based dataset for the purpose of training the final
BPANN classifier. For both studies, HIV Informed Consent
and Termites, each vector in the final dataset covered a 1-
second time period and represented the state changes for the
compiled non-verbal channels over the period. Each channel
was normalised in the range +1 to -1. The last attribute in each
vector was the desired classification, with discrete values of +1
for comprehension and -1 for non-comprehension. The
following training parameters (determined from previous
exploratory cross-validation sessions) were used to train the
single hidden layer neural network in the Fathom training
application:
Topology: 40:20:1
Accept value: 1.0 (output >= 0.0 equals comprehension
AND output <0.0 equals non-comprehension)
Maximum epochs: 10,000
Checking epochs: 250, i.e. at every 250th epoch the total
Classification accuracy (CA) was checked and if there was
no improvement training was terminated.
Learning rate (ƞ) was set at 0.005.
Weight initialisation: automatic range (0±1/sqrt(fan-in))
where fan-in represents the number of inputs entering the
neuron.
Cross-validation: 10-folds
For study 1, eighty randomly selected participant videos (
from the 292 obtained in the study) comprised the HIV
Informed Consent dataset containing 71,787 vectors with
63.5% comprehension and 36.5% non-comprehension. For
study 2, the forty participant videos yielded 16,951
comprehension vectors and 23,857 non-comprehension
vectors. The study 2 Termites dataset was composed of 40,808
vectors with 41.5% in the comprehension class.
B. Using a C4.5 Decision Tree as the final classifier
In order to use a decision tree as a comprehension
classifier the final, BPANN, comprehension classifier
shown in Figure 1 was replaced by the C4.5 decision tree
algorithm.
1) The experiment consisted of a series of trials using
different degrees of pruning to find the optimal C4.5
decision trees and determine the extent to which they
could be pruned. The Weka implementation of C4.5,
known as J48 was used. This was achieved by
establishing a baseline decision tree for each scenario
setting the pruning parameters to Confidence Interval
(CI) = 0.25 and minimum number of objects = 2 cases
per leaf (i.e. the default settings)
2) This was followed by fixing the minimum number of
objects (MNO) at 2 and conducting a series of trials over
a range of confidence interval values to determine which
provides the greatest improvement in CA over the
baseline tree performance.
3) Then the complementary process was performed, fixing
the CI at 0.25 and conducting a series of trials over a
range of values of MNO.
4) Finally, further experiments were performed varying
confidence interval and MNO independently, to find the
most severely pruned tree for each dataset, which, was
not significantly worse than the baseline in terms of CA.
The initial ranges used for the experiments were, for CI: 0.25,
0.2, 0.15, 0.1, 0.05, and for MNO: 2, 5, 10, 15, 20.
V. RESULTS
A. BPANN Comprehension Classifier
Table I shows the overall best BPANN Classifiers for both
studies. Comprehension (C%) and Non-comprehension (NC%)
are the percentages of comprehension and non-comprehension
vectors, respectively, which were classified correctly. Overall
% is the total normalised percentage of comprehension and
non-comprehension vectors classified correctly.
TABLE I: BPANN RESULTS
C%
NC%
Overall CA
%
Study 1: HIV
Informed Consent
88.08
86.79
87.44
Study 2: Termites
72.77
84.09
78.43
B. C4.5 Comprehension Classifier
Table II shows the results of varying the Confidence Interval
used for pruning in decision tree construction (Pruning CI).