An approach to predict the task efficiency of web pages

Abstract

Usability is generally considered as a metric to judge the efficacy of any interface. This is also true for the web pages of a website. There are different factors - efficiency, memorability, learnability, errors, and aesthetics play significant roles in order to determine usability. In this work, we proposed a computational model to predict the efficiency with which users can do a particular task on a website. We considered seventeen features of web pages that may affect the efficiency of a task. The statistical significance of these features was tested based on the empirical data collected using twenty websites. For each website, a representative task was identified. Twenty participants completed these tasks using a controlled environment within a group. Task completion times were recorded for feature identification. The one Dimensional ANOVA study reveals sixteen out of the seventeen are statistically significant for efficiency measurement. Using these features, a computational model was developed based on the Support Vector Regression. Experimental results show that our model can predict the efficiency of web pages’ tasks with an accuracy of 90.64%.

Introduction

With the recent advancements of ICT (Information and Communication Technology), web pages have become a part of our daily life [15]. It has been reported that a total of almost 1.7 billion active websites exist to date (https://websitesetup.org/news/how-many-websites-are-there/). They cater to the different needs of almost 4.15 billion internet users (https://www.statista.com/statistics/617136/digital-population-worldwide/). Due to the widespread applications of websites with their massive users, design of web pages is an important issue [3, 22, 27, 36]. Web page design is generally associated with how people use it. In order to check the effectiveness of a designed web page, it should be tested. Usability [37] is the standard metric used to judge the efficacy of a web page. It has been reported [6] that poor usability often leads to some serious problems which include - - wastage of users’ valuable time, discouraging users from further exploration, and increasing unnecessary traffics.

An interesting thumb rule states an investment of one dollar in usability can return ten dollars [21]. Over the years, researchers have defined usability in different ways. Out of these, Nielson’s [40] approach is the most popular and most cited by different researchers.

Nielson [40] defined usability by the following five factors -

• Efficiency: It is defined as the time required to complete a task.

• Memorability: Memorability is a measure to identify the proficiency of a user when he/she uses an interface after a long time.

• Learnability: It is defined as how quickly a new user can learn a system.

• Errors: The number of errors and severeness are also identified as another key factor for usability.

• Satisfaction: Satisfaction is subjective in nature. This is defined as the aesthetics of a design.

Shneiderman et al. [46] also reported all most in the same line that usability depends on the speed of performance, speed of learning, retention over time, errors, and subjective satisfaction.

Usability evaluation of a web page is extremely important in order to judge its effectiveness. Over the years, researchers came up with different ideas to measure it. The different usability evaluation techniques are listed below :

• Reviews of guidelines

• Heuristic based evaluations

• Think aloud analysis

• Cognitive walk-through

• Computational model based evaluations

Reviews of guidelines is a simple approach where a set of design guidelines are checked after the design of a web page. Any violation of such guidelines found in during the testing is generally updated by the designer. Faraday [14] proposed a set of visual guidelines for web page design. Although this is a simple straightforward approach, it is often time-consuming which in turn slows down the development.

Heuristic-based web page evaluation needs expert designers in order to find out the issues with a web page design. There are many heuristics reported in the literature to date. Among them, Nielsen [39] proposed ten heuristics for user interface design. A detailed review of heuristic development and testing is presented in [42].

Think aloud is another technique for usability testing of any GUI including web pages. In this method, a subject has to constantly verbalize their thinking while working on a given task. Nielesen [38] reported 4 ± 1 subjects are good enough for identifying the usability issues in think aloud study. Think aloud testing can be classified into two types - CTA (Concurrent Think Aloud) and RTA (Retrospective Think Aloud). In the CTA, a subject talk about their thoughts while performing a task, whereas in RTA he/she verbalizes about a task finished earlier [5]. Alhadreti et al. [2] compared the classical CTA with its two variations - the speech communication and active intervention. They found that these three variations were helpful to identify the similar number of usability problems. Heuristic-based evaluation is also time-consuming and costly. There is also a need for expert designers in order to find out the usability issues.

Cognitive Walk-through Method (CWM) [48], relies on the idea that the subjects can characterize the user perspective in terms of usability problems. It is generally considered as same with the task analysis in an interface. Blackmon et al. [4] proposed a CWM,- termed as CWW (Cognitive Walk-through for Web). Using Latent Semantic Analysis [26], they computed the similarities among the headings/link texts of a web page with its’ goal statement. A detailed study on the evolution and state of art CWM are reported in [30].

Except for the above mentioned techniques, Computational models for usability evaluation are also another approach to measure the usability of an interface. This is also true for web pages. The main advantage of these models is they can work on the fly [33]. In other words, a web page developer can use them during the design phase. He/she can redesign the web page based on the feedback from those computational models. It has been observed that most of the reported models were developed to capture the aesthetics (subjective satisfaction) of web pages [31–35, 47]. The reported work on memorability were particularly developed for image and video - as reported in [7, 12, 49]. A computational model of learnability by using NGOMS model [24] is proposed in [28].

Over the years, many works were reported to measure the task completion times. The Keystroke-Level Model, popularly known as KLM [9] is reported long back in the year 1980. Later, in 1983, Card et al. proposed theGOMS model. A comprehensive list of these models is reported in [8]. All most all the models of efficiency measurement are forward in nature [41]. In other words, these model are developed first. These models are time consuming and assume that the users will not make any errors. Therefore to measure the efficiency of a web page these models may not be a good choice. To the best of our knowledge, we did not find any such computational model to measure the efficiency (in terms of task completion time) in the literature. Consequently, there is a necessity to develop a computational model which can help a web page designer in order to predict the time required to complete a particular task.

The hypothesis of this work is that efficiency (in terms of task completion time) of a web page can be measured and predicted, although it is subjective in nature. The main contributions of this study include -

• Initial identifications of the factors that affect the efficiency (in terms of task completion time) in a web page

• Statistical significance testing of those identified features using one Dimensional ANOVA.

• Based on those significant features and empirically collected data, proposed a computational model of efficiency.

• Tested our proposed model and observed high accuracy of 90.64%.

Related work

There are many computational models reported to predict the efficiency of an interface. Fitts and Paul [16] proposed to model the human motor performance. They have used the following mathematical formula to model the movement time (MT) from one object to another object.

$$MT=a+b{\log }_2(D/W+1)$$ 1

In the above equation D and W represent the distance and the width of the target objects respectively, whereas a and b are the regression coefficient. Hick-Hyman law [18, 20] is developed to model the reaction time (RT) of a user based on the following mathematical expression.

$$RT=a+b{\log }_{2}N$$ 2

where N is the number of options present in an interface, a and b are two constants. A comparative study of the Fitts’ law and Hick-Hyman law is reported in [45]. The author claimed that applying these rules in HCI are difficult due to the complex stimuli. Their claim was later reasserted by Liu et al. [29]. It has been reported that due to the logarithmic growth of the temporal data, reaction time can not be modeled properly by Hick’s law.

A number of cognitive models were also reported to identify the different usability issues in HCI. Card et al. [10] proposed the GOMS (Goal, Operators, Methods and Selection) model. There are different types of GOMS models reported till date. The first model in this group, popularly known as Keystroke-Level Model (KLM) was reported by Card et al. in the year 1980 [9]. They proposed seven operators as listed in Table 1. The first five operators are considered as physical or motor operators. For each operator, they proposed a predefined time. The total time duration for carrying out a task is presented as the total time required by the operators.

Table 1.

Operators of the KLM model

No	Operator	Comments
1	K	Key press
2	P	Mouse pointing to a target object
3	H	Homing between Keyboard to Mouse or vice-versa
4	B	Button press or release of a mouse
5	D	Draw a line
6	M	Mental operator required to take any mental decision
7	R	System response time

In 1983, Card, Moran and Newell [10] proposed CMN-GOMS model. The main idea of this model is to represent a task in terms of the following four items -

• Goals

• Operators

• Methods

• Selection Rules

There are other modifications of GOMS models - Cognitive Perceptual Motor GOMS (CPM-GOMS) [23], and NGOMSL [25] also reported in the literature. A comprehensive analysis of these models may be found in [8]. It has been reported [28] that all the models -KLM, CMN-GOMS, CPS-GOMS and, NGOMSL assume that the execution of a task is error free. In other words, we can say mostly they modeled the expertise users. All these models also overlook the issues of learning, decision making, and error solving.

The above-mentioned drawbacks encouraged us to explore the possibilities to develop a computational model for efficiency (in terms of task completion time), particularly for web pages. Galitz [17] in his book mentioned the differences between GUI and web pages. It has been reported the user tasks on any web page are different than that of other GUIs. In the following section, we have briefly discussed the features that may have an effect on the efficiency of a web page.

Materials and methods

Based on the observations reported in literature [16, 44] and our understandings, we have considered seventeen features, reported in Table 2 to develop the proposed model. In the following, we briefly discussed about them.

Table 2.

Features of efficiency modeling

No	Feature’s Type	Feature’s Code	Feature Name
1	Operational	NOK	Number of Key Pressed
		NOB	Number of Mouse Button clicks
		NOS	Number of Mouse Scrolled
		TDS	Total distance by Mouse Movements
2	Structural	NOO	Number of Total Objects
		STO	Size of the Target Object
		POT	Proportion of the Target Object
		TOP	Total Number of Pages
3	Visual	BGH	Background Hue
		BGS	Background Saturation
		BGL	Background Lightness
		FGH	Foreground Hue
		FGS	Foreground Saturation
		FGL	Foreground Lightness
4	Profile	AST	Average Surfing Time
		AGE	Age of a user
		GEN	Gender of a user

Feature identification

The features can broadly be categorized into the following four groups - Operational, Structural, Visual and, Profile.

Operational features

Different operations performed on a web page has a direct impact on the time required to complete a task. We considered four operational features in our study. The total number of keys pressed (NOK) is an important factor in this class. In the same line, the number of times the mouse button is pressed (NOB), and the number of times the mouse scrolled (NOS) are also considered two other important features of efficiency. The total distance (TDS) required to move a mouse from a source to a target is also considered as another parameter in our work. In order to carry out a task, a web user may have to access a number of web pages. In Fig. 1, we have presented the way adopted by us. For the first web page - represented as P₁, the distance D₁ between the top left corner from the page to the top left corner of the target object O₁ is considered. In the next web page P₂, D₂ represent the distance from the top left corner of the object O₁ to the top left corner of the target object O₂. The same logic was adopted for the next pages. Therefore, the total distance traversed (TDS) is computed using the following mathematical equation- (3).

$$TDS=\sum _{i=1}^n{D}_i$$ 3

Fig. 1.

Distance between target objects

Structural features

Web page structure is also may be a key factor for efficiency measurement. In our work, we considered four structural features for efficiency computation. A densely populated web page generally needs more visual search time in order to find out the target object than that of a web page having few objects. Therefore the total number of objects (NOO) may be an important feature for efficiency measurement. Accordingly, NOO is considered as a feature in our work. A larger size of object attracts more visual attention than a smaller object. So, the size of the target object (STO) is another feature used in our work. In the same line, the proportion of the target objects (POT) as represented in Table 2 is considered as another factor for efficiency modeling. In the following equation, TOA and WPA represent the area of the target object and the area of the web page respectively.

$$POT=\sum _{i=1}^{n}TO{A}_i/\sum _{i=1}^{n}WP{A}_i$$ 4

The total number of pages (TOP) can also be an important factor for efficiency. In order to carry out a task, if the number of the pages required to access is less then few pages need to be loaded using internet which in turn can help a user to quickly carry out his/her required task. Therefore, TOP is also considered as a feature in our work.

Visual features

Visual features include the visual properties of a web page, more specifically the target object of the page. It has been reported [34] contrast is a key factor for aesthetics modeling. Hill [19] presented an extensive work on the role of color contrast on readability. Richardson [43] shown the significant role of colour and contrast are significant for the readability of a text. Improvement of readability can be referred as the improvement in efficiency. The HSL colour model [1] is popularly used to represent the perceived colour. In our work, we considered considered six visual features of the target object - background Hue (BGH), background Saturation (BGS), background Lightness (BGL), foreground Hue (FGH), foreground Saturation (FGS), foreground Lightness (FGL).

Profile feature

The profile of a user may also be an important factor for efficiency measurement. An expert participant generally can complete a task more quickly than a novice user. In our work, we considered the average web page surfing time (AST) as a feature. Age (AGE) and gender (GEN) are also taken as two profile features in our study. Altogether, seventeen features as listed in Table 2 were initially identified to develop our proposed model of efficiency.

In order to verify the statistical significance of the features and to develop a computational model of efficiency, we performed an empirical study presented next.

Empirical study

In order to carry out the empirical study, we considered twenty web pages from different applications. For each web page, a representative task was identified, as shown in Table 3. Twenty participants volunteered for our study. The participants’ profile (minimum age = 15, maximum age = 66, average age = 28, SD age = 10.75 )is listed in Table 4. Out of the twenty participants, ten were male, and the rest were female. They belong to a wide group of professions - students (school, UG, PG), servicemen, businessmen, school teachers, professors, and doctors. Inclusion and exclusion criteria for participants’ selection are listed as below -

• Participants are regular users of computers. (inclusion #1)

• Capable of accessing internet and web pages (inclusion #2).

• 6/6 normal vision (inclusion #3).

• able to read and understand English (inclusion #4)

• differently abled (exclusion #1)

Table 3.

Representative tasks of twenty web pages

No	Type	Website	Task
W1	E-Commerce	tatacliq.com	Find the price of the smartphone Xiaomi Mi 11X 128 GB, 8 GB RAM.
W2	Educational	iimtrichy.ac.in	Search and note down the name of the CVO (CHIEF VIGILANCE OFFICER) of IIM Tiruchirapalli.
W3	Automobile	mycarhelpline.com	Find the ex-showroom price of the new car make: Hyundai, model: Creta and variant: Sx Diesel in Delhi.
W4	News	guwahatiplus.com	Find the address of the GPlus office on the webpage.
W5	Travel	visitqatar.qa/en/home	Mention one of the thing we can do while visiting Qatar.
W6	Sports	sportskeeda.com	Search the schedule for ICC World Test Championship final match in cricket category.
W7	Research Org.	isro.gov.in	Name one of the latest missions of ISRO.
W8	IT industry	tcs.com	Name one company from Retail sector partnered with TCS.
W9	World History	timemaps.com	Locate India and South Asia using ATLAS option in 2500 BCE.
W10	Music	soundcloud.com	Play the first song trending free.
W11	Financial	apanadhan.com	How much the Express Plan costs for NRIs?
W12	Flight bookings	skyscanner.com	Search for today’s best flight fare from Guwahati to Kolkata.
W13	Job	shine.com	Search for the total number of Software Development jobs present in Bangalore for Freshers.
W14	Online Videos	sonyliv.com	Go to the sonyliv originals category and play the first episode of the web series SCAM 1992.
W15	Social Net.	twitter.com	Read the latest tweet of the twitter handle @PMOIndia (you can directly use google search if you don’t have twitter account).
W16	Health	cowin.gov.in	Search and note down the available number of 18+ vaccination center by entering pincode 783370.
W17	Crowd Sourcing	worldometers.info	Find the number of active cases of Covid-19 in India till now today.
W18	Weather	accuweather.com	Search for the current temperature and air quality index (AQI) in the location Kokrajhar, Assam, India.
W19	Magazine	outlookindia.com	Find the cost of the minimum subscription plan for the digital edition.
W20	Cultural	culturalaffairs.assam.gov.in	Name one of the museums of Assam in the information and services menu.

Table 4.

Users’ Profile

Participant No	Age (years)	Gender	Profession	AST (hours)
1	25	Male	Student (PG)	15
2	24	Male	Student (UG)	12
3	31	Male	Service	6
4	25	Male	Student (PG)	12
5	32	Male	Teacher	6
6	26	Male	Teacher	8
7	25	Male	Business	12
8	66	Male	Professor(Ret.)	3
9	24	Male	Student (PG)	10
10	23	Male	Student (UG)	15
11	34	Female	Doctor	8
12	22	Female	Student (UG)	16
13	25	Female	Student (UG)	12
14	27	Female	Student (PG)	15
15	16	Female	Student (School)	2
16	15	Female	Student (School)	8
17	20	Female	Student (UG)	10
18	26	Female	Service	6
19	29	Female	Service	4
20	45	Female	Teacher (School)	5

We conducted a small training session before the start of this study. In this session, we asked the participants about their familiarity with the web pages used by us. It was observed that all the participants have never used these web pages. This was done in order to avoid any learning effect on efficiency. We also briefed them about their tasks. The participants’ profile - age, gender, occupation, and average web page surfing time were recorded in this session.

All the participants completed their assigned tasks which are reported in Table 3 in a controlled environment with a laptop having 14 inch visual display, Intel(R) Core (TM) i7-7500U CPU and Windows 10 operating system in two sessions in a single day. All of them used Google Chrome browser for carrying out these tasks. In order to avoid any biases, Latin Square Method was used. In this method, each participant P_i started with the task W_i and carried out the consecutive task. For example, the task sequence for the participant P₁₀ is

$$W_{10}->W_{11}\mathrm{.....}->W_{19}->W_{20}->W_1->W_2\mathrm{.....}->W_8->W_9$$

This implies the task sequences for all the users were unique. In order to avoid any discomfort, participants were allowed to take a break when they finished a task. Stopwatch was used to record their task completion times in terms of seconds. Altogether, four hundred data points which are the task completion times of the twenty representative tasks by the twenty participants were used for further analysis.

Feature identification

The identified features, as illustrated in Table 2, are statistically analyzed here. To do this, we have used the empirical data of task completion time. The minimum, maximum, and average task completion times, along with the standard deviation (SD) of all the twenty subjects are reported in Table 5. For example, the cell [W1, minimum] = 18 denotes that the minimum task completion time among all the twenty users is 18 seconds. Similarly, the cell [W20, Maximum] = 36 represents the maximum time taken by all users in order to complete the representative task in the web page W20. The lowest average task completion time - twenty seconds is observed for the web page number W10, whereas web page W3 needed the highest of 84 seconds.

Table 5.

Task completion time (in Seconds)

Web Page No	Minimum	Maximum	Average	Standard Deviation
W1	18	47	31.15	8.04
W2	21	70	51.25	11.54
W3	38	84	60.90	14.13
W4	23	59	41.10	8.97
W5	16	39	25.05	6.30
W6	28	58	41.90	7.67
W7	10	36	23.35	6.89
W8	46	72	59.15	7.24
W9	29	77	53.80	11.93
W10	9	40	23.55	6.76
W11	27	56	37.35	7.72
W12	20	50	32.60	8.02
W13	31	58	46.15	8.39
W14	23	65	45.00	11.38
W15	22	48	33.90	6.88
W16	14	38	24.60	6.76
W17	29	57	44.35	9.53
W18	16	34	24.40	5.16
W19	15	42	25.90	7.30
W20	12	36	21.60	6.92

Altogether, four hundred data points which are the task completion times of twenty subjects on twenty websites were used for statistical analysis. In order to carry out the statistical significance testing, we formed our null hypothesis as below -

H₀: FeatureF_i has no effect on the task completion time, ∀i,where 1≤ i≤17, i denotes the all seventeen features listed in Table 2.

We performed one dimensional ANOVA in order to check the statistical significance of the seventeen features reported in Table 2 using anova1 command of MATLAB 2021. Box plots of the two features - number of mouse buttons clicks (NOB) and total pages accessed (TOP) are shown in Fig. 2. It may be noted that the X axis denotes categories and Y axis represent the NOB and TOP for each category in Fig. 2(a) and (b), respectively. We observed out of seventeen features, except the age of a participant (AGE), all the other features are statistically significant. The observed p - values for the fourteen features - NOK, NOB, NOS, TDS, NOO, STO, POT, TOP, BGH, BGS, BGL, FGH, FGS and FGL were < 0.001, whereas the p - value of AST and GEN are 0.0232 (< 0.05) and 0.0185 (< 0.05), respectively. Therefore, it is evident that the null hypothesis for these sixteen features are not true. In other words, these sixteen features are statistically significant for the efficiency of a task. In contrary, the p - value for the AGE feature is 0.1037 (p > 0.05) satisfies null hypothesis - means AGE is statistically not significant for efficiency modeling. Based on this observation, we considered the sixteen statistical features - NOK, NOB, NOS, TDS, NOO, STO, POT, TOP, BGH, BGS, BGL, FGH, FGS FGL, AST and, GEN for the development of the proposed computational model of efficiency, as discussed in the following section.

Fig. 2.

Box plots of two features - (a) Number of Mouse button clicks (NOB) and (b) Total Pages Accessed (TOP)

Computational model of efficiency

A computational model based on the sixteen statistically significant features - as discussed in the previous section is presented here. As we were interested to predict the task completion time - which is a continuous quantity, the Support Vector Regression [13] was considered. In order to implement the proposed model, we considered the Regression Learner app of MATLAB 2021. The six different kernels - Linear, Quadratic, Cubic, Fine Gaussian, Medium Gaussian, Coarse Gaussian of support vector regression were used. The box constraint and epsilon were set to automatic. The five-fold cross validation technique was adopted for our model validation.

Results

The observed results in terms of RMSE, R-Squared, MSE, MAE, prediction speed, time are reported in Table 6. It may be noted that the minimum RMSE value is observed for the medium Gaussian kernel. Even the training time of this kernel is the second minimum among all (The minimum training time - 0.5445 seconds can be observed for the coarse Gaussian kernel). We also noticed a significant R squared validation result of 0.67 for this kernel. Accordingly, we considered the medium Gaussian Kernel for our model. A screenshot of the experimental results found using this kernel is presented in Fig. 3. The deep colour (blue) dots represent the true task completion time whereas the light colour (yellow) dots represent the predicted time by our model and the straight lines represent the observed errors.

Table 6.

Experimental results of the different kernels of Support Vector Regression

No	Kernel	RMSE	R-Squared	Speed (obs/sec)	Time (sec)
1	Linear	10.43	0.52	5800	4.1269
2	Quadratic	9.06	0.63	13000	1.5736
3	Cubic	10.20	0.54	10000	10.2120
4	Fine Gaussian	9.61	0.59	8100	0.8300
5	Medium Gaussian	8.55	0.67	11000	0.6383
6	Coarse Gaussian	10.73	0.49	170000	0.5445

Fig. 3.

True and predicted rating observed by the medium Gaussian SVM

Our proposed model can predict the efficiency of a task with RMSE of 8.55 and MAE of 6.84. The relative error of our model can be computed with the help of Equation 5. In this equation, E_R, E_O denote the relative and observed error respectively, whereas T_Max and T_Min represent the maximum and minimum task completion time. During the empirical study, we observed the T_Min was 9 seconds (for web page W10) and T_Max 84 seconds for web page W3 as reported in Table 5. Accordingly, the relative error - E_R in our study is 0.0936 by considering the observed error, E_O with the MAE of 6.84.

$$E_R=\frac{E_O}{T_{\text{Max}}-T_{\text{Min}}}$$ 5

The accuracy of the proposed model (A) can be computed by subtracting the relative error with 1 as shown in the following -

$$A=(1-E_R)\times 100\%$$ 6

Using the above formula, we got an accuracy of 90.64 % by our proposed model.

Discussion

A flow chart of our proposed work is presented in Fig. 4. In this work, we found that the efficiency (in terms of task completion time) of a task in a web page is associated with the sixteen statistically significant features - belonging to the four different groups. We also observed that although efficiency (which is key a parameter of usability) is considered as subjective, it can be measured and predicted using our proposed model. This in turn validates our hypothesis. Twenty web pages were used for the model development and validation. It may be noted that these web pages are used for different applications. Therefore, we may claim they can be considered as representative samples of web pages.

Fig. 4.

Flow chart of our work

In order to develop our proposed model, four groups of features were considered. In the operational features- Number of Key Pressed (NOK), Number of Mouse Buttons clicked (NOB), Total Distance by Mouse Movement (TDS) features were considered based on the KLM model. The NOS (Number of Mouse scrolled) is considered as a separate feature. This is because scrolling a mouse generally changes the present contents within a web page. Consequently, a user needs to rethink about the different actions to be taken by him/her which in turn may need more time to finish the desired task. Unlike scrolling, a mouse click or an enter button can lead a user to a new page. In order to address this, we considered the Total Number of Pages (TOP) as a separate feature in the structural group.

We also have a significant observation regarding the number of keys pressed. It was identified that for the fifteen out of the twenty web pages, most of the users completed the representative task without the use of keyboard. For the other five web pages, they have used keyboards. Nowadays, web designers generally try to design a task in such a way so that a user can do it quickly. A classical example can be usages of the check-boxes, radio buttons, and selection from a menu. Even there are voice-based systems getting popularity among the web surfers. In future there may be a chance that we do not have to press any key for surfing web pages.

The structural features address the complexity of a task in a web page. In a web page, a user has to select his/her desired task from a set of tasks. The total number of different tasks - presented in terms of objects (NOO) is an important factor for task completion time as reported by us. A large size object attracts users attention more than that of a smaller object. In our work, we noticed the size of target objects (STO), proportion of target objects (POT), total number of web pages (TOP) are also important factors for efficiency modeling.

The six visual features considered in our work were found statistically significant for efficiency modeling. From the informal discussion with the subjects who participated in our empirical study, we came to know that not only the size of an object but also the colour property of those objects attracted their visual attention.

Unlike most of the empirical studies where expert levels were considered, in our work we consider the Average Surfing Time (AST). We observed the AST of three participants (No-4, 10, and 14) with 15 hours which is quite high. Interestingly, all of them were students. Informally, we asked them about the reasons for such high surfing time. They reported due to the ongoing pandemic they have to do all the activities - education, and entertainment in online mode. These may be a serious matter of concern in the context of the present pandemic. Although, there was a lot of debate regarding the gender effect on usability, we found that gender (GEN) is statistically significant whereas age (AGE) is not. This may be due to the acceptability of the web platform among elderly people.

In order to develop our proposed model, we need to consider a technique that can predict the task completion time. Task completion time is a continuous random variable. Accordingly, we avoided the classification (which predicts in terms of a class, for example - good, bad) and decided to go with a regression technique. Support Vector Regression (SVR) is a popular regression technique and even we found it suitable for human data, as reported in literature [32]. Accordingly, we decided to go with the SVR for the development of our proposed model of efficiency. Unlike the existing models which are suitable for expert users, our proposed approach can model different types of users. This is because of the consideration of the two statistically significant profile features - GEN and AST for our model development.

Due to the high accuracy, our proposed model can be a good choice for the web page designer during the early design phase. Once a web page designer comes up with his/her initial design, our proposed model can be used to determine the efficiency of each task. The designer can feed the operational, structural, and visual feature values to our model. He/she can also assume a virtual user whose profile features’ values need to be supplied. Consequently, our model can predict the required time to finish the task by the virtual user.

On a web page, a number of tasks may be performed and there is always a need to how quickly we can perform a task. Interestingly, not all the tasks are often used and there may be a few tasks used mostly. In order to design an efficient website (in terms of task completion time), it may be a better idea to perform these mostly performed tasks quickly. A web page designer can check using our model whether the task completion times for these tasks are acceptable to him/her. Otherwise, if an important task needs huge time, the designer can redesign the website. Thus our model can be used during the design flow and can significantly reduce the usability testing time and acts as an indicator of efficiency in terms of a task completion time by a user.

In order to test our model, a five-fold cross-validation technique is adopted. Here the whole four hundred data points are divided into five equal groups, each having a size of eighty. One group is treated as a test case whereas the four others are used for model validation. This process was repeated for all five groups. In other words, all the groups are considered once as a test case. Cross-validation [11] is extremely beneficial in order to avoid the over-fitting of a model.

We noticed an interesting optimization problem regarding the placement of the web page objects in a website. Let us consider a situation when there is a requirement of placing N objects in a web site. This can be done in a variety of ways. A web designer can place all these N objects in a single web page. However, another designer can do it using two web pages, where one page has N₁ and the other has N₂ objects. Similarly, a designer may use n number of web pages to place the all N objects. This can be represented with the following mathematical expression as shown in equation (7).

$$N=N_1,N_2,\mathrm{.....},N_{n-1},N_n$$ 7

From the above expression, it may be noted that increasing the total number of web pages n leads to fewer objects in a web page and reduces the searching time. However, this also needs a significant navigation time which in turn can increase the task completion time.

In this work, we have tried to measure the efficiency of usability in terms of task completion time. The proposed work can be considered as a part of measuring the usability of a web page. Although, a number of works are reported on the perceived beauty - aesthetics of web pages, we found very few works were reported for the other metrics of usability - learnability, memorability, errors - particularly in the context of webpage. A future study can be carried out to identify the different features associated with these metrics. Similarly, computational models for these metrics can also be developed. It may be also an interesting area to figure out how these models are co-related.

In order to develop our proposed model support vector regression was used. Except for the support vector regression, a future study can be carried out by the other machine learning techniques. In this work, we considered web pages used on a laptop. However, a similar study can be carried out on the web pages for mobile devices (as the typing scenario is different). Our study was carried out in a controlled environment. It will be interesting to investigate the effect of an uncontrolled environment on efficiency.

Conclusion

An approach to model the efficiency of modern web page is reported in this work. Sixteen features of web pages - from four different groups were found statistically significant and used for the model development. The proposed model developed with the medium Gaussian kernel of the Support Vector Regression. In order to test our model, five-fold cross validation technique was adopted. It was observed that our model can predict the efficiency of a web page with an accuracy of 90.64% (MAE = 6.84). Due it its’ high accuracy, wide selection of different web pages (in terms of domain) and participants (in terms of gender, profession, expert level) we can claim that our model can be a good choice for the web page designer. In future, we planned to model the other parameters of usability. This in turn can help to develop a model of usability by combining all of them.

Declarations

Conflict of Interests

The authors have no relevant financial/ no financial interest to disclose.