“I think you need help! Here’s why”: Understanding the Effect of Explanations on Automatic Facial Expression Recognition

FER is a widely adopted, non-contact method for emotion recognition and has garnered growing research interest in the past few years [43, 16, 6]. FER has been shown to be naturalistic, unobtrusive, economical and easy to deploy and maintain using commercially available sensing systems [26, 22, 27, 35]. One commonly adopted FER method is the Facial Action Coding System (FACS) [14]. This method encodes human facial expressions as facial muscle activations associated with a person’s emotional state. Each FAU represents a visible facial muscle movement. We used this method as it provides a formal representation of facial expressions through muscle activations which are easy to explain and grounded in anatomy.

Literature in FER has focused on improving classification accuracy and rarely investigates the perception, reliance and trust of users [11]As the field continues to advance rapidly, it is crucial to provide a deeper understanding of machine learning models so that end users can have a greater sense of autonomy when using these systems [1]. Aside from enhancing the system’s intelligence and social acceptance, empowering users to make informed decisions and retain control is equally vital.

Explanations are crucial for imparting knowledge and sharing experience among people [24]. It can be defined in multiple ways such as an assignment of causal responsibility [23] or simply as an answer to a “why-question” [18]. In the field of artificial intelligence (AI), explainability (also sometimes equated with interpretability) [13] helps to create a shared understanding [32] between end users and the AI system that is being used.

The two main approaches to explanation generation in XAI are intrinsic explanations and post-hoc explanations. Intrinsic explanations can be generated for white-box models, systems that inherently have some amount of interpretability “built-in” that accompanies the model’s output [34], whereas post-hoc methods are techniques applied “after” a model’s prediction is made [13]. In our research, we focus on intrinsic explanations for any FER model that utilizes FAUs as inputs to the expression recognition system. Specifically, we use SVMs [9] to perform FAU prediction that feeds into an expert system [36] for emotion prediction. We then generate human-readable explanations for the predicted emotions based on the activated FAUs to improve system interpretability (seen in Sec. IV-B).

Explanation can also vary depending on the type of insight provided to users. These are, namely, Global and Local explanations [13, 32]. Global explanations provide the user with general knowledge, such as the features a model uses to make a decision. Local explanations on the other hand provide knowledge about the specific prediction that was made by a model. As there has been no prior literature on which explanation type would be beneficial to end-users in the context of FER, in this research, we use both global and local explanations to provide a complete understanding of our system to end-users. More details on the global and local explanation design is provided in Sec. IV-B as it is embedded in the testing environment.

Miller [32] provides insights on what constitutes a good explanation, specifically that “probabilities and statistics don’t matter” and that “explanations are social”. We incorporate these insights by showing the user which FAUs lead to a prediction accompanied by natural language explanations. This allows the explanation to take the form of a conversation or interaction with the user.

Past studies of XAI found explanations to improve users’ understanding of AI systems [5, 8]. However, empirical results regarding its impact on users’ subjective experience such as trust [44, 41] and acceptance [17] are varied and particularly rare when it comes to the domain of FER.

In the context of explanations for expression recognition, very few research publications have been produced. One example of how explanations of facial expressions can be beneficial is shown in [15], who used patient-clinician simultaneous fMRI scans synchronised with facial expression recordings in order to correlate active brain regions with FAUs associated with pain stimulus. The authors used SHAP [30], to determine the features most highly associated to pain and correlated them to studies that identified brain regions most associated with pain. While this work is relevant in understanding how pain is related to involuntary facial expressions, it may not generalize to other interaction scenarios. However, this work demonstrates that explanations can be utilised to extract context-specific knowledge necessary to better understand end-users.

Most relevant to our work are [38] and [10], in which the authors themselves received explanations (generated through Grad-CAM, SoftGrad, LIME and CEM methods) on emotions identified by CNN models and proceed to qualitatively analyse the explanations. The explanations highlight the features (superpixels) which contribute most to the predictions being made. The authors distinguish between the different XAI methods through qualitative comparison obtained by the authors themselves. However, how helpful these explanations are to an average, non-expert user remains unclear. Crucially, these studies do not empirically evaluate the perceptions of non-expert users and the usability of the explanations with regards to the emotions predicted and the underlying mechanisms of the system. Moreover, these studies do not evaluate the performance of FER models in a real world setting, nor did they measure perception and trust of the users towards the FER when exposed to explanations [39, 33]. Although there has been work in comparing some XAI techniques in the literature [10] with regards to FER, utilizing FAUs to explain facial expression in an empirical setting has not been carried out to the best of our knowledge and their current benefits or usability are unknown.

Various evaluation methods have been proposed to evaluate the effectiveness of XAI systems from the end users’ perspective [21]. These evaluations have not been conducted in the context of explanations for FER, which is the focus in this work. By moving experiments from a human-grounded approach to an application-grounded approach  [13], we expect effective XAI systems can be developed to better suit the context of emotion recognition. In this work, we leverage these evaluation methods and construct an empirical user study to evaluate the need and effects of explanations on FER systems, thereby providing a foundation for future research on explaining FER systems to end-users.

We implemented a FER system that streams video data and detects faces using the face predictor implemented with the dlib library [25]. We then utilize OpenFace’s SVM library [3] to predict FAU activations from the detected facial region. FAU activations are fed into the HERCULES production rules expert system [36]. This expert system computes the confidence of the facial expression being associated with each of the “Big-6” emotion categories: happy, sad, angry, surprised, disgusted, fearful. If the confidence value of all emotions is below 50%, the system classifies that facial expression as neutral. Otherwise, the emotion with the highest confidence value is selected as the classification output. The system has an accuracy of 86.3% when tested against annotations by 5 certified FACS coders [36] and 89% on the benchmark Cohn-Kanade dataset [16]. The output of the emotion prediction is used by the Hint System as seen in Fig. 1.

The Hint System stores the emotions inferred within the last 2 seconds and uses majority voting to determine when to trigger a hint for the user. If the majority of the emotions inferred within the 2 second window was negative emotions (anger, fear, sadness, disgust) or surprise, a hint is triggered.

The Hint System was tested in a separate pilot study with 20 participants who did not participate in the main study. It was conducted as a within-subject experiment to measure the usability of the hint mechanism. The pilot study tested the use of a manual hint button compared with the proposed automatic hint mechanism triggered by facial expressions. Each participant interacted with 3 versions of the system in random order – the autohint system using a 2-second timing window (decided based on previous research [31, 7, 37]), a manually triggered hint system and a system without any hints. A pairwise Wilcoxon signed-rank sum test was used (non-normally distributed data) to compare the means of the game score, number of collisions and hints triggered between each version. The results of the test indicated no significant difference (p $>$ 0.1) between the manual and autohint system. This suggests that the autohint system has similar usability to the manual hint system and therefore we adopted the autohint system in the main study.

The task chosen for this study involves a multi-agent navigation game referred to as the Explorer Game (see Fig. 4a). The participant (blue dot) navigates in a grid-world with obstacles (grey boxes) and enemy agents (red dots). The aim of the game is to navigate to the end zone (green box) while making the least number of moves possible and also avoiding collision with enemy agents or obstacles. The participant begins with 30 fuel units and can move up, down, left and right (not diagonally) or skip their turn. The enemy agents’ moves are random and cannot be predicted by the participant (forcing the participant to place more reliance on the hint option which will be explained in the next section).

Each move the participant makes consumes 1 fuel unit. Colliding with an obstacle or the edges of the grid-world results in a penalty of 1 fuel unit, while colliding into an enemy agent results in a penalty of 3 fuel units. Each game instance had 16 obstacle boxes and 3 enemy agents. If the participant runs out of fuel, they lose that particular game trial and continue to the next. Winning a game involves getting the participant’s agent to the end zone before the fuel runs out. The amount of remaining fuel after completing all trials determines the participants’ score and was used to populate a leader-board which was shared among all participants. Each participant played 9 trials each with different maps configurations.

During the game, if the participant triggers the hint function through the mechanism explained in the previous section, the system will generate a pop-up window stating that the participant may need a hint and asking if the participant would like one (as seen in Fig. 4b (top)). If the participant selects “Yes”, we consider this as hint “acceptance”. If the participant selects “No”, the window will close and the participant continues to play the game. For participants who “accepted” the hint, they are then given information on the enemy agents’ next moves as well as a recommended move to make (as seen in Fig. 4b (bottom)). If the participant selects “Yes” to follow the recommended move, the system will automatically make that move for the participant. This is defined as hint “compliance”.

In addition to the Explorer Game, we designed a Test Game to evaluate the participants’ knowledge of the FER system. This gives us an opportunity to estimate each participant’s particular understanding of how the FER works, following their experience with the Explorer Game. In the Test Game, there are no enemy agents (red dots), obstacles (gray squares) or hints. The agent has a pre-planned set of moves used to get to the end zone and the keyboard is disabled. The participant can only move the agent by triggering the FER system with the same facial expressions used to trigger the hint system in the Explorer Game. The objective of the Test Game is to get as far as possible within a time limit of 1 minute. The farther the participant moved, the higher the score. This Test Game enabled us to objectively measure the users’ understanding of the FER system, through the score.

The study was designed as a between-subject study with 2 cohorts:

1. 1.

   AutoHint Participants were provided with hints only (no explanations). Participants were instructed prior to the experiment that video data will be used to detect emotions which, in turn, trigger the hint system. No further details were provided.

2. 2.

   XAutoHint Participants are exposed to global (mandatory) as well as local (optional) explanations of the FER system explained in Sec. IV-B.

Participants performed the experiment individually on a laptop in a private, well-lit room. An onboard camera was used to capture the video stream for processing. Participants used the onboard keyboard and a mouse to interface with the system. After a few practice runs, both cohorts proceeded to play the same set of maps for the Explorer Game with a randomized map order to control for any ordering effects. Participants in the XAutoHint cohort were able to view local explanations as described in the previous section. After playing the Explorer Game, participants proceeded to play the Test Game which aims to evaluate their understanding of the FER. During the games, the experimenter exited the testing room to ensure that participants could express themselves freely and to reduce experimenter bias. After both games were completed participants answered the the Trust Scale [21].

After the questionnaires were completed and prior to conclusion, participant exit interviews were conducted. Participants were asked to explain how the hint mechanism worked and which facial muscles or emotions were needed to trigger the hint mechanism in the exit interview.

Thematic analysis [4] was used to evaluate the interview responses and identify themes and patterns in the data. Following the thematic analysis method outlined by [4], the qualitative data from the interviews were analysed in depth to extract common themes. We identified themes using a data-driven/inductive approach. The themes were updated throughout the review of all interview data. We reported the identified themes and added relevant extracts from participants’ own ‘verbatim’ transcripts within an analytic narrative [2]. This narrative of the qualitative analysis is presented in Sec.  VI-B.

The study was conducted with a total of 20 participants, 10 participants in each cohort (5 M, 5 F). No significant differences were found between the cohorts in terms of age ($U=44.5$, $p=0.70$), gender ($U=50$, $p=0.97$), self-reported navigation capabilities ($U=38$, $p=0.38$) and self-reported emotional expressiveness/suppressiveness ($U=42.5$, $p=0.60$) using a Wilcoxon rank-sum test.

The in-game performance data for the Explorer Game was analysed to evaluate H1 (A user will have better task performance when provided with explanations). Since the hints helped users avoid collision with enemy agents we compared the amount of collisions experienced by participants in each cohort. We found a statistically significant difference in the number of collisions ($W=5434.5$, $p<0.001$) between both cohorts with the XAutoHint cohort experiencing a significantly lower number of collisions (average collision of 0.32) compared to the AutoHint cohort (average collision of 0.99) as can be seen in Fig. 5 ²²2This significance held during post-analysis when outliers were removed ($W=3967.5$, $p-value<0.001$).). The size of the effect was medium with a Cohen’s d value of 0.681.

When analysing overall game performance as determined by the game scores (amount of fuel leftover at the end of the game) and completion rate we found no significant difference for the scores ($W=3522$, $p=0.13$) and completion rate ($W=53$, $p=0.84$) using the Wilcoxon rank-sum test. This, however, is not necessarily a reflection on the helpfulness of the hints since fewer collisions does not directly correlate to less fuel usage. The hints allowed participants to avoid collisions (which have a higher penalty), however the extra moves made to avoid collisions still costs fuel (lower penalty) and may have lowered the scores of participants in the XAutoHint group which resulted in no significance being detected for the overall scores or completion rate.

We next attempt to address H2 (A user will report and demonstrate better understanding of the system when provided with explanations). This can be ascertained by comparing the number of hints triggered during game execution between both cohorts. A higher number of triggered and accepted hints indicates a better understanding of which expressions can trigger the system and which cannot. The number of hints triggered in the Explorer Game is shown to be significantly higher in the XAutoHint cohort ($W=1906.5$, $p<0.001$) as can be seen in Fig. 6a ³³3This significance held during post-analysis when outliers were removed ($W=1636.5$, $p-value<0.001$).. The average number of hint triggers for the XAutoHint cohort was 2.76 whereas for the AutoHint cohort it was 0.66. The size of the effect was large with a Cohen’s d value of 1.03. We further evaluated the overall scores of the Test Game which is determined by the number of times the hint system was activated. When comparing between the AutoHint and XAutoHint cohorts, the scores for the Test Game are significantly higher for participants in the XAutoHint cohort ($W=17.5$, $p=0.015$) indicating more hints were triggered, as seen in Fig. 6b. The average score was 53.8 for the XAutoHint cohort and 26.7 for the AutoHint cohort. The size of the effect was large with a Cohen’s d value of 1.23.

Qualitatively, we compare the interview responses from participants explaining how the hint function works in their own words. For the AutoHint group, the themes found indicated that participants in this group have a ‘vague’ and sometimes ‘incorrect understanding’ of what emotions/expressions seem to trigger the hint mechanism. Although 4 out of 10 participants in this group mentioned that the “eyes getting smaller, not happy face, and squeeze eyebrows” (P1, P5, P13) seem to trigger the hint, 9 out of 10 of the participants also mentioned a mix of different thoughts that indicate incorrect understanding of how the hints work. For example, some mentioned that a “thinking, anxious, concerned or even confused face” (P3, P9, P13, P15, P17, P19) seemed to work in triggering the hints. For the participants in the XAutoHint group however, the understanding of the system is more ‘concise’ and ‘accurate‘. All 10 participants mentioned that the emotions that worked for them were “mainly negative emotions” (P6) such as “disgust, surprise, anger, fear” (P8, P10, P12, P14) whilst others went even further to mention facial movements that worked for them such as “raising/lowering eyebrows, tightening their mouth or opening it, and opening their eyes wider” (all XAutoHint participants).

To evaluate H3 (A user will report higher trust towards the system when provided with explanations), we analysed perceived trust as obtained from the survey data of the Trust Scale [21]. The overall Trust Scale scores are significantly higher for the XAutoHint cohort ($W=17$, $p=0.014$). The size of the effect was large with a Cohen’s d value of 1.29.

A further indication of demonstrated trust towards the FER is obtained by analysing the number of hints “accepted”. To compare the hint acceptance (measured by Yes/No button clicks) we performed a Welch’s T-test, as the number of hints triggered were different for both groups and equal sample tests could not be carried out. The results indicate that participants demonstrated a significantly higher hint “acceptance” ($t=-5.7162$, $df=40.694$, $p<0.001$) in the XAutoHint condition ⁴⁴4This significance held during post-analysis when outliers were removed ($t=-7.7042$, $df=32$, $p<0.001$).. The analysis of the data supports H3, indicating that users have a higher trust towards the system when provided with explanations. We also note that participants in the XAutoHint condition had higher hint “compliance” ($t=-2.3647$, $df=19.77$, $p=0.028$) ⁵⁵5This significance held during post-analysis when outliers were removed ($t=-3.2929$, $df=18$, $p=0.004$)..