Properly.3.four.two. Incorrect predictionsFrom the 10-fold evaluation in the SVM-based predictor, there had been a total of 62 episodes resulting in incorrect predictions. Inside the following paragraphs, we describe the traits of 4 identified categories of these incorrect predictions.3.four. Qualitative AnalysisTo additional comprehend how our intention predictor produced correct and incorrect predictions within the collected interaction episodes, we plotted the probability of each glanced-at ingredient over time, aligned together with the corresponding gaze sequence received in the gaze tracker, for every interaction episode (see OPC 8212 site Figure 2 for an instance). These plots facilitated a qualitative analyses of gaze patterns and additional revealed patterns that were not 
captured in our created options but may signify user intentions. In the following paragraphs, we present our analyses and discuss exemplary instances.three.4.2.1. No intended glancesAmong the incorrect predictions, there were 23 episodes (37.10 ) throughout which the prospects did not glance in the intended components (Figure 4, First row). There are actually 3 motives that could clarify these situations. Very first, the consumers had produced their decisions in prior episodes. For instance, after they have been glancing about to choose an ingredient, they may have also decided which ingredient to order next. Second, their intentions were not explicitly manifested by way of their gaze cues. Third, the gaze tracker didn’t capture the gaze with the intended ingredient (i.e., missing data). In each and every of those situations, the predictor could not make right predictions since it didn’t have the vital info concerning the intended components.3.four.1. Right predictionsTwo categories–one dominant selection and the trending choice– emerged in the episodes with appropriate predictions (see examples in Figure three).TABLE 1 | Summary of our quantitative evaluation of the effectiveness of distinctive intention prediction approaches. Predictive accuracy Opportunity Attention-based SVM-based four.35?1.11 65.22 76.36 Anticipation time N/A N/A 1831 ms3.4.2.2. Two competing choicesSometimes, clients seemed to possess two components they had been deciding between (Figure 4, Second row). Within this case, their gaze cues had been similarly distributed between the competing ingredients. Hence, gaze cues alone were not adequate to anticipate the customers’ intent. We speculate that the determinant variables in these circumstances have been subtle and not wellcaptured via gaze cues. Therefore, the predictor was likely to make incorrect predictions in these scenarios.six July 2015 | Volume six | ArticleFrontiers in Psychology | www.frontiersin.orgHuang et al.Predicting intent making use of gaze patternsFIGURE three | Two key categories of correct predictions: one dominant choice (major) as well as the trending selection (bottom). Green indicates the components BQ-123 chemical information predicted by our SVM-based predictor that were the exact same as theactual components requested by the consumers. Purple indicates gazing toward the bread and yellow indicates gazing toward the worker. Black indicates missing gaze information.3.4.2.three. Multiple choicesSimilar towards the case of two competing alternatives, the clients occasionally decided amongst multiple candidate ingredients (Figure 4, Third row). As gaze cues had been distributed across candidate components, our predictor had difficulty in picking the intended ingredient. More information, either from different behavioral modalities or new functions of gaze cues, is necessary to distinguish the intended ingred.Successfully.3.4.two. Incorrect predictionsFrom the 10-fold evaluation from the SVM-based predictor, there were a total of 62 episodes resulting in incorrect predictions. Inside the following paragraphs, we describe the qualities of four identified categories of those incorrect predictions.three.4. Qualitative AnalysisTo additional realize how our intention predictor created correct and incorrect predictions within the collected interaction episodes, we plotted the probability of each glanced-at ingredient over time, aligned with the corresponding gaze sequence received from the gaze tracker, for every interaction episode (see Figure two for an example). These plots facilitated a qualitative analyses of gaze patterns and additional revealed patterns that weren’t captured in our created options but might signify user intentions. In the following paragraphs, we present our analyses and go over exemplary circumstances.three.4.2.1. No intended glancesAmong the incorrect predictions, there had been 23 episodes (37.ten ) in the course of which the buyers didn’t glance at the intended components (Figure four, Very first row). You’ll find three causes that may explain these cases. Initially, the consumers had made their decisions in prior episodes. As an example, once they have been glancing around to choose an ingredient, they might have also decided which ingredient to order subsequent. Second, their intentions were not explicitly manifested by way of their gaze cues. Third, the gaze tracker didn’t capture the gaze 
in the intended ingredient (i.e., missing information). In every single of those circumstances, the predictor could not make correct predictions because it didn’t have the required information concerning the intended ingredients.3.four.1. Right predictionsTwo categories–one dominant option and also the trending choice– emerged in the episodes with right predictions (see examples in Figure 3).TABLE 1 | Summary of our quantitative evaluation on the effectiveness of distinctive intention prediction approaches. Predictive accuracy Chance Attention-based SVM-based 4.35?1.11 65.22 76.36 Anticipation time N/A N/A 1831 ms3.4.2.two. Two competing choicesSometimes, customers seemed to have two ingredients they had been deciding in between (Figure four, Second row). In this case, their gaze cues had been similarly distributed involving the competing ingredients. Consequently, gaze cues alone weren’t sufficient to anticipate the customers’ intent. We speculate that the determinant components in these conditions have been subtle and not wellcaptured by means of gaze cues. For that reason, the predictor was most likely to create incorrect predictions in these situations.6 July 2015 | Volume 6 | ArticleFrontiers in Psychology | www.frontiersin.orgHuang et al.Predicting intent working with gaze patternsFIGURE three | Two key categories of right predictions: one dominant option (leading) plus the trending selection (bottom). Green indicates the components predicted by our SVM-based predictor that were precisely the same as theactual components requested by the prospects. Purple indicates gazing toward the bread and yellow indicates gazing toward the worker. Black indicates missing gaze data.three.four.two.3. Various choicesSimilar to the case of two competing options, the shoppers at times decided amongst a number of candidate components (Figure 4, Third row). As gaze cues had been distributed across candidate components, our predictor had difficulty in deciding upon the intended ingredient. Additional information and facts, either from distinct behavioral modalities or new options of gaze cues, is necessary to distinguish the intended ingred.