Content uploaded by Stefano Fasciani
Author content
All content in this area was uploaded by Stefano Fasciani on Sep 23, 2017
Content may be subject to copyright.
Towards Automated Optimization of Web Interfaces and
Application to E-commerce
Aoun Lutfi
Faculty of Engineering and Information Sciences
University of Wollongong in Dubai
Dubai, UAE
al702@uowmail.edu.au
Stefano Fasciani
Faculty of Engineering and Information Sciences
University of Wollongong in Dubai
Dubai, UAE
stefanofasciani@stefanofasciani.com
Abstract—In this paper, we present a system that supports the
design of web graphical user interface by finding the optimal
placement of interactive elements. The definition of optimal
placement is context specific; it aims at maximizing measurable
aspects of the user experience, and it is derived using expert
knowledge embedded in the system, which is based on HCI
principles, user studies, and data analytics. We use a novel
modeling technique to represent the layout of web user
interfaces. The model is unsupervisedly computed on
information extracted by several image processing algorithms.
The system identifies the website category, builds a layout
model, compares it with the relevant optimal model, and
recommends an alternative layout of interactive elements. To
prove this concept, we present a study on e-commerce websites
where placement of the checkout button has a significant impact
on the online sale process conversion rate. The system identifies
non-optimal placement, and it recommends an alternative
position that is likely to improve the conversion rate. The results
will be validated based on the Visa Checkout user experience.
The system is implemented as an open-source software, and it
currently supports the re-positioning of a single interactive
element. In the paper, we discuss a further generalization of this
approach to support a larger number of interactive objects in a
wider spectrum of scenarios.
Keywords—User Interface Modeling, User Experience
Optimization, Human Computer Interaction, Complex System, E-
commerce, Image Segmentation, Patter Recognition,
Classification.
I.
I
NTRODUCTION
Providing user-friendly systems is one of the key
objectives for designers of interactive computer-based
systems. As technology advances, novel interaction
modalities emerge, providing new engagement opportunities
as well as new challenges in the design process. Previous
works on Human-Computer Interaction (HCI) have
demonstrated that it is possible to influence the behavior of
humans using computer systems [1], [2]. In the past, extensive
user studies had to be carried out to collect sufficient data to
identify effective strategies for influencing the users. Today,
most personal devices are connected to the Internet, and user-
activity logs are often collected with anonymous profiles.
Large Internet companies collect a significant amount of
information and appropriate large data analytics can identify
specific users’ preferences and patterns. This enabled the
definition of context specific HCI design guidelines to
improve system usability. The application of this expert
knowledge has opened opportunities in a wide spectrum of
fields. In this paper, we focus on web interfaces because of
their popularity, universal accessibility, and nearly ubiquitous
presence. In particular, we propose a system to automate the
embedding of expert knowledge in the design of website
layout. E-commerce, which is mostly based on web interfaces,
has recently seen a tremendous growth. Worldwide sales are
expected to double from $1.9 trillion in 2016 to $ 4 trillion by
2020 [3]. This requires that e-commerce platforms continue to
evolve to increase the demand. E-commerce companies need
to develop more sophisticated strategies to attract and
maintain customers. An approach to this issue is to improve
the conversion rate, which is defined as the ratio between
completed transactions and the total number of transactions; it
also accounts for abandoned transactions. Studies from key
players in the market suggest that User Experience (UX) and
User Interface (UI) have a major impact on the conversion
rate. User-friendly UIs and simple UXs ensure the users are
engaged and attached to the website [4]–[6], and in the case of
e-commerce, this improves the likelihood of a sale. We target
e-commerce websites to test and validate our approach
because of the availability of data and expert knowledge, the
relative simplicity and consistency of layouts, and the
presence of a parameter measuring the user experience (i.e.
the conversion rate) we aim to maximize.
A. System Overview
The process of changing an existing web UI to a more
appealing one is often time consuming and cumbersome,
because it involves several subjects and it requires several
stages, such as assessment of current interface, change
proposal, approval, implementation, verification, and
deployment. The system we propose addresses this issue by
reducing the effort of designers in implementing web layout
improvements, allowing them to focus on more challenging
and higher-level aspects of the UX process. This also provides
an opportunity to combine HCI principles with image and text
processing techniques into the field of automated UI design.
To address this issue, we propose a system that:
• identifies the type of website;
• builds a model of the website layout;
• compares the computed model against optimal
reference models, providing a quantitative
evaluation for the current design;
• and provides a recommendation to improve the
UI based on the expert rules.
In the case of e-commerce, we focus on the placement of
the checkout button, which has a significant impact on the
conversion rate. The checkout button’s optimal placement
follows different expert rules (aka. rule set) based on the type
of goods being sold online. The above scenario allows for a
complex system that combines various subsystems to achieve
a certain target. The system uses a classifier, pattern matching
algorithm, image modelling algorithm, and an optimization
algorithm. The system is partitioned in to the following
modules:
1. Website capture.
2. HTML code extraction.
3. Screenshots capture.
4. HTML based website classification
5. UI elements identification via image
segmentation and pattern recognition.
6. UI model generation.
7. Evaluation of the current layout.
Major contributions of this work are in modules 4, 5, 6,
and 7. In the fourth module we include a lightweight classifier
that helps identify the type of website using only the HTML
code of the website. The image pre-processing module
requires combining various segmentation and pattern
matching algorithms to achieve a comprehensive module that
can identify and locate all the relevant interactive UI elements
in the website layout. For the image modeling, we developed
a novel modeling technique and algorithm that are based on
human visual perception. Finally, the evaluation module
involves complex model matching and comparison techniques
to compute the UI score and provide recommendations for
improvement.
The system we developed is generic and may find
application in a wider spectrum of scenarios. The system can
be extended to model and adjust specific features of any
graphical interface. The role and nature of the image
acquisition, classification based on metadata, image
processing, and model generation would not differ, as they can
be used to describe the relevant elements in other interactive
system. It is important to note that the description is based on
human visual perception rather than computational models.
The system also relies on expert knowledge that we encode in
the last two modules of the system.
II. R
ELATED
W
ORKS
When supporting the optimization of distinct types of
website following heterogeneous design rules, it is essential to
compare the website layout against the correct reference
model. This is automatically achieved using a two stages
classifier, using keywords from the HTML content to describe
the webpage. First, the website is scrapped to obtain the
keywords, which are then used to perform the classification.
For the scrapping, the Document Object Model (DOM) tree
based approach [7] provides satisfactory results, while we
select the bag-of-words model for the text based classification
[8], [9]. Since the early 1990s, it has been shown that any
approach that involves text classification requires hard
categorization of keywords [10]. A bag-of-words approach
would satisfy this criterion because it results in a histogram
that describes all the words in the text, along with their count.
Also, the identification of a website’s interactive elements,
such as buttons and forms, requires two steps. First, the layout
is segmented and then we use pattern recognition to find
which segment is associated with UI elements. For
segmentation we found that Felzenszwalb’s algorithm [11]
provides the better performances among those considered in a
preliminary study targeting websites images. This is visible in
Fig. 1, where the segmentation results are provided also for
Simple Linear Iterative Clustering (SLIC) and Watershed
segmentation algorithms [12], [13]. It is evident that
Felzenszwalb’s algorithm segments the image along edges
and borders that describe UI elements, whereas SLIC does not
identify the background and a large element, resulting in an
excessive segmentation of the image. After segmenting,
pattern recognition is used to identify each segment.
Figure 1. Comparison of three segmentation techniques: Felzenszwalb’s
algorithm (top left) performs better than SLCI (top right and bottom left) and
better than Watershed algorithm (bottom right). The segments borders are
shown in yellow.
Scale Invariant Feature Transform (SIFT) method is a key-
point matching algorithm that is scale and orientation
invariant and capable of matching key-points across multiple
images [14]. Although faster pattern recognition algorithms
exist, such as Speeded-Up Robust Features (SURF) [15],
which is more suitable for real-time applications, SIFT
provides higher accuracy, which is the key selection criteria.
However, in some cases, key-point matching does not perform
well, especially when the template lacks in feature complexity
and in complex edges. In such case, alternative methods
include Optical Character Recognition (OCR) of text (if text
exists) [16] or analysis of color histograms.
For UI modeling, there are graph based approaches [17]
that are easily combined with pattern matching. Deep Neural
Networks with Markov Random Fields [18], and Long Short-
Term memory [19] have been used in image modeling, but
they have been shown to work best in regenerative image
processing rather than visual based modeling. Other works in
image modeling include the Fisher Vector representation,
which uses a set of low level descriptors to generate a global
model [20], [21]. In the literature, there is lack of image
modeling based on visual perspective, which can be employed
in our system. The importance of the visual based model is
essential in this work. The UI design highly depends on the
human perception of the website layout, which if considered
for, would invalidate the model itself and the overall
approach.
Figure 2. Flowchart illustrating the dataflow and dependencies between
the five key modules of the system.
TABLE I.
E-COMMERCE CLASSIFICATION CATEGORIES AND
KEYWORDS
Category Keyword
Airlines airline, boarding, ticket, tickets, travel, airplane,
fight, flights, booking
Hotels hotel, hotels, room, rooms, night, nights, booking
Tickets ticket, tickets, cinema, show, shows, performance,
performances, movie, movies
Food food, delivery, meal, meals, order, combo
Generic generic, electronics, flowers, fashion, kids, delivery,
phone, TV, computer, toy, toys, flower, florist
As mentioned earlier there is a strong correlation between
a proper UI layout and usability [1], [2], [4]–[6]. Previous
works have proposed fuzzy logic to control the UI and
dynamically mutate the layout in real-time [22], [23]. These
demonstrate that fuzzy logic control is suitable to compare a
UI design given a set of design rules. The importance of
controlling the UI can also be derived from the three
paradigms of HCI [24].
In particular, the first principle states that human computer
interaction is a classical cognitivism and information
processing problem which requires the understanding of how
one influences the other to properly understand the interaction
between both. Understanding this interaction allows some
systems to automatically change to adapt, in real-time or
offline, as we propose in this work.
III. I
NTERFACE
O
PTIMIZATION
A
LGORITHM
The entry point of the system, as visible in the dataflow of
Fig. 2, is the URL of a website. This represents the input for
the website capture, which includes both the HTML code (i.e.
the meta-data) and the screenshot image. These represent
respectively the input of the classifier module and the image
processing module. The image processing module produces a
list of relevant elements in the image, which are fed to the
model generation module. The computed visual model and the
website type are sent to both the evaluation module and
optimization module, which are also provided with a set of
website type-dependent rules on the optimal layout properties.
The result of the system is a quantitative evaluation of the
original layout, and a set of recommendations for
improvement based on the input rule set.
A. Website Capture Module
The website capture module is the entry point to the
system. It takes a website URL, and it fetches the HTML code
of the website using an HTTP request. The code is then passed
to the website classification module. We use a webdriver
(such as Selenium webdriver) to render the HTML and obtain
an image of the webpage. Then the webdriver produces a
screenshot of the webpage with a parametric geometry and
passes it on to the following image processing module. This
approach shows its main limitation with websites that
dynamically generate some of the elements after delivering
the HTML webpage. This would result in these elements not
being captured in the initial HTML request and resulting in
high likelihood of misclassification.
B. Web Classification Module
The classification algorithm developed in this system is
based on the bag-of words approach. This approach was
chosen because it provides high accuracy with text based
classification. Before applying the bag-of-words approach,
the HTML text is cleaned from tags, scripts, and styles. It is
then passed to a natural language processing toolkit to remove
simple words such as “and”, “or”, “with”, “is”. These words
are known as stop-words and are irrelevant for classification
purposes. The bag-of-words approach generates a histogram
of frequencies for each keyword. Based on this histogram, a
score is given to each category based on the number of words
assigned to each category. The categories requiring different
layout optimization rules in our proof-of-concept application
to e-commerce website are shown in Table 1 together with the
related keywords used for classification. These categories
were also chosen because the vast majority of e-commerce
websites fall into one of these. As for the keywords, they were
selected based on experimental results with a website test set,
and are sufficient to describe and correctly classify websites
of each category. The classifier then returns the type of
website and passes it to the evaluation and optimization
modules.
C. Image Processing Module
This component of the system processes the screenshot of
the website (checkout page in particular) captured in the
previous stage. At first, the algorithm segments the image
using Felzenszwalb’s algorithm, as in Fig. 3. Then, each
segment is analyzed separately to identify whether it contains
any UI relevant element. Pattern matching is used to identify
elements of unique layout, such as the Visa Checkout button
which is used for validation. The matching process is based
on SIFT to find key-points and on Random Sample Consensus
(RANSAC), which uses Homography to find the most likely
correct key-points. The module reiterates over each segment
and tests the different templates. To pass the test, templates
need at least 10 matched key-points and a statistically
sufficient number of good matched key-points defined by
Lowes ration test [14]. Lowe’s test indicates that more than
70% of the key-point descriptor mask must match. Then the
segment is localized in the image frame and we extract the
features listed in Table 2.
Figure 3. A sample e-commerce website’s checkout page segmented using
Felzenszwalb’s algorithm.
TABLE II. FEATURES EXTRACTED FROM THE SEGMENTS
KEY DESCRIPTORS AND THE METHOD OF EXTRACTION
Feature Method of Extraction
Type From template type
Boundaries Maximum and minimum
key-descriptors positions
Center Coordinates Calculated from boundaries
Dimensions Calculated from boundaries
Text (if applicable) OCR
Colors Color histogram analysis
The key-points based approach, works only with elements
that contain a larger number of key-descriptors, which are
usually detected in edges-rich segments. For simpler
elements, such as basic buttons, we take another approach.
First, a search for rectangular boxes is performed. Boxes have
been shown to represent the majority of buttons and one of the
most suitable shapes for a button [4]–[6]. After discarding all
segments that do not contain such elements, the remaining
elements are analyzed using OCR, looking for the strings:
“Checkout”, “Continue”, and “Payment”. These words
usually identify UI elements for progressing towards the
completion of the transaction [4]–[6]. For product-related
images, we take a different approach, performing a color
variance analysis in each segment. Those with high variance
likely contain an image of the product being purchased. After
performing the analysis on each segment, we extract the
features listed in Table 2.
D. Model Generation Module
After receiving the UI image elements from the image
processing, this module generates a graph based model that
describes the relation between the image elements. The model
can be described as a set of nodes connected by links. Each
node in the model will describe an element in the UI. The node
structure will be as such:
• Node Type: (Checkout or Continue or Image)
(text)
• Node ID: (number)
• Node Coordinates: (x, y) (number, number)
• Node Boundaries: (up, down, left, right)
(number, number, number, number)
• Node Dimensions: (h, w) (number, number)
• Node Text: (text)
• Node colour: (number)
The connections between each note will consist of the
distances between each centre and the direction, which results
in a 2-dimensional vector. Any node can have any arbitrary
number of links to all the other nodes describing the relative
placement of each. Consequently, comparing any two models
will involve comparing the links and nodes. As such,
comparing two links can only be done if both links connect
two nodes of the same type. The comparator computes
differences in the distance and direction (a difference vector).
The comparator also calculates the differences between nodes
as well; two “checkout” nodes can be compared and the result
would also be a comparison between the different node
components (text, dimensions, coordinates, and color). Fig. 4
shows the resulting model after the segmentation performed
as in Fig. 3.
As visible in Fig. 4, the links describe the relation between
each relevant element whereas each node describes each
relevant element. The website modelling technique we
propose here represents only elements that are relevant to a
human observer, which is a significant advantage over those
proposed in literature. However, extracting such information
can be challenging: experiments, tests and data analytics could
indicate which elements are more relevant to a human
observer in certain situations. In e-commerce in particular and
when addressing the conversion rate problem, it has been
shown that the layout of the checkout button, continue
shopping button, back button, and the shopping items have the
greatest impact on the conversion rate [1], [2], [4]–[6].
Figure 4. A sample graphical representation of the model based on the
segments in Fig. 3
E. Scoring and Evaluation Module
After computing the nodes and links of the model, the
evaluation module processes it according to the website type
and according to the provided expert rules. These rules are
parsed according to the follow syntax:
1. Colour intensity of X with respect to Y
2. Location of X relative to Y (above, below, right
of, left of)
3. Text of X contains “text”
4. Distance between X and Y
5. X alignment with respect to Y (center align, right
align, left align)
6. Size of X with respect to Y (same, larger, smaller)
X and Y in the rules indicate two different nodes
(elements) identified by the node type and optionally ID, the
keywords in bold help the parser identify the two nodes in
question, and the words in italic help identify the property
being measured. Both X and Y use the node type as a method
of identification since generally in e-commerce websites
elements, such as the checkout button, are not redundant in the
UI. The evaluation module utilizes these rules to assess the
generated model. This is done comparing each parsed rule
with all nodes and links, and checking if these nodes and links
match the rule. If so, a score is incremented. Nonetheless, the
system also requires a small margin of error to account for
minor inaccuracies in the system. The evaluation is presented
with a score out of 100.
The evaluation module also computes the differences
between the desired reference model generated by the rules
and the actual model and then uses the differences to generate
a set of recommendations to improve the UI. The
recommendations indicate the rule broken and show the
required change to satisfy the rule.
1
Available on https://github.com/aounlutfi/E-commerce-Opimization
IV. I
MPLIMENTATION AND
R
ESULTS
The system discussed above was implemented as open-
source software written in Python
1
. The system has been
tested on both Windows 10 and Linux Ubuntu 14.04.1. In the
partially-optimized current implementation modules are
executed as parallel threads when possible, and verbosity level
of the screen output can be controlled to reduce the execution
time. For the website capture we use the Python libraries
urllib2 to retrieve the html code and Selenium webdriver to
get a screenshot of the website rendering. However, due to the
dynamic scope and different configurations of individual
websites, the automatic navigation to the checkout page was
not always possible. Therefore, we include an extra input
parameter which represents the URL of the checkout page.
In the classification module, the HTML code obtained is
pre-processed using bs4 BeautifulSoup, re, collections, and
nltk. Bs4 and re are used to clean the code from tags, scripts,
and styles. And the natural language processing toolkit ntlk is
employed to remove stop-words. The histogram is generated
using the collections Counter method. The implementation of
this module has shown a 100% accuracy with a test set of 20
valid URLs and 3 outlier URLs. As mentioned before, the
limitation of this approach is the incompatibility with dynamic
HTML documents. Any change in the HTML after the code
is fetched is not considered for evaluation.
The image processing module is the most computationally
intensive module. The average execution time is about 90
seconds across all websites. The module integrates OpenCV
(for SIFT and Homography), PIL Image (for image
transformations), and skimage (for image segmentation). The
OCR functionality is implemented using the tesseract-
ocrwrapper of the Tesseract OCR engine.
TABLE III. IMAGE PROCESSING AND MODELING RESULTS
Sample Correct
number of
Elements
Obtained
number of
Elements
Model
Accuracy
False
Positives
Sample 1 3 3 100% -
Sample 2 4 4 100% -
Sample 3 3 3 100% -
Sample 4 6 4 100% -
Sample 5 2 3 66% 1
Sample 6 4 7 57% 3
Sample 7 2 2 100% -
Sample 8 1 1 100% -
Sample 9 4 3 50% 1
Image processing produced an overall accuracy of 95%.
This is due to some text elements being identified as buttons.
The pattern matching could detect all Visa Checkout buttons
with 100% accuracy. The accuracy of the computed model
with respect to the website layout is strictly dependent on the
performance of the element identification in the image
processing module. Table 3 shows the model accuracy for
each of the 9 samples used to assess the system. As it can be
seen in the table, most of the samples were 100% accurate.
Three samples present false positive mostly due to elements
(especially text) visually similar to buttons. These false
buttons can be considered constructively towards the UI
improvement goal. In fact, these suggest that the website may
contains visual elements that can be mistakenly identified by
some user. The data set used to assess the image processing
and modeling algorithms is composed of 9 sample
screenshots, one of which is an ideal implementation
(template) of the Visa Checkout button.
After computing the model of the website’s layout, the
system evaluates it against the following rule set (provided
just for validation):
• element 1 is visa checkout
• element 2 is checkout
• element 1 right of element 2
• element 1 less than 600px from element 2
• element 2 contains "checkout"
As expected, the template Visa Checkout implementation
scored 100%, where as other implementations resulted in a
lower score. It is interesting to note that despite some
implementations were correct for their own website type,
when compared to this specific ruleset for the “generic”
website type, they produced a lower score. This highlights the
importance of the classifier to identify the type of website
using the correct set of rules for that type. The novelty of our
approach and the lack of similar modelling techniques of
graphical layout make it difficult to compare our system
against others. However, our evaluation is reliable as it was
carried out using reference data assessed by experts.
V. C
ONCLUSION
In this paper, we presented a generic method to optimize
the design of web interfaces by finding the placement of
interactive elements that maximize measurable usability
parameters. The method has been implemented as an open-
source software and customized to automatically optimize the
position of the checkout button in several types of e-
commerce websites using Visa Checkout guidelines. The
algorithm is based on the combination of several image
processing techniques and a novel perceptually-relevant
visual model of the website’s layout. Results have shown that
the system operates at a high degree of accuracy. In future
work, we will address technical improvements of the system.
The current image processing module can be enhanced to
improve accuracy, reduce execution time, and detect more
complex elements such as tables, fields, and text blocks. This
can be achieved by optimizing the current implementation or
by integrating additional image processing techniques. The
evaluation module will be extended to automatically modify
the HTML code to implement the repositioning of the
interactive element following the expert knowledge
embedded in the system. Finally, the current implementation
can be extended to support the simultaneous optimization of
multiple interactive elements, suitable for more complex
application scenarios and advanced interactive systems.
R
EFERENCES
[1] J. D. Gould and C. Lewis, “Designing for usability: key principles
and what designers think,” Communications of the ACM, vol. 28,
no. 3, pp. 300–311, 1985.
[2] B. Shneiderman and C. Plaisant, Designing the User Interface:
Strategies for Effective Human Computer Interaction, Fourth.
Pearson Addison Wesley, 2004.
[3] “Worldwide Retail Ecommerce Sales Will Reach $1.915 Trillion
This Year,” eMarketer. [Online]. Available:
https://www.emarketer.com/Article/Worldwide-Retail-Ecommerce-
Sales-Will-Reach-1915-Trillion-This-Year/1014369. [Accessed:
05-Mar-2017].
[4] C. Harshman, “A/B Test Ideas for E-Commerce Call to Action
Buttons,” Optimizely Blog, 07-Dec-2013.
[5] C. Holst, “Fundamental Guidelines of E-Commerce Checkout
Design,” Smashing Magazine, April 2011.
[6] P. Laja, “How to Design an Ecommerce Checkout Flow That
Converts,” ConversionXL, February 2014. .
[7] V. B. Kad am and G. K. Pakle, “A S urvey on H TML Str ucture Aware
and Tree Based Web Data Scraping Technique,” International
Journal of Computer Science and Information Technologies, vol. 5,
no. 2, pp. 1655–1658, 2014.
[8] Z. S. Harris, “Distributional structure,” Word, vol. 10, no. 2–3, pp.
146–162, 1954.
[9] Y. Zhang, R. Jin, and Z.-H. Zhou, “Understanding bag-of-words
model: a statistical framework,” International Journal of Machine
Learning and Cybernetics, vol. 1, no. 1–4, pp. 43–52, 2010.
[10] F. Sebastiani, “Machine Learning in Automated Text
Categorization,” ACM Computing Surveys, Oct. 2001.
[11] P. F. Felzenszwalb and D. P. Huttenlocher, “Efficient graph-based
image segmentation,” International Journal of Computer Vision, vol.
59, no. 2, pp. 167–181, 2004.
[12] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Süsstrunk,
“SLIC Superpixels Compared to State-of-the-Art Superpixel
Methods,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 34, no. 11, pp. 2274–2282, Nov. 2012.
[13] Jianbo Shi and J. Malik, “Normalized cuts and image segmentation,”
IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 22, no. 8, pp. 888–905, Aug. 2000.
[14] D. G. Lowe, “Distinctive image features from scale-invariant
keypoints,” International journal of computer vision, vol. 60, no. 2,
pp. 91–110, 2004.
[15] H. Bay, T. Tuytelaars, and L. Van Gool, “Surf: Speeded up robust
features,” in European conference on computer vision, 2006, pp.
404–417.
[16] R. Smith, “An overview of the Tesseract OCR engine,” in Document
Analysis and Recognition, 2007. ICDAR 2007. Ninth International
Conference on, 2007, vol. 2, pp. 629–633.
[17] P. Foggia, G. Percannella, and M. Vento, “Graph Matching and
Learning in Pattern Recognition in the Last 10 Years,” International
Journal of Pattern Recognition & Artificial Intelligence, vol. 28, no.
1, p. 1, Feb. 2014.
[18] Z. Wu, D. Lin, and X. Tang, “Deep Markov Random Field for Image
Modeling,” in European Conference on Computer Vision, 2016, pp.
295–312.
[19] L. Theis and M. Bethge, “Generative image modeling using spatial
lstms,” in Advances in Neural Information Processing Systems,
2015, pp. 1927–1935.
[20] J. Sánchez, F. Perronnin, T. Mensink, and J. Verbeek, “Image
classification with the fisher vector: Theory and practice,”
International journal of computer vision, vol. 105, no. 3, pp. 222–
245, 2013.
[21] K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside
convolutional networks: Visualisin g image classification models and
saliency maps,” arXiv:1312.6034, 2014.
[22] A. Agah and K. Tanie, “Intelligent graphical user interface design
utilizing multiple fuzzy agents,” Interacting with computers, vol. 12,
no. 5, pp. 529–542, 2000.
[23] S.-M. Chen and J.-M. Tan, “Handling multicriteria fuzzy decision-
making problems based on vague set theory,” Fuzzy Sets and
Systems, vol. 67, no. 2, pp. 163–172, Oct. 1994.
[24] S. Harrison, D. Tatar, and P. Sengers, “The three paradigms of HCI,”
in Alt. Chi. Session at the SIGCHI Conference on Human Factors in
Computing Systems San Jose, California, USA, 2007, pp. 1–18.
