Content uploaded by Alla Abdella
Author content
All content in this area was uploaded by Alla Abdella on Mar 17, 2020
Content may be subject to copyright.
FLAIRS-31 Poster Abstracts
Vasile Rus, Keith Brawner
Editors
A Survey about Machine Learning
Algorithms in E-Commerce
Alla Abdella
(Florida Institute of Technology)
This survey paper covers two e-commerce applications,
marketing selling enhancement, and riskiness fortification.
In online marketing, e-commerce sites like Amazon, eBay,
etc., utilize data mining algorithms profusely to assist its
customers in finding what their purchase needs. With the
growth of the e-commerce domain, credit card usage has
become a common phenomenon. This gives a chance for
fraudulent to commit fraud. This paper presents two topics
- recommendation systems and fraud detection - and shows
how sophisticated machine learning approaches can en-
hance both. In particular, it presents a survey about collab-
orative methods as part of an attempt to improve recom-
mendation systems. It then surveys support vector ma-
chines as an attempt to enhance credit card fraud detection.
Also, this paper elaborates on how the aforementioned al-
gorithms enrich both users' online buying experience and
e-commerce sites' safety.
Value-Aware Recommendation
with Multiple Stakeholders
Himan Abdollahpouri, Robin Burke, Bamshad Mobasher
(DePaul University)
The main goal of recommender systems typically is to pro-
vide a personalized information access for users. That is,
showing products to each user, tailored based on their indi-
vidual taste and preferences. In many real-world applica-
tions, however, that is not the only consideration. An
online grocery, for example, has thousands of products
from hundreds of brands each of which may have a differ-
ent commission rate with the platform owner for recom-
mending their products to the user. In this situation, the
recommender system must not only take users’ preferences
into account, it also should try to maximize the system rev-
enue by considering the value of a recommendation.
Moreover, it also has to, in some cases, meet item provid-
ers’ (brands or other owners) preferences regarding which
group of users to target. In this work, we introduce a sys-
tem design for the value-aware multi-stakeholder recom-
mendation that is able to generate recommendations by
taking into account item profitability and brand preferences
as well as users’ preferences. Our system design is general
enough that could also be used for other domains which
have multiple stakeholders involved in delivering or re-
ceiving the recommendations such as online retail stores,
apartment sharing businesses, ride sharing systems etc.
Framing Impacts and Avoidance Techniques
for Group Decision Support Systems
Badria Alfurhood, Marius Silaghi
(Florida Institute of Technology)
Framing refers to untoward ways in which questions or al-
ternatives are presented to decision makers for deliberation
purposes. Framing is an attack method used to manipulate
people into preferring a particular alternative. Deliberative
framing threatens the integrity of group decision support
systems (GDSSs) and has negative impacts on ultimate de-
cisions. Various studies have shown that, unfortunately,
GDSSs are vulnerable to framing, as early posted infor-
mation can sway many later arriving participants. Howev-
er, framing avoidance or reduction is possible. Framing
impacts are studied to get insights for providing mecha-
nisms to avoid negative effects within GDSSs. The GDSSs
structure and mechanisms can be designed in ways that re-
duces framing effects. GDSSs can be designed to incorpo-
rate resistance to framing attacks. In this research, we ad-
dress the GDSSs framing behavior, the techniques people
use to frame choices or arguments, factors that influence
framing in GDSSs, and the design practices and recom-
mendations that eliminate or lessen the negative framing
effects.We investigate several approaches to designing
GDSS tools for providing resistance to framing, namely
The Thirty-First International Florida
Artificial Intelligence Research Society Conference (FLAIRS-31)
547
based on: intrinsic intelligence via input constraints, col-
laboration filtering, and intelligent rendering interfaces.
Mainly, the effort is focused on proposing tools for exper-
iments, identifying GDSSs features that reduce framing,
and measuring the framing effects in GDSSs. De-biasing
framing in GDSSs has big potential impact on society as
people rely heavily on social networks and related technol-
ogy for gathering information to be used in deliberation
and voting purposes, as demonstrated by recent elections.
Can a Computer Learn from a
Natural Conversation with Humans?
Awrad Mohammed Ali and Avelino Gonzalez
(University of Central Florida)
One goal of artificial intelligence is to develop a conversa-
tional agent that can communicate efficiently with the user
without many restrictions or predetermined patterns. Fur-
thermore, it should be an agent that can learn from an
open-ended natural conversation as do humans. However,
existing agents require inputs with a clear description or
specific words in a specific order to trigger the learning
process. This makes the procedure tedious and dependent
mostly on the human. In this work, we present a new con-
versational learning approach that aims to learn from an
extended natural dialogue with a human. We focus our dis-
cussion on how a computer agent can benefit from using a
meaningful representation for its knowledge base, such as
a modified version of a semantic network that links sen-
tences that have a high similarity measure rather than rep-
resenting the nodes with words only. This approach makes
linking the acquired knowledge to existing information
more efficient and accurate. This work also discusses the
benefit of using majority voting among multiple existing
classification algorithms to decide whether a provided in-
put should be learned or only considered a chatting state-
ment and therefore neglected. We also propose using a se-
cond level of classification based on the word level when
there is no unanimous agreement between the classifiers on
the label of the user input (relevant information or chatting
statements). This step improves the classification process
and allows additional information to be captured and
learned. This research makes strides toward more natural
and robust conversational learning systems.
Models and Inference Techniques
for Diagnosis of Embedded Components
Timothy Atkinson, Marius Silaghi
(Florida Institute of Technology)
In largescale software projects, it is not commonly known
how to make the product perfectly secure. Often software
is first written with the sunny day scenario; and then each
rainy-day scenario the software/system engineers can think
of is explored and tested until the company feels that they
have a product they can sell. Anyone who has shipped a
piece of software knows that once the software hits the real
world, new scenarios they never thought of could happen
occur. One approach is to perform more analysis prior to
release. An alternative technique which we propose, is to
have the system engineers provide a parallel formal model
as part of the development process. Then as the software
runs, the model can alert the user if there is ever a signifi-
cant discrepancy between the model and the running soft-
ware, helping the user know that there is a problem as early
as possible. We are experimenting with using statistical
models to determine if a trained model is different from the
current observed model using logical and probabilistic in-
ference techniques.
Bridging the Gap between Artificial
and Spiking Neural Networks
Sylvain Chartier
(University of Ottawa)
Since the development of Spiking Neural Networks (SNN),
it seems that they should be the next successor of artificial
neural networks. SNN is based on temporal aspect of spik-
ing dynamics instead of the mere presence or absence of
spikes; or mean firing rates. Therefore, they should be used
in more applications than standard artificial neural net-
works in machine learning. However, training SNN to en-
code desired behaviours reveals to be more challenging
than expected. Usually, researchers will used spike timing
dependent plasticity to reinforce synapses which is difficult
to implement in a network using distributed representation.
On the other hand, Recurrent Associative Memories
(RAMs) can store information easily in such representation
but lack the dynamical aspect of spikes. Few studies were
able to use recurrent associative memory to display spiking
behaviour and when they do they simply replace the output
function with a spiking one. This has a consequence of still
having two separate models (spiking and non-spiking). In
this study we propose to use an output function that allows
both spiking and non-spiking behaviour within the same
model. Results show that the resulting model can adapt to
context and display essential RAM properties (e.g. patterns
storage, noise tolerance) while being able to output spike
responses. Therefore, the proposed model opens the door
to bring the best to both worlds and allow a wider range of
applications in machine learning.
548
Improving User Acceptability of
Recommendations through Opinion Mining
Arman Dehpanah, Jonathan Gemmell
(DePaul University)
Recommender systems have become increasingly popular
in many different fields to enhance the user experience,
however some users, despite receiving personalized rec-
ommendations, end up reading hundreds of reviews of an
item in order to make an informed buying decision. In this
research, we use Word2Vec to analyze the text reviews and
discover meaningful features. Word2Vec includes two im-
plementations of a shallow neural network; Continuous
Bag of Words (CBOW) and Skip-gram. We rely on
CBOW for the purpose of word representation. CBOW is a
neural network that predicts a target word given the sur-
rounding context words in a sentence. Representations ob-
tained from CBOW are used as input for clustering. For in-
stance, when analyzing comments containing words such
as “cried”, “end”, “hero” and “died” for movies, using
CBOW we can derive a topic such as “Sad Ending” for
those movies and form a cluster containing movies with
sad endings. After forming different clusters, each cluster
is assigned an importance and sentiment score to determine
their ranks. Finally, we propose that presenting these topics
along with the recommendations offers a more satisfying
experience for the user.
Performance Evaluation of a
Real-Time Clustering Algorithm
Gabriel Ferrer
(Hendrix College)
"Clustering algorithms are unsupervised learning algo-
rithms that employ a distance metric to organize their train-
ing inputs into clusters. The classification of a previously
unseen input is determined by the closest distance from
that input to a cluster’s reference value. A real-time clus-
tering algorithm updates the reference values for its clus-
ters in a constant amount of time for each training input.
We are currently assessing the performance characteristics
of the recently introduced real-time clustering algorithm
Bounded Self-Organizing Clusters (BSOC).
When it was originally presented, the classification per-
formance of BSOC was compared to a real-time formula-
tion of the Self-Organizing Map (SOM). We find this
comparison flawed. SOM clusters are deliberately moved
close to the reference values of their neighbors, while
BSOC tries to maximize distances between the reference
values of all clusters. Because of this, we believe that k-
means, which also tries to maximize these distances, is a
more appropriate clustering performance benchmark.
The goal of this study is to determine the relative effec-
tiveness of BSOC and k-means on a range of clustering
problems. We are using the k-means++ initialization algo-
rithm. So far, we have found that the mean clustering error
is very similar when both are applied to the color quantiza-
tion problem. We next plan to evaluate performance on the
problem of extracting convolution kernels from input im-
ages. Our final comparison will be for clustering video im-
ages, comparing k-means++ with BSOC at several differ-
ent checkpoints in the image sequence."
Hybrid Goal Selection and Planning in a Goal
Reasoning Agent Using Partially Specified Preferences
Michael Floyd, Mark Roberts, David Aha
(Knexus Research, Naval Research Laboratory)
Goal Reasoning agents are not restricted to pursuing a stat-
ic set of predefined goals but can instead reason about their
goals and, if necessary, dynamically modify the set of
goals that they will pursue. For a solitary agent, goal selec-
tion is guided by the agent’s own internal motivations.
However, an agent that is a member of a team also needs to
consider its teammates’ preferences when selecting goals.
In this work, we propose an online approach to estimate
the utility of goals based on a teammate’s partially speci-
fied preferences and use the estimated goal utilities to
guide goal selection. Specifically, we consider the situation
where an autonomous agent is teamed with a single opera-
tor in a supervisor-supervisee relationship. At the start of a
mission, the operator will provide an initial set of goals
and, optionally, partially defined preferences for how the
mission should be performed. After the initial interaction,
the agent acts autonomously to pursue the provided goals.
While acting autonomously, the agent may encounter situa-
tions that present opportunities for new goals and must de-
termine if those goals align with the operator’s preferences.
Estimated goal utilities are used during a hybrid goal selec-
tion and planning process to select a subset of goals for the
agent to pursue. We report evidence from an empirical
study which demonstrates that our approach outperforms
several baselines in scenarios drawn from a simulated hu-
man-agent teaming domain.
eSense 2.0: Modeling Multi-Agent Biomimetic
Predation with Multi-Layered Reinforcement Learning
D. Michael Franklin, Derek Martin
(Kennesaw State University)
Accurately modeling predator-prey relationships in a bio-
mimetic was can be a daunting task. Applying machine
learning to solve such a modeled task is even more diffi-
cult. What is needed is a system that uses multiple simple
models to accurately represent complex biological multi-
549
agent behavior. To do so, eSense 2.0 is being introduced.
Building on the success of the eSense BioMimetic model-
ing done in (Franklin and Martin 2016), eSense 2.0 ex-
pands the modeling to include a stronger predator / prey re-
lationship. eSense provides a powerful yet simplistic rein-
forcement learning algorithm that employs model-based
behavior across multiple learning layers. These independ-
ent layers split the learning objectives across multiple lay-
ers, avoiding the learning-confusion common in many mul-
ti-agent systems. The new eSense 2.0 increases the number
of layers and the amount of separation between agents so
that the behaviors for each agent can be more highly cus-
tomized and adds specific additional layers for behavior-
only learning. In other words, each agent now has multiple
layers to model each aspect of their behavior (e.g., obstacle
avoidance, prey observation, prey seeking, etc.). This new
abstraction of breaking various agent behaviors into multi-
ple levels furthers speeds learning and clarifies the objec-
tives the agent is considering. This builds on the general
goal of eSense (splitting out multiple agents into their own
levels) because now the agent’s behaviors are also split in-
to multiple layers. The learning is now more expressive,
faster, and less noisy. This paper presents this new multi-
level learning system for multi-agent systems and experi-
mentally confirms its performance.
firstGlimpse: Learning How to Learn
through Observation via Memory
Modeling with Reinforcement Learning
D. Michael Franklin, Ryan Kessler
(Kennesaw State University)
While there are many systems extant that take well-known
models and refine how to learn with them, there are very
few that deal with the initial stages of raw model for-
mation. How do we, as human being, learn? Can we teach
machines to 'learn to learn'? In this paper, the authors in-
troduce just such a system. firstGlimpse is a hierarchical
memory-modeling technique that can learn from scratch
via observation. It uses reinforcement learning to develop
its models and reinforce which of these models is viable,
which are most likely, etc. This hierarchical approach to
memory modeling develops from symbols to words to
phrases to sentences, thus creating a modeling system for
general learning that can be built from the ground up via
observation. As a new symbol is observed, it is recorded.
Likewise, that symbols connections with other symbols is
also recorded in another layer of the memory. Next, these
connected symbols are viewed as words, and the words are
observed as being connected to other words. In this man-
ner, phrases are added. This continues upwards to higher
levels of understanding. firstGlimpse uses the game of Set
to show, starting from scratch, that it can learn to play by
observing humans playing the game and recording its ob-
servations within its hierarchical memory. The experiments
will show that this method is successful both in observing
the links of the memory graph and in the results of success-
fully playing the game.
Data Mining Approach to Estimate Field Popularity
from the US College Scorecard Data
Fazel Keshtkar, Shiromani Neerudu, Md Suruz Miah*
(St. John's University, *Bradley University)
The College Scorecard is launched by the U.S. Department
of Education in September 2015 that provides information
about different colleges. It has a rich resource of infor-
mation intended to help students and parents to attend col-
leges in USA. The data set includes information from 1996
through 2015 for all undergraduate degree granting institu-
tions of higher education, and is updated each year. In this
paper we aim to explore The College Scorecard Data in-
formation based on college affordability and values. Our
goal is to extract patterns in these data that can provide
students to make better decisions to which college to attend
in United State. College Scorecard data is an online tool
created by U.S. Department of Education that launched in
2015. The Data contains a variety of information about
admissions, student body, cost, department, completion
and earnings by institution. We used unsupervised learn-
ing, clustering analysis, PCA, and data visualization ap-
proaches to explore, underlying the structure of data, and
pattern extraction. The model shows the percent of total
enrollment for each field ranked highest to lowest. The
Business Program was enrolled 18.23% and second highest
enrollment was Health Program with 12.83%. The analysis
will use the number of enrolled undergraduates and the
percent of degrees awarded in each field to estimate the to-
tal number of enrolled undergraduates for each program.
The PCA component is finding schools with students who
did well on the SAT score, complete college at a high rate,
and from high income.
Compensating for Rating Distribution
through Percentile Transformation
Masoud Mansoury, Robin Burke, Bamshad Mobasher
(DePaul University)
"Recommender systems use information from user profiles
to generate personalized recommendations. User profiles
are either implicitly inferred by the system through contex-
tual information, or explicitly provided by users. In the lat-
ter case, users are asked to rate different items based on
their preferences and may have individual differences in
how they use explicit rating scales: some users may tend to
rate higher, while some users may tend to rate lower. It is a
well-known phenomenon in recommender systems that ex-
550
plicit user ratings are biased toward high ratings, and the
users differ significantly in their distributions. This is usu-
ally compensated for through mean-adjustment in rating
normalization or the inclusion of a user bias term in factor-
ization models.
In this work, we propose a rating transformation model
that compensates for skew in the rating distribution by
converting ratings into percentile values as a pre-
processing step before recommendation generation. This
transformation flattens the rating distribution to reduce its
skew, which better compensates for differences in rating
distributions and improves recommendation performance.
Also, each percentile value associated with an item reflects
its rank among all of the items that the user has rated.
Thus, the percentile captures an item’s position within a
user’s profile better than the raw rating value or a mean-
adjusted version of the same, and therefore compensates
for differences in users’ overall rating behavior. A com-
prehensive set of experiments is performed, showing im-
proved ranking performance for the percentile technique
across a variety of state-of-the-art recommendation algo-
rithms in four real-world data sets."
Genetic Algorithms on Tensor Network
Contraction Order Finding
Reamonn Norat, Annie Wu, David Anekstein,
Jonathan Jakes-Schauer, Pawel Wocjan
(University of Central Florida)
We analyze the use of genetic algorithms on the problem
of tensor network contraction order finding, which can be
reduced to a permutation problem. The tensor network can
be represented as a connected graph with nodes corre-
sponding to tensors. Each edge incident to a node corre-
sponds to a particular index and each weight corresponds
to the number of values assumable by that index. Tensor
network contraction involves a series of edge contractions;
contractions occur until there is only one node remaining
in the network. Each contraction has an associated cost,
and the sum of these costs represent the approximate com-
putational cost of contracting the entire tensor network.
This total cost can vary greatly depending on the edge con-
traction order. In practice, the current state of the art is
guaranteed to find an optimal solution, but the time re-
quired will scale exponentially with network size. We in-
vestigate the use of a genetic algorithm and compare it
against the state of the art approach to tensor network con-
traction ordering. Within the genetic algorithm, we exam-
ine the performance tradeoffs among three different repre-
sentation strategies.
The Game of Chicken and Bitcoin Trading
Marius Silaghi, Badria Alfurhood, Timothy Atkinson
(Florida Institute of Technology)
Modeling large-scale social phenomena related to Bitcoin
trading and processing using theoretical games has the po-
tential to predict and even control outcomes. Efforts have
also been in the past directed to the modeling of other sig-
nificant world problems such as nuclear war, Ponzi
schemes, and the theory of money. The Bitcoin phenome-
non has overtaken the world attention through an unusual
combination of idealism and utilitarian interest. As with
other recent crypto currencies, Bitcoin trade arguably adds
intrinsic value by supporting anonymous transfers between
idealistic or criminal entities. Entities banned from selling
their coal or other energetic resources, can convert them to
anonymous value on an open and still largely unregulated
international market. There exists also a significant specu-
lative side of this trade, which alone would amount to a ze-
ro-sum game, if not for the price fluctuations and high en-
ergetic costs of transactions. This is superposed with the al-
ready studied game of the miners and traders, into a com-
plex environment influenced by available information. We
propose an agent-based simulation method based on itera-
tions where the result of each long-term simulation is used
to refine the pay-off matrix of each type of participant, to
capture the impact of media and information on the game
structure. Initial simulation inputs are currently approxi-
mated with “reasonable values” but can be refined with
corresponding social studies. The result of our simulations,
while currently heavily dependent on approximations for
initial participant structure and pay-off matrix values and
dynamics, can help improve predictability and stability in
the market.
Fairness-Aware Recommendation Systems
Nasim Sonboli, Robin Burke, Farzad Eskandanian
(DePaul University)
The basic literacy is a complex phase, composed of several
stages that require dedication to the conclusion of the pro-
cess. Nowadays, with the evolution of electronic devices
and with the advent of data mining technology, respective-
ly, the teaching and monitoring of the stages of literacy can
be facilitated by the introduction of its principles into the
educational process. Based on this, this paper proposes the
presentation of a learning object, consisting mainly of a
mobile application and a follow-up system, named Object
of study of Literacy, Objeto de Estudo de Letramento
(OEL), for teaching and monitoring the initial stages of the
process of basic literacy.
551
Comparing General and Domain-Specific LSA
Classifiers in the Context of Virtual Internships
Zachari Swiecki,1 Vasile Rus,2 Zhiqiang Cai,2
Dipesh Gautam,2 David Williamson Shaffer1
(1University of Wisconsin-Madison, 2University of Memphis)
Virtual internships are simulations that give students the
opportunity to experience realistic professional practices
and problem solving in an online environment. As part of
the virtual internship experience, students submit short
written reports of their work, which are assessed by human
raters. To implement virtual internships at scale, automated
assessment methods are needed to reduce this human ef-
fort. Our prior work trained classifiers on student data to
automatically identify domain-relevant concepts in student
reports and assess their quality. However, educators can
now develop virtual internships for new domains using au-
thoring tools. In such cases, student data does not initially
exist for classifier training. In other words, to include au-
tomated assessment models in their new internships, edu-
cators need a way to bootstrap these models before stu-
dents participate in the internships and before any student
data is available to train classifiers. To address this prob-
lem, we developed a method for generating classifiers
based on limited information provided by educators during
their authoring process, including concept definitions and
exemplar reports. Our classifiers rely on a latent semantic
analysis (LSA) approach using a domain-general corpus.
Prior results showed that classifiers developed using this
corpus had mixed performance. In this work, we compare
the performance of the domain-general LSA classifiers to
LSA classifiers developed using domain-specific corpora
of varying document and dimension sizes, and we investi-
gate how concept and exemplar features relate to perfor-
mance in each case.
Cooperation Protocols for Ad-Hoc Robot
Teams Composition in Labyrinth Exploration
Muntaser Syed, Marius Silaghi, Rajaa Rahil,
Sam Kellar, Shakre Elmane
(Florida Institute of Technology)
One’s cohesive team of robots may participate in disaster
rescue operations where other teams from other organiza-
tions/countries are also ready and willing to help. Such ro-
botic teams may not be specifically prepared to work to-
gether, but they can apply ad-hoc collaboration mecha-
nisms if appropriate communication interfaces and deci-
sion-making mechanisms are defined and supported. Sig-
nificant effort has already been laid into the design of on-
tologies and their standardization. Further, physical com-
munication interfaces have also emerged that are widely
supported, such as wireless and Bluetooth. The challenge is
to now come out with collaboration mechanisms that can
exploit such opportunities. Protocols can be standardized to
support a flexible engagement in such decision-making.
Flexibility here is understood from the perspective of the
depth and amount of details of the involvement. We take as
departing point the advancement in Distributed Constraint
Reasoning, which converged to a set of flexible protocols
based on the assumption of privacy of constraints. This
privacy assumption inspired the definition of protocols
leaving agents the flexibility of selecting their degree of
collaboration within limits that maintain guarantees con-
cerning decision-making processes. Protocols are devel-
oped for guaranteeing that the robots advance in their laby-
rinth exploration tasks while respecting physical con-
straints and pursuing progress, exchanging map infor-
mation, avoiding collisions, and reducing overlapped ex-
ploration.
Recognizing and Exemplifying
Gender Bias in Online Articles
Khonzodakhon Umarova, Eni Mustafaraj
(Wellesley College)
There is so much information online, with millions of new
websites being created continuously. The notion of credi-
ble sources is being lost, as users rely on search engines
and social networks for encountering news or fulfilling
their information needs. This has lead to an increase in the
spread of misinformation, which takes many forms and
disguises, from pseudoscience and conspiracy theories, to
fake news and propaganda. At the same time, numerous
studies have shown that young people lack skills necessary
to critically assess and evaluate online information that is
consumed so readily. Middle-, high-school and even col-
lege students struggle to identify biases in the information
sources; differentiate between fact and opinion based texts;
or recognize promotional pieces. In order to develop
web/news literacy skills and educate young people, we im-
agine augmenting search results or web pages with particu-
lar signals that will raise awareness and invite users to rea-
son about the credibility of the source. One such signal can
refer to a possible bias in the source. Examples of biases
are: ideological, cultural, or social. How to learn to recog-
nize biases automatically at web scale? We explore the fea-
sibility of a new approach, Word2vec, which computes
vector representations of words and has been shown to be
effective in determining words that are semantically related
to each other. We we will use it to identify examples of
gender bias in articles, not only to recognize such instanc-
es, but to extract concrete phrases that can be used to edu-
cate through examples.
552
Detecting Vehicular Patterns
Using a Graph-Based Approach
Sirisha Velampalli, Lenin Mookiah, William Eberle
(University of Hyderabad Campus, CA Technologies,
Tennessee Tech University)
Extracting knowledge from heterogeneous datasets is a
complicated but useful task. The Visual Analytics Science
and Technology (VAST) 2017 challenge, deals with unu-
sual activity at a wildlife preserve. It is identified that in
this preserve, there are a number of nesting pairs of Rose-
Crested Blue Pipit, a popular local, singing bird with at-
tractive plumage. In the VAST competition, one of the
challenges is to discover vehicular traffic patterns for un-
derstanding the reasons behind a decrease in the number of
nesting pairs. One approach to solve this problem is a
graph-based approach. Graph-based approaches enable one
to handle rich contextual data and provide a deeper under-
standing of data due to the ability to discover patterns in
databases that are not easily found using traditional query
or statistical tools. In this work, we present a graph-based
approach that analyzes the data for structural patterns in
the data. Our approach first reports the normative patterns
in the data, and then discovers any anomalous patterns. Af-
ter analyzing the patterns, we are able to hypothesize some
of the reasons that include: (a) most vehicles in the park
are 2-axle cars, emitting more pollution than other types of
vehicles; (b) vehicular traffic is at its peak during the
months of July and August - the prime breeding months for
these birds; and (c) traffic is at its peak in the mornings,
which effects the atmospheric conditions surrounding
when the birds sing – effecting the migratory patterns of
the birds.
An Analysis of WordNet's Coverage of Personality
Disorder Terms Using A Personality Disorder Corpus
Morgan Wixted, Amanda Hicks, Mary Cate Espinosa
(University of Florida)
The healthcare industry as a whole is turning to natural
language processing for fast solutions to complex prob-
lems. Personality disorders, a subset of the mental health
field, has limited and underdeveloped computational re-
sources that could benefit researchers if developed. Word-
Net is becoming an increasingly popular resource in the
mental health and natural language processing industries.
With the addition of personality disorder characteristics in-
to WordNet, there could be an enhanced version of Word-
Net that is designed to be more helpful to its various users,
including researchers and the general public. In this project
we analyze the coverage of various mental health terms in
WordNet gathered though corpora of six different person-
ality disorders: antisocial personality disorder, borderline
personality disorder, histrionic personality disorder, narcis-
sistic personality disorder, paranoid personality disorder,
and schizoid personality disorder. We then discuss some
key considerations and next steps for adding personality
disorder terms and their meanings to WordNet.
553