>> Ashish Kapoor: So it’s my pleasure to welcome Jacopo Staiano, Saltiano --. >> Jacopo Staiano: Staiano. >> Ashish Kapoor: Staiano, my apologies, who is visiting us from University of Trento. He is very interested in research and human behavior modeling, specifically with focus on effective computing. And today he is going to talk about some of the interesting work that he has done on social networks. Okay, Jacopo. >> Jacopo Staiano: Thank you. So first of all I have to prepare you. So this talk was that I will give at [indiscernible] in a couple of weeks and therefore it’s quite short. So I will talk about this work and then I will give you some hints on other works I have been doing lately. So this is involving data from smartphone’s and afterwards I will give you just a quick outline of a data collection we have been doing using [indiscernible] sensors. And then finally I will talk about a system with a device for user experience simulation using a web cam. So it’s, there will be a jump, I have to fill one hour some how. So, okay, the title is basically what we try to do is we asked ourselves if there is any relation between the social networks that a person builds around itself and is included within. And basically we asked how effective egonetwork structure features are in predicting user’s personality? And most importantly what are the benefits of using activity-based data verses surveybased data? So personality is seen as a static and internal determinant of behavior that means the difference is our disposition towards belief, actions and attitude information. And it’s often used to explain and predict behavior. The model we have been using is the “Big Five” factor model, which basically defines the five traits that are chosen as being the most [indiscernible] as possible. And those five traits are extroversion or introversion which basically describes a person on his/her sociability verses shyness; emotional stability verses neuroticism, so if a person is calm, unemotional or anxious. Agreeableness or disagreeableness, so if friends, cooperative or it’s a faultfinding attitude. Conscientiousness on the other hand is a trait of people who are organized while unconsciousness people don’t really care much and are inefficient in doing their duties. And finally a trait related to creativity which is openness to experience. So distinguished people who are intellectual and insightful verses people who are unimaginative. So why do we go for personality? We go for personality because it’s linked to preferences and attitudes. It’s relevant for user modeling and recommendation systems. And it basically is a factor if you want to devise a persuasion system because adoptions that I might take as an extrovert person are probably driven dynamics if I compare it to a guy who has a different personality than me. One thing we have to, I want to stress about, is that there have been previous works in both the social psychology and the social computing fields. The main difference between these two fields is that while social psychologists use traditional social networks in the meaning that those networks are built using [indiscernible] which you indicate friendship ties or relational ties with other people. And from the social computing perspective we instead use social events that are basically like a call I make to someone or if I am wearing a device that tells me I am around your right now. Those are events, I am not self reporting that I am friends or that I know any of you, but there is a device that’s logging somewhere that something is happening between me and the other people that are around me. So, [indiscernible] I have found that psychological predispositions, so personality, are linked to other network characteristics. Also there have been works made on mobile phones that are related to the work on [indiscernible]. So Chittaranjan and [indiscernible] found out that actorbased features, that means code durations, how many codes they do, can be predictive of the Big Five traits. On the other hand De Oliveria and Oliver found that using those actor-based features and augmenting them with the networks structure features you up the significant performance improvements. So we basically had access to a database which was collected at MIT a couple of years ago, last year. It’s composed of an international sample of more than 400 students. And those students are married graduated students living in a student residency. And in the experiment there are this unit and this partner taking part of the experiment. So you have data also from people who are not really into that community, even though if you look at it tends to be quite [indiscernible]. So this is one of the limitations I will talk about later. There is a vibrant community life, so social events and facilities. And there are some that are sub communities, so social groups based on the ethnical or geographic province of the participants. So a rough view of the system that we use to collect the data. It’s basically this is an application that is fairly downloadable on Android systems. It’s called the funf and basically it allowed us to gather some data which I will describe in the next slides. And that data was augmented through the users Facebook app which was getting information about the people who allowed us to do that. And importantly it was, the data we had was augmented with the [indiscernible], leasing [indiscernible] of the people who were participating in the experiments. Of course all of these were complemented with the surveys that we use to submit to the participants. So the platform itself is composed of a number of probes which are, let’s say the sensors for which it interfaces with the subject. And then there are applications. There is a market application, there is basically all means to know what the user is doing with that smartphone. So if I install a new app how long it takes for you, my best friends, to adopt it too. If I keep using it how long it takes to spread over my --. This is one example. And of course all surveys and notifications, or whatever logged as well. So this work is based on the pilot phase of the study. So we isolated the pilot subjects, 53 subjects. We are using the --. We are logged for 8 complete weeks. The data we used are the Bluetooth hits, call logs and sociometric survery data. So the Bluetooth hits were basically driven on, derived from obviously the Bluetooth sensors, but they were filtered on the signal level in order to try to avoid, as much as possible, false positives. The Big Five questionnaire data was used to gather the ground truth for the personality traits. So from this data we built three networks; one was based on call logs, the other on Bluetooth and then we have the survey one. And we also merged the call and Bluetooth data by taking the intersection of the two networks and built, let’s say a digital network based on the entire digital information we had; while the last one was built on the sociometric survey, which was basically, by the traditional means, asking what your relation is with the other people. So just at a glance the distribution of the ground truth which we had used to [indiscernible] and we extracted a set of features that are related to the structures of the network the subject was in. We focused mostly on the egonetwork structure features because our hypothesis was that the structure of the ego-network of a guy who tells something that could [indiscernible] the prediction ability of a model. So we started centrality measures, small world efficiency measures and the triadic and transitivity measures. In particular, sorry this is a fixed slide. In particular we have started the standard centrality measure. Then we have started the local efficiency, nodal efficiency and then the mean local/nodal efficiency of the ego-network of the subjects. And with respect to the triadic and transitivity measures we extracted two different definitions of the triadic [indiscernible]. The first was formalized by Davis and Leinhardt’s and basically it’s account of different configurations of triplets in your own subject ego-network. While in the Kalish and Robins’ definition this configuration is based on a strong/weak, strong/weak assessment of a given tie and in a different configuration in which the ego would be present. Other than that [indiscernible] created the [indiscernible] efficient, the global transitivity and the mean/local transitivity. Meaning the mean of the local transitivity so, for alters of the ego i. So, to give quickly some of definitions; the centrality measures are the degree which is defined just as the number of alters [indiscernible]. Closeness, I don’t know if you are interested in this, but I wouldn’t spend much time on this, so it’s the [indiscernible] of the sum of the distance to the alters, betweeness, [indiscernible] which is a relative drop in efficiency when i is taken out. Then we have efficiency measures which measure basically the response of the egonet when the node, the ego is removed. And another feature which is the inverted harmonic mean of the path length. So the closer the i is to its alters the higher its nodal efficiency. And then we calculated values relative to the entire egonet of i. Finally we computed the triatic measures that I described before and transitivity measures. So our experimental setting was based on the 4 network types. So, the network based on the co-network, the co-data, the one based on the Bluetooth data, the [indiscernible] and then finally the survey one. We evaluated 8 feature types, so centrality, centrality plus efficiency measures, transitivity, [indiscernible] the combination of them. After 8 we had the binary test classification task for which we used the Random Forest algorithm and we used the bootstrap, a bootstrap strategy. So basically we embedded these in a leave-one-out setting, so basically at each iteration we will take a subject off then for other times resample the set; the data set that was remaining and then using that one as training and to predict the left one out subject. And this was suggested by a paper --. >>: What is the binary classification test for? >> Jacopo Staiano: So it’s basically for each trait. introvert. So you has extrovert, >>: I see, so you didn’t do like a 5 class classification for access? >> Jacopo Staiano: No, no, you have 5 binary classification tests basically. >>: Got it. >> Jacopo Staiano: And then we tried to explain our results by running some correlation tests, so global surveillance analysis and some analysis of marginal means. So the results are, I will skip the boring tables and go more into the juice. So, global analysis surveillance was performed. So we ranked the accuracy figures and analysis surveillance was run on a design which was 5 traits by 4 networks by 8 feature sets that we defined. What we got out of this analysis was that basically the Bluetooth network, the centrality with efficiency features on openness and agreeableness were coming out as being the best performing ones; in particular running a pairwise comparisons marginal. On marginal’s you find out that this is the ranking of the, for example, for the network types you will find that Bluetooth network would be more in informative, then the merged one, then the survey, then the call one. Then regarding the feature set, as I said before, the centrality and the centrality plus efficiency were above the other ones. And with respect to the trait then openness and agreeableness were over the extroversion and consciousness. Now going to the traits we ran 5 ANOVA’s, so one per trait. And what we found out is that basically the network type is a significant influence on all traits as you can see from this table. And the stable i accuracy of the centrality measures somehow counteracts with the good performance of transitivity measures, especially on extroversion of Bluetooth and on the other social traits. So agreeableness as you see here and openness, which kind of makes sense because those are social traits. So the transitivity gives a measure of how close together are your others. So for example this is a graft showing the performance on extroversion. The blue line indicates the Bluetooth network and this is the peak over there which indicates the transitivity set of features. >>: When the Bluetooth is telling you how, I can go back, the Bluetooth data is telling you how --. Yeah, so the Bluetooth data is telling you how extroverted they are and how are you figuring that? Is that by how many times their Bluetooth is within the presence of others? >> Jacopo Staiano: So basically the Bluetooth data is telling you, “Shall we use it as a proxy for proximity”? >>: Okay. >> Jacopo Staiano: And basically it was, so the smartphone will act in an opportunistic way. So each, at most, at least every 5 minutes it would sense who’s around and log them. And then as a post processing step you would use some filters to eliminate a lot of false positives because with Bluetooth you have like a 10 meter range or so. And if you have a really tiny wall you also have a heat. But then if you filter by the signal level, for example, and you, you can get a decent proximation of the proximity network. So for example one result that came out was that basically, as you can see the extroversion for example now here we are focusing we saw on the peak on the previous slide, so extroversion on the previous network. And you can see there is a positive correlation between extroversion and the full, the complete triad we can say. What is negative otherwise and this just makes sense because it’s a measure of the [indiscernible] efficient. And we explain these values by saying that probably extrovert people like to, let’s say filter out people who don’t fit in their network. So like extrovert people are probably better off with other extrovert, which also somehow makes sense. And it’s actually supported in the literature by a work from Hallinan and Kubitscheck. So they show actually that friendly students, so extrovert students, show a lower tolerance for intransitivity triads and tend to remove them over time, so this might be an explanation of the results we got. We also tried to compare our results with the actor based features we had. We couldn’t directly compare with other works, for example the one by Kubitscheck, because we, in our data we only have width; we don’t have for example the duration of a code. That’s not represented in our data unfortunately so we couldn’t really extract all the features they used. Anyway we extracted all the features we could from our data and what we found out was that basically the network features, network based features always led to better results on all traits. Turning to the limitation of this work, one is definitely the small sample sites because we only have 53 participants. The population is quasihomogeneous and importantly we have no logs about interaction data with nonparticipants. So we, we can’t really tell if a guy has a really sparse network because he just doesn’t use his phone for calling or because he only calls people who are not participating in the experiment. So this is a pretty strong limitation to our work. So the take away messages are basically the structural egonet features are important, look important for personality studies. The data, activity based data, seems to be more informative than surveys based data. And triadic/transitivity measures are very relevant for extroversion, especially in a [indiscernible] network like the one we built on the Bluetooth data. And centrality and efficiency were relevant for neuroticism and the agreeableness, overall on all traits, but on neuroticism the proximity based network was more relevant than the one built from call lots. So this is the first part of the talk. If I don’t have any questions on this then I will go on with the other. Yeah? >>: I think you actually answered this, but I might have missed it. So for introverted people would, if you looked at like their local social network would their social network graph have lower diameter than extrovert people? >> Jacopo Staiano: Sorry? >>: Would the social network graph of someone who’s introverted have a lower diameter than that of someone extroverted? >> Jacopo Staiano: A low diameter? So you mean there are more --? Yeah, no, it should be the opposite I think. Yes. I guess it should be the opposite because basically if we go back to the, to this one. So for example here we are talking about the extroverted people; so the opposite of the introverts. You are on either one of the spectrum. And it’s very negative [indiscernible] for example with let’s say this setting of triads or this one. Let’s say the non-closed triads. So I expect, no your right, I would be longer. You mean the diameter? >>: For extroverted, it would be longer for extroverted people? >> Jacopo Staiano: No for introverted. You mentioned that right? >>: No I said it the other way around. The diameter for an introverted person, I am sorry, should be, I feel should be shorter or smaller. And for an extroverted person it should be wider. >> Jacopo Staiano: Yeah in --. >>: So it’s opposite then? >> Jacopo Staiano: Yeah, our results suggest the opposite, because it looks like the extroverts have a tighter network so the diameter is short right? >>: Yeah, okay. It’s just, it feels like --. >> Jacopo Staiano: So let me see if you have a sparser network --. >>: I feel like if someone is extroverted they should be more likely to make friends quicker and with wider groups so they have like a longer path for their social network. So they have more [indiscernible] components in their social network. >> Jacopo Staiano: So introverts should have a bigger diameter? diameter you define the --? So by the >>: Yeah, but my point if someone is introverted and they have a small close group of friends who are all probably next to each other so their diameter is probably lower. >> Jacopo Staiano: All right, yeah, that could be the case too. >>: Okay, so yours is just the opposite of what I --? >>: I think what you are saying is like extroverts friends, all like different people, also know each other. >>: Yes. >>: So it’s shorter because it’s --. >>: Because it’s more tight-knit, so that was my intuition. are opposite of what my intuition is, so I was curious. But his results >> Jacopo Staiano: Yeah. So if there are no more questions left I will give you, so I will give you a quick glance of actually our latest work which was basically a data collection we ran with the [indiscernible] sensors. So the sensors are the ones you see over there and the SocioMetric Badges. And those sensors are equipped with microphones, the Bluetooth, an infrared sensor, and an accelerometer. What we did, again, is that basically we recruited the volunteer, actually entire search units in our private research center, and we got over 55 participants over 6 weeks that were wearing these badges Monday through Friday on the working hours. And we used an exponential sampling strategy so that every day, three times a day they would receive on their mobile phones, or on their e-mail box, a short questionnaire meant to be like 2 minutes, 3 minutes. And so we have a bunch of variables logged into survey data. We have the phone and e-mail logs and of course the sensory logs. The data we get is a behavioral from the badges and from the mail and phone logs. So, how people interact with each other by technological means or physically in the case of infrared sensors that takes, for example, the face to face communications. On the other hand we have the survey data which are pretty comprehensive and they include personality, affect and loneliness, creativity and productivity and situational items which we will use in order to reconstruct the mood, the context in which the person was acting within. So this was just a quick glance. And now we will turn to a paper I presented a couple of months ago in Newcastle which is quite different from what I said before is still related to behavioral analysis, automated behavioral analysis. In this case instead of using smartphone’s and data gathered through mobile phones we used basically a webcam and the user is actually a subject testing a new prototype for a softer system, so a whole new interface. So in contrast with the feature you have on the left there in user experience studies the emotions which are listed are typically low-intensity, mixed nature, they are dynamic and simplified. The research goal for this work was trying to understand how emotion-based mere measures could support the user experience evaluation. We tried to find the validation for new measurement tools and the overall goal was basically to build some non-intrusive methods in contrast with the EMG, for example, to seek on the subjects face in order to exploit emotion in user experience evaluation. So emotions are used to be measured in different ways. One way is the questionnaire measures, which has pros and cons. So on one end it’s ecologically valid on the other hand it doesn’t capture interaction dynamics because it only happens afterwards, when the interaction is over. On the other hand psycho-physiological measures like EMG sensors have the benefit of capturing interaction dynamics, but on the other hand they require expensive equipment and they are not ecologically valid because you have to call a subject, call him in the lab, and stick the stuff on it which might bias him and so on. So what we propose is a system that basically automatically monitors the emotional state of the users during the interaction. It is cheap in the sense that it only needs a consumer level webcam and a computer. It’s noninvasive because there is no sensors attached to the user and it can potentially be run at the user’s place. What is provided by the core of the system is basically the detection of minute changes in facial activity during the interaction. So this is how the initial prototype is designed. So in a very simple way there is the user in front of the camera. The camera gets video feed. There is a module that tracks the facial motion. And those facial motion features that are extracted, they are indexing. So basically there is a postprocessing step. Then there is a --. We devised a calibration phase and a prediction phase. Basically in the calibration phase we followed literature and the assigned two really quick tasks who were prone to elicit the reactions that we wanted to predict on the, for the automatic UX assessment. So I will go into details now. So the facial features are based on the [indiscernible] facial action coding system by Ekman, which in ‘78 defined the set of more than 40 action units that are responsible for building for facial expression taxonomies. Our system uses motion units. Motion units basically are a subset of those action units. And the relation between motion units and action units is that while action units gives you a coordinate of a certain point in time our system gives you also the [indiscernible] component so the amount of distance you had from the previous position to the next one. So this is an example that shows the muscles on the face of a human face which are the most important muscles in the [indiscernible]. Anyway, this is what the system looks like. This picture here only shows that it works, that it used to work with really low, low-fi cameras, which are not in commerce anymore right now. So that’s a good thing. That’s how it looks like for real. And now I will show you a video. So this is the tracking system in action. So basically how it works is that it fits a 3D mask on the users face and basically we use it in order to track certain points which are the ones at the joints of the face which are most related with the changes in the facial expression. Like this one is an impression that we want to take. We want to detect exactly those kinds of things, while the user is playing with a new interface. There should be another one, like this. So within those are signals that might, you know, combine some information about the problems that the user is experiencing with your new interface. Okay, so basically, sorry for the mess with the video. So basically we have our focus was on usability and in the literature we found out that some combination of features, facial feature motion, were found to be correlated with the confusion and frustrations. And since those were relevant to our study we decided to basically following the leaders, to build compound variables over the features that we had. So here is a combination of them. And so we computed those indexes for confusion and frustration, but it’s important to say that we don’t want to detect frustration or confusion. We only use previous works in order to build a new feature based on the features that we extract and then go on with it. The activation indexing stage that I spoke about before basically it uses a technique which we found it literally to be very common on EMG signals. So basically a feature would be consider active to given frame in time if its value is above it’s main plus sum deviation value. So we stick with that definition and we post-processed all the features at the start and that’s how we built our vectors of features. So we ran a couple of studies for this software and the first one we used 15 MSc students who had never used before a set of media players that we chose. We gathered performance measures and user experience self-reports, and of course we got UX_Mate output from the video data. What we found is that there were significant correlations between the data we extracted and the number of errors. We identified which were the relevant motion units which were mostly related to the mouth corners and the eyebrows, as expected. There was a weaker relation between subjective evaluation and the number of errors. And it looked like the, the questionnaires were biased by the last task which was the most challenging. Also there was a strong inter-individual variability. What we also did is that we took one-third of all the face videos that we recorded and we gave it to three observers and asked them to do UX_Mate’s job to see if she could say, “Okay, the guy is experiencing some difficulties”. And what turned out is that two of them were kind of random. One was performing well so we asked her what, do you have any specific training on the facts systems on the Ekman coding system. And it turns out she had. So we were like, “Okay, this might work”, because I mean if you were random as well it does make sense that we try to make a system that approximates that stuff. But these actually gave us the motivation to further investigate the problems. So we designed an evaluation study with twice as many students. So we got 31 undergrad’s, we used 3 academic social networking websites because we didn’t have the resources to build the interfaces. And so those were Academia, [indiscernible], and ResearchGate. And we introduced the calibration tasks. So we designed two tasks in order to get basically the training data for building our models. And while both the screen action of the user and the synchronized face video have been released so they are online if someone wants to work on it. So this is an example of the calibration faces. So basically the guy would be randomly, the subjects would be randomly shown this sequence; either first that one or first that one. The order would be random. And here following [indiscernible] you can see that the wild one is the cluttered one over there. So it’s more difficult to find a double room at the Holiday Inn in Bradley in that set up, rather than in this set up where everything is like, you know, visually organized. So we used these in order to gather our training set. This is an example of what happened to the guys. So that’s what the guy is watching. So the non-cluttered one, that’s, you know --. So there is something there. And actually when we checked the numbers we had we found that, so these motion units are the ones related to eyebrows and to the mouth corners, those are the compound ones that I talked about before, the frustration and confusion one. So you can see that in the difficult condition there are higher activation levels of the facial motion; so we ran, we built our models. We tried to predict using all of the calibration phase data. We tried to predict the actual experience of the user on the 3 networking websites. The results were not so great, even though not either too bad for a first run. So we plan to actually improve UX_Mate and this can happen, for example, in trying to fix better calibration tasks more suited for what you want to detect on the real, on the real tasks. Probably we should extend these to other interaction feelings, for example engagement during video games might be an option. And probably new compounded variables need to be identified in order to extend it. So after all this is probably your face now. So the outline is that we present a cheap, non-invasive way to try to get the automatic user experience assessment. And we released the database and hopefully somebody will use it. So, that’s it now. >>: Do you feel that there is room for improvement in terms of the actual, like, mark-up of facial features? I mean do you think that tool is good enough? And I am sorry, I like missed a little bit of how you developed that tool or what that would look like. >> Jacopo Staiano: Okay, so I think there is --. I mean the tool does a good job, but right now I think that in that field you have a lot of tools that do actually probably also a better job. So, I mean from a research point of view I wouldn’t go improving that tool as so much researching it. >>: Yeah, I mean even just like looking at the video for example, it looked like it wasn’t looking that much at the forehead where it looks like a lot of these kind of confusion wrinkles are. So, yeah I don’t know. What we are looking at here, how do you pull the actual number out from --? >> Jacopo Staiano: So what the number represents is that at each frame you would have a displacement which gives you the direction and magnitude of a movement of a certain facial feature point with respect to any [indiscernible] point, which is the moment before the calibration task there was a cross in the frame, at least there at me. The screen should --. But yeah, I mean this system is not perfect, but it works reasonably well for out purposes. It’s true that there are a bunch of other systems or commercial systems or I mean I think of the ones that Google has, Microsoft has, Facebook has. They are for sure 10 times better than this one, but we don’t have access to them so. >>: Have you considered like how this thing would play out in like more natural settings instead of like, this is fairly artificial. You bring people in --. >> Jacopo Staiano: No, this is --. but --. Okay some of them are running in offices, >>: [indiscernible]. >> Jacopo Staiano: Oh, you mean [indiscernible] here, the guy. >>: I mean it seems, I don’t know, it seems like there aren’t much distractions. The person is really like doing what they are doing right? >> Jacopo Staiano: Oh, yes. >>: But in more natural settings it might be more complicated where you are in a social situation and there are marked different types of signals that you are exposed to which --. >> Jacopo Staiano: Yeah, more noise. >>: Yeah, which --. I mean detecting emotion from facial features and those types of environments. Do you have any idea as to how --? >> Jacopo Staiano: Well the thing is that here the, the whole point of this is to have some evaluation of a new interface prototype that I have. So I would coax people in and say, “Hey”. With what we foresee is that with a system like this, which can run on your laptop, I can just give you access to my interface prototype and tell you, “Hey do it at home when you have 5 minutes”. You would do it; of course, if you started drinking your coffee and other stuff we will get random results, but yeah. It’s the trade off with the, you know, ecological ability here and --. >>: Right. >>: How much does it depend on how --? top of the monitor? I mean if I moved the camera to the >> Jacopo Staiano: The camera is on top of the monitor. >>: Oh, is it on the top or --? >> Jacopo Staiano: Yeah, this is [[indiscernible] like this. embedded camera. So you have the >>: The people are really having bad movement set-up. >>: Yeah, I mean it seems like that. How sensitive is it too --? well I guess all the [indiscernible] cameras over there, but --. I mean, >> Jacopo Staiano: You mean on the position of the camera? >>: Yeah, yeah. >> Jacopo Staiano: Well it’s, I mean it’s --. I would say like you have at least a moment. If you talk about a tail I think you have a kind of --. >>: What about if I like move this piece? Is it [indiscernible]. >> Jacopo Staiano: Yeah, I mean it gets up to I would say like this distance, like 1 meter and a half from the camera should be fine. And the problem is with the partial occlusions like if you start to do this. I mean the tracking would still be there. It is [indiscernible], but you should filter them out because you don’t need it. >>: So I am just wondering do people actually show a lot of expressions during these tasks or? >> Jacopo Staiano: It’s low intensity. It’s precious. >>: So I am just wondering is this worth spending effort into facial expression analysis when people are, I mean there are other things that you can do right? You can look at maybe their keyboard movements, maybe some physiology. That might be an easier signal than some things, just a thought? >> Jacopo Staiano: Yeah, yeah, probably, but I an [indiscernible] professor and she was like, usually done with 2MG sensors, one on the head they measure”? They measure the activation of measure because if it’s not visually you can’t mean the idea was coming from “Look arrangement studies are and one on this and what do those, which we can’t really get it through the webcam. >>: Actually the psychology literature there is a strong volume work around the facial features and that emotion, like [indiscernible] Ekman, you sited that [indiscernible]. But I think that you are right, that was the precomputer used times, so it’s not clear how those things play out in today’s information world. >>: So maybe if you had like those badges. So when people are interacting with each other it seems that people would be more expressive with their facial expression because they are trying to communicate a certain thing to a different person. So in those kinds of situations these might be more explicit. >> Jacopo Staiano: So for example, so for this, so for this study. So you have both the videos of the guy and you have the video of the screen of what the guy is doing because that’s what annotators used later to say, “Hey it didn’t take the right part to upload the photo over there”. But of course, I mean, you could devise your interface prototype or whatever, in order to log if the guy was doing a wrong action. Then you wouldn’t need facial expression, you wouldn’t need to be a self contained story. >>: Have you thought about just running this over time with what works and logging how they felt throughout their --? So I know that I get passively frustrated, I don’t realize that I am frustrated when I am working, and it would be really interesting to go back see that I was frustrated when I was writing this e-mail, or like when this person texted me, or something like that. >> Jacopo Staiano: All right. >>: Something like that? >> Jacopo Staiano: I mean, no. >>: So self reflection. >>: Yeah, yeah. >> Jacopo Staiano: Do you think you would find people to participate in experiments like that? >>: Yeah, I think people would be interested in doing it for themselves. Because I think in the social setting you have a more loudly expressed facial, you know, communication. But I think on your own you have a more honest one, because there is no one there so you probably --. >> Jacopo Staiano: Right, you are less biased. >>: Yeah. >>: I have a feeling that when I am on my own, when I don’t like have --. But I may be wrong; it might be like a subconscious thing. >>: Yeah, feel that way and then when someone walks in I am like, “Oh I must adjust my face”. >>: There is a guy who did like an art project where he constantly took pictures of himself, and he got in a lot of trouble because he took pictures of, he got all the laptops at the Apple store in Manhattan to record everyone and he got in a lot of trouble, but he said he wanted to do that project first because he recorded himself and he found that he didn’t have any change in facial expression. So that kind of goes along with what you were thinking, that people don’t really change their facial expression that much. But, it’s one example. >>: Okay, let’s thank Jacopo again. >> Jacopo Staiano: Okay, thank you.