A decision in the presence of uncertainty means that the outcome cannot be fully predicted before the decision is made. Multiple possible outcomes can occur, and our knowledge of the probability distribution only allows for a limited characterization of uncertainty. Following Knight (1921), Ellsberg (1961), and Camerer and Weber (1992), we can distinguish between two fundamental aspects of uncertainty, aforementioned ambiguity and risk. The simple definition of risk is uncertainty with known probabilities, a certain a priori probability for a given outcome. Ambiguity is also probabilistic but less well defined, generally associated with events that the decision maker has even less information about than the risk of outcomes. The two aspects are also called expected and unexpected uncertainty. Dealing with unexpected uncertainty involves a more subjective evaluation of probabilities. In the case of ambiguity, less information is available, and expected utility is harder to estimate. Not knowing crucial information, such as the probability distribution of the outcomes, is a frightening prospect which explains why most people are ambiguity-averse (Ellsberg, 1961). The two forms of uncertainty are so different that dealing with risk and ambiguity are supported by distinct neural mechanisms in humans (Huettel et al., 2006).

Apart from this probabilistic nature of decisions in an uncertain environment, there is an even deeper form of uncertainty: the kind we normally refer to in the context of quantum mechanics. Some measurements on a quantum system are simply incompatible: measuring one aspect of the system prevents us from learning more about another aspect thereof, explored by a different measurement—this is known as the Heisenberg uncertainty principle (Heisenberg, 1927). Using the formalism of quantum mechanics in decision theory is not novel either: recently there has been a surge of interest in this topic (Bruza et al., 2009; Khrennikov, 2010; Busemeyer and Bruza, 2012; Wittek et al., 2013).

We are especially interested in how risk and ambiguity appear in sequential decisions. Simultaneous or coordinated decision making, on the other hand, is more complex, being less common among animals because it involves comparative evaluation. Pointing at a major difference between the animal kingdom vs. man, Kolling et al. (2012) showed that humans are able to choose between these two models in uncertain environments. A foraging scenario is a good example of sequential decision making: food resources are available in patches, and a forager must find an optimal strategy to consume the resources. There is a cost associated with switching from one patch to another. Uncertainty relates to the quality of the current patch, the quality of background options—the opportunity cost of not foraging elsewhere–, while the environment is also subject to changes. The forager has to minimize the tradeoff between exploitation of a patch vs. exploration of background options. This pattern is not restricted to food consumption: for instance, it pertains to mate selection, retrieving memories, and consumer decisions. In fact, the same neural mechanism can serve these different functions (Adams et al., 2012).

Optimal foraging theory establishes the strategy to follow if the probabilities can be estimated and updated by the forager (MacArthur and Pianka, 1966; Charnov, 1976). Ambiguity alters the behavior: for example, unexpected forms of uncertainty may trigger more exploration (Cohen et al., 2007). We would like to see how ambiguity and risk can be minimized in sequential decisions, and how that affects exploration and exploitation.

Many decisions require an exploration of alternatives before committing to one, and exploiting the consequences thereof. This is known as foraging in animals that face an environment in which food resources are available in patches: the forager explores the environment looking for high-quality patches, eventually exploiting a few of them only. Such decisions take place in an uncertain environment: ambiguity about the quality of patches and the risk of not foraging at better patches force the forager to accept a tradeoff.

Risk-sensitive foraging is not exclusive to animals, human subjects also show similar behavioral patterns (Pietras et al., 2003; Rushworth et al., 2012). An optimal solution between exploration and exploitation is generally not known, except in cases with strong assumptions about both the environment and the decision maker (Cohen et al., 2007). The tradeoff between exploration and exploitation is also known as the partial-feedback paradigm, linking the decision model to the description–experience gap (Hertwig and Erev, 2009): people perceive the risk of a rare event differently if the probability distribution is known (decision from description) vs. when they have to rely on more uncertain information (decision from experience).

To take the next step in our working hypothesis, below we shall look at a scenario where seeking was exercised by gaze fixation at segments of user interfaces with significant elements of content, and show that underlying the seemingly random walks of eye gaze on the screen, there is order in the patterned data inasmuch as a certain typology of user behavior applies to them.

The information foraging nature of the data was recognized by eye tracking analysis, based on the concept of information scent, operationalized as “the proportion of participants who correctly identified the location of the task answer from looking at upper branches in the tree” in a study of user interactions with visualization of large tree structures (Pirolli et al., 2000). Pirolli et al. (2001) provided further theoretical accounts for scanpaths from cognitive perspectives in which users were able to find information more quickly when strong information scent was detected. Chi et al. (2001) built a computational model for user information needs and search behavior based on information scent, and the model and algorithm were evaluated by simulated studies. More recently, the modeling of user search behavior using eye tracking techniques has focused on levels of domain knowledge, user interests, types of search task and relevance judgments in search processes (Cole et al., 2010, 2013; Gwizdka, 2014; Vakkari et al., 2014; Zhang et al., 2015). However, there is still limited understanding of the effect of individual differences and user perceptions of search tasks on eye gaze patterns in information search. White (2016a,b) provided a review of information foraging and user interactions with search systems.

The eye gaze patterns, an indicator of user attention and cognitive processes have been extensively studied for designing user interfaces, such as the functional grouping of interface menu (Goldberg and Kotval, 1999; Brumby and Zhuang, 2015), faceted search interface (Kules et al., 2009; Kemman et al., 2013) and comparison of interface layouts (Kammerer and Gerjets, 2012). Information retrieval researchers have been concerned with users' attention to the ranking position of documents and different components of search engine results page (SERP) (Cutrell and Guan, 2007; Lorigo et al., 2008; Dumais et al., 2010; Savenkov et al., 2011; Kim et al., 2016). These studies generally suggest that there is no significant difference in users' eye gaze patterns on comparisons of search interface layouts, and users' attention to elements of interfaces depends on the length and quality of snippets on SERPs, as well as the displayed position of search results.

This study has been informed by uncertainty and foraging decisions in information seeking, since patches of information can be found in different elements of search interfaces in potentially uncertain search environments. The hypothesis that unexpected forms of uncertainty may trigger more exploratory behavior can be approached from user perceptions of task difficulty in information search (Cohen et al., 2007). Exploitation of a patch vs. exploration of background options in foraging can be observed by eye gaze and user search behaviors, and explained by the concept of information scent (Chi et al., 2001; Pirolli et al., 2001).

Based on the concepts of uncertainty and foraging decisions and their relationship to information seeking, we propose four research hypotheses in this study:

Participants searched on four different search interfaces, with a single search system behind the scenes. The four search interfaces were distinguished by whether MeSH terms were presented and how the displayed MeSH terms were generated, partly because we were concerned with how users search for patches of information, using different information seeking strategies such as exploitation of a patch vs. exploration of background options. The complexity of interface design characterized different perceptions of risk of information patches.

Each interface was labeled with a simple figure: a square, circle, diamond, or triangle, which was referred to in the exit questionnaire. A save icon alongside each retrieved document was provided to collect user perceived relevant documents.

The search topics used were a subset of the clinical topics from OHSUMED (Hersh et al., 1994), originally created for information retrieval system evaluation. They were slightly rewritten to read as instructions to the participants (see Figure 2 for an example). Topics were selected to cover a range of difficulties.

Participants were given brief instructions about the search task and system features, followed by a practice topic and then the searches proper. They were informed that the test collection is incomplete and out-of-date since the OHSUMED test collection (Hersh et al., 1994) was used, with MEDLINE data from 1987 to 1991. Recorded user interaction data included all queries, mouse clicks, retrieved and saved documents, time spent, and eye movements. Electroencephalogram (EEG) readings were also captured.

Background and exit questionnaires collected demographic information and asked participants about their perception of the search process. Participants' opinions of the tasks and the interfaces were sought. Finally, information on their cognitive styles was also collected by a computerized test (Peterson et al., 2003; Peterson, 2005), which took a further 15 min to complete.

The search system was built on Solr³, with the search results ranked by default relevance score. The MeSH terms were not specifically weighted.

Eye gaze data was recorded from two Sony VFCB-EX480B infrared (IR) cameras which were controlled by Seeing Machines FaceLab 4.5 software⁴ and attached to a dedicated machine running Windows 7. At the same time, EyeWorks Design and EyeWorks Record⁵ were used to present instructions for the corresponding search tasks during the experiment. Gaze points were recorded at 60 Hz, and the eye gaze data included the x and y coordinates of where the eye was looking on the screen, as well as the time of that recording. EEG data were recorded with an Emotiv headset⁶ to monitor brain activities throughout the search session. A Windows 7 computer was dedicated to the cognitive styles test.

Recordings were analyzed to see how often there were fixations in different parts of document surrogates (i.e., different elements of the interfaces), and therefore how often people looked at each part.

Four common areas of interest (AOI) were specified: title, author, abstract and MeSH (except for Interface A, without MeSH) to investigate which elements received attention. EyeWorks Analyze⁷ was used to specify the AOI, and fixations were specified as gazes within a 5-pixel radius which lasted at least 75 ms (Marshall, 2000). The EEG data was not included since our analysis was focused on eye gaze and search behaviors.

Since we were interested in the relationship between user perceptions of search tasks vs. system features, and eye gaze behavior, the data were analyzed by a logarithmic cross-ratio analysis (Fleiss et al., 2003) for dichotomous outcomes. The independent variables of user traits of cognitive styles and perceived search task difficulty, as well as dependent variables of search behavior and eye gaze measures, were converted into binary variables for further analysis. This data analysis technique was chosen because it is resistant to sample selection bias and applied in the study of the relationship between user traits and search performance (Saracevic et al., 1988).

In the study a post-search questionnaire was used to assess user perceptions about the search processes, in which search task difficulty was also identified as important moderator of eye gaze behavior (Toker et al., 2013). No follow-up interview was conducted in this study. As such, our data analysis was focused on information search in a difficult environment with a high degree of perceived uncertainty.

Table 1 reveals that there was a statistically significant relationship between user perception of search task difficulty and proportion of fixations in reading time on all elements in interfaces. One-way ANOVA indicates a significant interaction effect of interface and task difficulty on the fixations time spent in title [F_{(3, 248)} = 3.72, p < 0.05] and MeSH terms [F_{(3, 248)} = 3.71, p < 0.05], but it is not the case for the element of author [F_{(3, 248)} = 1.69, p > 0.05] and abstract [F_{(3, 248)} = 1.55, p > 0.05]. These results support our first research hypothesis that users will attend to different elements of search interfaces, when they are under perceived difficult search environments.

In the study the E-CSA-WA (Extended Cognitive Style Analysis–Wholistic Analytic) test was used to determine user's cognitive style. A Wholistic Analytic ratio (WA ratio) for each participant was produced (Peterson et al., 2003). The results suggest that there was no significant relationship between the users' cognitive style and eye gaze across all elements in interfaces in terms of proportion of fixations in reading time.

One-way ANOVA of the effects of search task difficulty, search interface and cognitive style and their interactions on eye gaze indicates significant interaction effects of difficulty and cognitive style [F_{(1, 240)} = 4.54, p < 0.05], and cognitive style vs. search interface [F_{(3, 240)} = 2.89, p < 0.05] in terms of fixation time on the element of abstract. We found significant interaction effects between search task difficulty and search interface [F_{(3, 240)} = 4.19, p < 0.01], and search interface and cognitive style [F_{(1, 240)} = 4.24, p < 0.01] for the element of MeSH terms. These results suggest that searchers with different cognitive styles may use different search strategies under an environment with uncertainty perceived as difficult and observed by their eye gaze behavior. Our second research hypothesis that user traits of cognitive styles will affect information seeking strategies in eye gaze behavior is supported.

Table 2 shows that when search tasks were perceived difficult, users tended to spend less time searching, issued less queries or typed queries, saved fewer documents and had fewer mouse clicks, but there was no difference in the number of MeSH queries issued and the number of pages viewed.

Overall, the results indicate that searchers made less mental effort when the search tasks were difficult, and they tended to optimize limited resources in information seeking, demonstrated both by eye gaze (Table 1) and search behavior (Table 2). Search behavior associated with expending mental efforts like issuing MeSH terms and viewing SERPs has not changed according to the uncertainty within the environment, such as perceived search task difficulty. Therefore, our third research hypothesis that search behavior associated with expanding mental efforts will change in uncertain environments is not supported.

Number of queries issued. Table 3 shows that there was a statistically significant relationship between the number of queries issued and the area of interest (AOI) of MeSH terms. In other words, when users issued more queries, they paid significantly more attention to the MeSH terms in search results. This might indicate that processes of query reformulations are very resource intensive, particularly for using MeSH terms.

Number of MeSH queries issued. Table 4 reveals that there was a statistically significant relationship between the number of MeSH queries issued and the element of abstract viewed. That is, when users issued more MeSH queries, they paid significantly less attention to the abstract in interfaces. This might reflect the fact that using MeSH terms can be more efficient in the search processes.

Number of mouse clicks. Table 5 indicates a statistically significant relationship between the number of mouse clicks and the element of title visited. That is, users who clicked the mouse more often were less likely to be interested in the titles.

Number of pages viewed. Table 6 brings evidence for the same inverse relationship between the number of pages viewed vs. titles inspected.

Number of documents saved. Table 7 reveals that there was a statistically significant relationship as regards the number of documents saved vs. abstracts and MeSH terms as document segments inspected: when users saved more documents, they devoted significantly more attention to the element of abstract, but less attention to the MeSH terms.

Summary of search behavior and gaze pattern types. Table 8 provides a summary of search behaviors and gaze patterns. These results clearly show that types of searching behavior, such as issuing queries with MeSH terms that imply notable mental effort and strive at the exploitation of resources, are correlated with changes in eye gaze patterns. Therefore, our fourth research hypothesis that search behavior that involve with notable mental efforts and exploitation of resources are correlated with changes in eye gaze patterns is supported.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.