>> Jonathan Grudin: Okay. Thank you for coming to the talk today. We'd like to welcome Conrad Albrecht-Buehler, if I pronounced that correctly. >> Conrad Albrecht-Buehler: Yep. >> Jonathan Grudin: Who I met -- I visited Northwestern, where he just recently got his PhD working with Don Norman and others and I visited Northwestern about a month ago and afterward had a conversation with Conrad, just prior to his actually defending his dissertation, where I realized that the overlap of his interests and interests in groups here was very high. And so when I got back I invited him to come and he begged off on showing up until after he defended his dissertation on Halloween. But he did that successful and now he's here to talk about Heed, which is a framework for Situation Aware Monitoring and I think also some other things. So thank you very much. >> Conrad Albrecht-Buehler: Thank you. Thank you, Jonathan. Thanks. Thank you all for coming, I appreciate it. What I want to do is I want to tell you four user stories about monitoring and some of the difficulties that these people -- that each of these people have. I'll discuss the situations that are important to each of them and how they each perform their various monitoring tasks. Each story is an example of a monitoring task that competes with many other responsibilities for these people. And is one for which alarms and dashboards don't always work. So the first story I'm going to tell you about is about monitoring servers, something that I think to some degree we all probably have a little bit of familiarity with. The second story is about monitoring business performance by reading financial reports. The third story I'll tell you about is about monitoring the user's experience in a discussion forum. And the last story that I'll tell you about is about monitoring for infectious disease outbreak at an animal shelter. And I'm going to try to make the case that alarms and dashboards are essentially incomplete. I'm going to base that on a distinction that I'll make that between monitoring activities that are analytical and those that are evaluative. In other words, those that try to reach an objective conclusion, compared to those that try to reach a subjective conclusion. So let me start out by telling you a little bit about a CTO that I work with. He's a CTO of a small entrepreneurship and the company's business is offering services online. So it's a very technical company and he's in charge of everything technical there. He's the lead developer and so he's created the website and most of the application that runs on it. He also has executive duties which involve raising funds and doing interviews, but he's also the systems administrator for the entire company. That means that he has to build and maintain all the computers, including all of the servers. So one of the things that he monitors is the performance of the web, the database and the mail servers. So to do that he has these powerful server dashboards that give him a huge amount of information. Now he ends up rarely using them at all. In part it's because it's a lot of work for him to find what is relevant to him in these. I mean, they can be enormous. And they end up requiring quite a significant amount of some mental effort to spot and assemble all the data that is relevant to a situation that he wants to be aware of. One of the other things that he has at his disposal are alarms and alerts to let him know that an incident has occurred. But he ends up filtering all of these alert messages that come in and put them into their own mailbox, which he basically never looks at. He really only looks at them if there's been an incident and he wants to sort of try to trace back to try to figure out where it's origin was. So if he doesn't really use his alarms and he doesn't really use the dashboards, how does he even know that something's ever wrong with his server? Well, he's got two ways. For one, sometimes he'll get someone from his customer support team will come by and tell him that they're getting angry phone calls about people who can't access the site or can't do something on the site. In some ways he let's his customers do his monitoring for him. But throughout the day he's also working with the servers. He ends up, you know, applying patches, updating application code and so on. And whenever he happens to be logged in into the service he'll run up-time just to check the server's load. In essence what he's doing is he's sampling the server load. If he happens to catch it at a moment where the load is high and if he has a little time, he might go and investigate the situation and try to understand what is happening. So in other words, he's essentially asking is the server load high enough that he needs to go investigate it. But high enough...what is high enough? That's a pretty subjective description. What might be high enough for me is probably not high enough for him. Let's take a look at a quick graph of server load. So is the load high? Was it high? Is it now? Will it be soon? A lot of questions about how the server load is doing. Why doesn't he just set an alarm? Well, for one we already know he doesn't really use them. There's another reason. He wouldn't necessarily have time to drop whatever he's doing in order to go and investigate why the server load is high. And when he does have time, an alarm doesn't necessarily tell him anything. It would just be silent if it's not high. So the problem is that this is what an alarm does. It divides the data range. It divides it into the data that he should ignore and the data he should attend to. Now remember that he'd investigate if the load was high enough. Alarm just tells him that it's high. And what's more, it tells him that once it's happened, nothing at all before it happens. The graph would tell him that a little bit more. He could see how it's changing. He could see where it is, where it's been, but it's at a cost. He has to interpret the graph. He has to spend time looking at it and trying to understand it and then he's already dismissed using his dashboard for pretty much the same reason. And it also takes up a lot of screen real estate and as we saw with the dashboards that can balloon really quickly. So instead what if we change how the data range is divided up from this to this? Instead of showing the whole graph, we could just show the region prior to the alarm threshold. Then the position in the region could identify the importance of attending to that situation. If the data is below the "ignore" threshold it might as well be at the "ignore" threshold. If it's above the "attend" threshold, might as well be at the "attend" threshold and somewhere in between it tells them how important it is that he might go look in on the situation. So I prefer sort of this range in between as Heed and it enables us to indicate this sort of degree of importance. And since our CTO wants to know when the load is high enough at that time, we don't necessarily need to show him the whole graph, we could just show him its current position in that range. And so now we could offer him a simple dashboard that relies on the way he already monitors his servers by looking at load. And one that doesn't give him too much detail and still will help him decide if he should turn his attention to those servers. And it does so continuously, so when he has time he can quickly identify if at that time it's worth investigating. And by making this dashboard tiny, easily and persistently visible, he can evaluate the situation quickly and develop some awareness of how his servers are loaded, how frequently they get loaded and so on. So what would end up happening when giving him this little dashboard that we let him use for a few weeks? So he ended up learning quite a bit about how his servers typically perform and that ended up increasing his confidence in their performance and their operation. And he explained to me how it ended up being used to actually catch an incident that could have been a problem, but it wasn't his to catch. He had -- since this thing was running in his office and his other developers go in and out of his office and meet with him from time to time, eventually enough had asked questions what this thing was and he explained it to them. I didn't get to teach them anything about it, but he did. And he told me about one incident where a -- one of his developer his walked by his office and peered in. And he wasn't there, but the developer looked at the display and saw that the database server had a high Heed and it suddenly reminded him that oh, he had left a database backup running and it was the middle of the workday so he ran back and quickly stopped the database backup so that the customers had more access to the site. So even -- to some degree that ended up showing that this was a relatively simple to interpret display and something that even his developer could use and then he explained -- the CTO explained to me how much he had wished he had this display several weeks before. They had gone through a major upgrade and they had replaced the entire web application. And despite having done months of testing once it had been launched there were these instances where the site would just grind to a halt and they would look in and they would see huge amounts -- tons of processes running on the database server. The web server was getting bogged down, but they didn't understand why. And it wasn't consistent, it would just happen sometimes. And this was such a problem that he ended up having to dedicate himself and three of his developers full time for three days watching constantly what was happening on these servers in order to catch the moment when this problem began so that they could try to understand and interpret it. And he had said how much he wished he had a display like this to give out to everybody in the company so that everyone could be watching for a time when one of the developers needed to go look in and see what was causing this problem when the load was increasing on all these things. Ethan? >> Question: Are you going to talk about longitudinal data? Like one of the neat things about the chart that you had earlier is if you're willing to roll back and look at the logs, you can see that historic -- or you are watching dashboard, you can see that historical trending. >> Conrad Albrecht-Buehler: Uh-huh. >> Question: It feels like here I can say, oh, the database load is really, really high ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: -- so I don't know that it's only very recently become high or it's been high for the last three hours because I went and took a long lunch after asking the database to back up? >> Conrad Albrecht-Buehler: Right. That's actually -- that's a great question. I will actually get to that in just a moment. There's -- essentially what you're asking is a person monitoring may actually care more about how something is changing than where it is. And so we'll get to that in just a moment. But it's also -- the way in which you actually asked the question is actually I think very telling. It's -you're doing an analysis of what's going on and a lot of people who monitor things, especially very technical things, will tend to do that. We want to understand what is going on. The only thing is that we don't have a way to identify when we should do that. And that's what I'm trying to cover primarily. So let me tell you a little bit about the next user. She -- so deciding to investigate the situation with the servers with the CTO is a tactical decision and it's one he may have to do several times a day. But what about problems and situations that you need to keep track of that happen far less frequently? So the CEO, she's the executive of a medium-sized business that also offers online services and she's launched several programs to grow the business. And she needs to keep an eye on how those programs are performing. However, she doesn't need to look in on them every day, every hour, she just -- once a week or so is enough for her to get a sense of how they're performing. But like our CTO. She has way too many other responsibilities and constantly sort of dealing with much more urgent and immediate problems. As a result she forgets to checkup on these programs and in fact told me that one time that she hadn't looked in on them for a month and a half, only to find that they were underperforming tremendously. Furthermore, these programs carry a lot of anxiety for her. She's consider worried about their performance and actually as a result that only discourages her from checking up on it, a little bit of avoidance. So the way in which she monitors these programs can she reads these reports that are generated for her and if the report shows that a program is not performing well enough or based on her expectations, she might go investigate and try to understand why something isn't -- why users aren't signing up or why they're not signing up faster. In many ways this is a lot like our CTO's problem, except that the values don't change as frequently. They change a little bit every day. They change a little more every week. But as a result since they're close enough, maybe we can apply similar solution and just show her how close each of her clients are to being underperformers. Now we could just show her these four indicators and let her know how each of the clients are doing, but what happens when this grows to 20 clients or 40 clients? We really want her to read the report and spend her time analyzing the report, so we need a way to sort of compact this display so she can just decide whether or not she needs to go read that report. So to do that we can just assign some values to the Heed range and just say, set 0 for the minimum Heed and set 1 to the maximum Heed and make it a scale. And now we can elect the maximum Heed value to represent the whole report. That means that instead of saying client M needs your attention, it says that the report needs her attention. And it also says that unless the whole program is performing well she should probably go and investigate the situation. So this is essentially an administrator operation applied to these four Heed values. Now she has a single representation that summarizes the report and again if we make it persistently visible she can continuously evaluate the program's performance. Also of note, because we're just assigning these values to the extent of the range we're -- we can combine data of any kind. It doesn't matter whether it's numbers of users or CPU load. We can simply combine them because we're just saying how important is it that you pay attention. Furthermore, because we've sort of compacted this display in a sort of predictable way, we enable her to delve into this and into the situation. Meaning that she can get an evaluation report and then she can delve in and get a little more detail about the -- to get sort of a more detailed evaluation report and then she can go and look at the very detailed actual report and raw data, which is essentially the sort of analytical process. So there's an evaluation that triggers the analytical process and then there's this act of analyzing what's going on. So working with the CEO actually had very surprising results. The biggest surprise was -- well, it was very surprising that when I would go to interview her after leaving the application running with her she would -- she would tell me that she had -- she would apologize and say she hadn't used the display she'd been so busy and had all these things going on. But when I questioned her about the performance of the program, when I questioned her even just about the position of the display, she knew exactly where it was. Despite the fact she didn't think she was using it, she ended up using it. And the fact that it was this very almost passive interface was a really surprising result, something that I certainly didn't expect. But the biggest surprise was the fact that it ended up reducing her anxiety over the performance of this program so much. It wasn't that the program was suddenly performing so much better, but she was aware of how it was performing. And she asked me to create many more displays for her, not because these were necessarily things that she needed to monitor, but these were aspects of her business that were creating a huge amount of stress for her. So she was looking at this as a way of decreasing her anxiety, which I thought was a very surprising result. So the third story I want to tell you about is about a community website owner. There are some situations that people monitor not because they're already familiar with them, but because they want to learn about them. This is one such case. Now he runs a small website -- well, medium-sized website now with map, which is primarily a large community-supported website where people can interact and inform each other and it's largely just a huge forum. He's also the primary developer for this site, he created all the software for it. He created the whole site himself, but the site runs on co-located servers, so it's not really his responsibility to go monitor the status and know how they're doing. He's more interested in had knowing what his users' experiences and trying to make sure they have a good one. The thing he's looking out for is if they're having a bad user experience. And he talked to me about two types of negative user experiences he was trying to look out for. Something that restricted access to the site for his users and the discussion on the site becoming what he called unbalanced. So let me start by telling you a little bit about the access restriction problem that he saw. So since he's the sole developer, he's frequently changing the site. He's changing the site code and he told me about this instance where he had made a small typo essentially in the code and a subset of his users couldn't log in. And at a different time where a subset of his users couldn't post to the forum. And so he was trying to monitor for this happening again. The way he found out the first time this happened was he happened to be talking to a friend of his who's also a forum member, who then told him, oh, I haven't been able to access this site for a few days, is something going on? And he had just had no idea. Even though all the data about the numbers of people logging in, the number of people posting, the number of people applying, were all -- was data that he actually in theory had access to because they were part of a sort of Ruby on Rails framework that he used to create the site. The only way he really had of seeing this data and querying it was through SQL queries and he had enough other things to do that wasn't something he'd end up looking into. So what he wanted to know was if he had something else that ended up reducing the ability for people to log in or reducing the ability for people to post or reply. The thing is that there might be a day when there's a lot of people posting, a lot -it's been a day where there was a lot of -- some national event happened which everybody had an opinion so there was a hot topic and everyone was replying or it was a day after a holiday and everyone was talking about something that was important about the holiday, so there were many new threads started. He doesn't -- it's not that he cares that neither people are not posting or replying, it's that -- I'm sorry, it's not he cares that people are neither posting nor replying, not just one or the other, which is essentially an "and" operation of those two. It's not a logical "and" which would sort of try to pick the minimum of the two values, but it's a compensatory measure instead. Because if people -- not posting doesn't mean that they can't post so maybe he should look in on the problem. But if they can't post and they can't reply he definitely should look in on the problem. Now this issue of monitoring for unbalanced posts, unbalanced discussion is a bit unusual. He talked about making sure the user sort of felt comfortable, especially when new users came to the site. And an unbalanced discussion forum he felt was one where people would be reluctant to want to contribute, especially as new users. And he described that as happening when there were a lot of posts, when there was just a ton of activity where it just felt so -- it might feel insulated to somebody who just came to the site. And so it might be too much -- a time when there's just too much activity he may want to go in and moderate the discussion a little bit. And the other thing he wanted to look out for, what he called Dud posts. These are when somebody posts something, but to which nobody replies. In other words, it's essentially kind of like an announcement of some sort. And that's a problem that he felt would maybe make somebody feel like it was a dead forum in which nothing was really going on. So he was looking out for this, but how did he really know that this was the case because he'd, you know, gone in and moderated before? Well, he used his gut to make that decision. He participates on the forum himself. He looks in on the forum F. It feels like there is too much of it, then he'll step in and do something. But it relies on him being a party to the site and being constantly involved. But more importantly he has -- he didn't have any idea what these numbers were. He didn't spend any time with those numbers. So the previous two solutions involved knowing what some of these values were, having familiarity with the data. So how was he going to even find that out? So what we did is we started out with the same sort of solution, but we used a very conservative estimate and just said maybe it's a very huge range. And if he had got high, he would go and investigate. He would say, okay, I'll go see if this discussion is sort of getting out of control. If you find, oh, no, my gut telling me this is fine, he would then go and adjust the value. He would modify the value to reflect what he was -- what his interpretation was of the information. And really what this interaction amounts to is just moving those thresholds. If -- I'm using the graph from the CPU load to illustrate this, but since the data value is fixed, the pulling the slider down is like moving that -- moving the "ignore" boundary up. So as -- I'm going to walk up here and show you this. So the data value is fixed, so pulling this down is like dragging this "ignore" boundary up. Likewise, if he were to move it up and say, you know what, there is something going on here you're not telling me. I'm going to make sure that next time this happens this is an important situation. He drags it up. It's actually just bringing the attend boundary down. So after a while he would do this and it would end up reflecting his experience with the website more and more. But to answer your question, the thing that he really wants to look out for is that it's this problem is building. That it's not just that it is getting close, but that it is building so he can nip it in the bud. So that is exactly where the graph would come in handy, like Ethan had pointed out. But we still want to -- oops. We still want to use the sort of -- try to make it an evaluation, so we'll add the same sort of boundaries and like I said before, once it dips below -- once the data is passed, one of the boundaries might as well be at the boundary. The other thing is that a graph, as we said sort of before, that there's too much detail in. It just needs to make a quick evaluation, so a lot of the detail's extraneous and so we can down sample the data. >> Question: (Inaudible) -- >> Conrad Albrecht-Buehler: Oh, it fell down. Sorry, Henry, thanks. So we can down sample they'll data and we end up with a much more simplified graph of what's going on. But we also want to be able to predict how soon he's going to need some attention and so I tried to offer him a sort of forecast of what Heed would be in the near future. I tried to look at some really clever learning systems that would do a good job with long-term data, but wasn't successful so I just used a common filter to at least give them a sort of short-term forecast of the data and change the display to reflect both of that, to show the current value, the history, as well as the forecast, and represent occupancy likelihood in this sort of grid with a gray scale. But this display is still -- takes quite a bit of screen real estate and its biggest problem is that the current value is a bit -- not included, but it isn't really obvious. So we want to know he still needs to know what the current situation is. So I created this little animation sort of shows the relationship between this graph view and the sort of slider view we were looking at before. And shows, highlights the current state. It shows how the data and changing and how it's likely to change in the future. It's intended to sort of convey motion, sort of similar to that cartoon, that method of drawing comic books where a ball is flying through the air and your little motion lines behind it, this is intended to sort of evoke that, as well. And users consistently interpreted the direction in which something was changing pretty readily. And so now this also, the display sort of adds another step to that delving subject. The subject can -- the user can see what the situation is, what the evaluation situation is, and then expand the display, sort of get a sense of how it's changing before then digging in deeper. This is some actual data from the access interference test that we did with the user in which we were looking for posts, but what is sort of interesting here is that I've all shown the 10 boundary at the top and the ignore boundary at the bottom, but in this case he care when is people aren't posting. That was the issue of user interference and so it's actually important when that number gets low, not when it gets high. But the -- obviously it's trivial to invert the actual display and make sure that Heed is always oriented with high Heed at the top for sort of a consistency to maybe better see how this historical data relates to the actual data. If we flip the graph it makes it a little bit easier to see how it's down sampled to this. But this inversion was a little bit confusing for this user and for a few others. So we added a little trick to sort of -- a little interface element to remind them what was high Heed and what was low Heed. I hate the red-green mapping, but it was what the user really remembered and could interpret. So we just used these little reminders, but after a while, he, too, disabled them. So our form owner ended up learning a lot -- learning quite a bit about his users' behavior with this. He ended up finding -- he would make these adjustments and then actually go in and look at the configuration file and see where these values were and actually find out how much people are posting and it gave him a much better sense of what was actually going on at the site. So much so that he ended up building his own version of the display that I had nothing to do with. He just simply rewrote the configuration file. He instrumented his website even more and tried to use it to learn about all sorts of other aspects of his -- of the website and how it was used, which was a pretty sort of exciting result. So the last story I want to tell you about is about the Chief veterinarian at a very large animal shelter. So there's some situations people wish they could monitor for, but they're just too complicated and this is one of them. The chief veterinarian also had a great many responsibilities. He's not just a veterinarian that has to care for the animals, but he's also an administrator and he has a lot of sort of budgeting duties and personnel duties. He also has to -ends up testifying in court cases in which he's had to talk about animal abuse, so he has an enormous amount of work to do. And this is a very large animal shelter. The problem that he's particularly concerned about is infectious disease outbreak at these -- at the shelter. The sort of infectious disease outbreak is a problem, not just because he cares about the well being and welfare of the animals, but it's very expensive when there's a disease outbreak. So he wants to try to identify the fact that an outbreak is about to happen and try to nip that in the bud, as well. So the reason outbreak happens is because that the shelter the animals are kept in close quarters, which unfortunately can spread the disease, but is necessary because they want to try to rescue as many animals as possible. They also, as a result, have very limited space for isolation. Any room, any kennel that is reserved for isolation is another kennel in which they can't rescue an animal. But, the animals get sick from time to time. They'll catch -- they'll come in with an infectious disease to the rescue, but it doesn't mean an outbreak is going to happen. It just means that an animals sick. An outbreak is when it starts spreading and there's rapid infection between many animals, which is something that's definitely happened to them in the past. If the disease seems to be spreading, there is some mitigating actions that he can do. He can call for thorough sanitizing of a room which the animals are kept and which they call nuking a room. But before he calls for that, which is both very expensive, it's very time consuming for the volunteers of which they never have enough anyway. He wants to be certain that all the animals that are in that room are not carrying the disease, either because they've been infected and are now immune or haven't gotten the infection yet. So he wants to be certain of that, but it's a really difficult problem for him to even determine that, which has been part because the -- many of thee infectious diseases have symptom-free gestation periods. So some of these gestation periods can be a week and a half long and the animal will be contagious to other animals, but nobody -- but nobody knows they've got the illness yet. Even if somebody notices that there are some symptoms, the animals are necessarily immediately brought for care and there's so many volunteers that information doesn't really get around very easily. Furthermore, even if an animals cared for or diagnosed by one of the vets, there's so many vets that all those people don't necessarily communicate that kind of information with one another. So the way in which he would know is he'd have to actually walk by the isolation room all the time, recognize there's an animal there and recognize it is a new animal being there. So the other way in which he could know is he could go through the records that they have for veterinary care, but there's a lot of them and he does not have time to go through the new records every single day. And lastly the problem gets made each more complicated because the animals get moved around. They are in one kennel one day and they get moved to a different one on another day. The volunteers have to sort of make these decisions based on some animals are not getting along with one another or one seems to be particularly uncomfortable in one kennel. So they get moved around. So just the act of actually investigating this problem is really difficult and very time consuming. So he needs to know when it's valuable to even go investigate this problem where he's going to end up dedicated quite a bit of time just to figure out if there is an outbreak about to happen or beginning to happen. So essentially the importance of him going investigating this is based on the number of incidents that have occurred in each room. So the display we made for him tells him if the disease is spreading through the shelter by showing him how many rooms are showing infection, are showing likelihood of an infection in there. If an animals diagnosed with the illness every room it's been in for about the last 10 days needs to increase it's likelihood for being infected, but adjusted for how many days since it's been there. To determine all this we need to know which animals have been diagnosed and then where they all have been. So luckily that kind of information is stored in those veterinarian records and the animals sort of private records for the -- at the facility and those are stored in an online sort of medical system. But it's a proprietary system and so the data isn't really readily available, but we can parse it by hand and luckily one of the volunteers was willing to do that for us. So if each incident sort of increasing the likelihood of the room being infected after several days that likelihood diminishes because now either all the animals have gotten sick or they're likely already immune. So if the likelihood gets high enough in any room then it's worth his time to investigate that situation. And the more rooms that get high and the more likely it is that he may have an outbreak on his hands and he may have to make time to investigate and make some is very important decisions. So what's interesting actually about this is that he didn't read the display as time for me to go investigate. He knew, he recognized that, but he also said doesn't that also tell me the danger that any of these rooms are at infecting new animals going in? If I could just put this information outside of every door then all the volunteers would know and when they would actually maybe take preventative measures, they'd be more care envelope that room. They might decide there's enough danger in that room that they won't put a new animal in there, they'll put them in a different room. And I thought that was interesting that he ended up mapping this sort of Heed to something else, in this case to danger. And he saw it as a way of actually not just increasing his own awareness, but increasing the awareness of many other people involved in the site at the animal shelter. So from all these stories and from all this work, this all leads me to describe what I think is needed in order to monitor a SWAG effectively. First the observer or the person who is going to do the monitoring must identify and describe the situation of interest to them in terms of some set of indicative conditions. Next they need a way to quickly perceive that situation as an evaluation in some way. And then they need a way to investigate the situation. Delve into it and analyze more. And finally they need a way to refine and modify that -- their understanding of the situation there or define and modify the situation and base, to reflect their changing understanding or the changing situation. I've sort of tried to make the case that with each of these stories that alarms and dashboards and reports aren't sufficient, that they're analytic tools and in the case of all four situations the users valid benefited from or do benefit from some kind of additional tool to support making evaluation. The evaluation is necessary because all these people have many responsibilities. They must constantly make decisions about where to direct their attention. And the closer a swag is developing, the more important it becomes to pay attention to it. And so tools that aid in making those evaluations quickly will help make monitoring easier. Thank you. Thanks all for listening. Thanks for tuning in. Thanks. >> Question: I was going to cover this earlier, but how is this distinct from a dashboard? >> Conrad Albrecht-Buehler: So for one thing that it simply becomes this thing, you can create, you can compact the display on and on to make a single indicator. Let me show you, I think you may have missed the slide in the beginning about -- no, no, no, please. So in this case we can, this was a report in which an executive was looking at the performance of several clients and you could just show this is a giant dashboard or we can actually create a process by which we consistently refine -- summarize them all into a single display that becomes an individual indicator. And the sort of framework -- actually I didn't go into a lot of the detail of the things I showed the "or" and an "and" operation, but with many other logical operators that allow you to just describe much more sort of detailed descriptions. Shouldn't say many more operators, two more operators, one is not. It is very important for a user to be able to describe what's typical and then say, I care when it's not typical, which is just a matter of essentially inverting it, taking one minus the value. And then they need sort of censor operators, not s-e-n, c-e-n-s-o-r operators to be able to say, I care when this happens, except when this other thing is happening. So they need -- so I have something called "and not" which isn't exactly an "and", but it works by essentially suppressing the highest Heed value and it takes the minimum of these values and allows the user to define the situation in terms of mitigating circumstances. That's how it's different than a dashboard, I don't know if that helps, but it also requires less mental effort to parse. So ->> Question: The server example, the massive dashboard versus the indicator, also shows the ->> Conrad Albrecht-Buehler: The dashboard shows raw data and this is intended to show an evaluation of the data. >> Question: What you could do on a thing called a dashboard ->> Conrad Albrecht-Buehler: Right. >> Question: They didn't -- you largely created that for all these people as they have this new dashboard thingy that they can look at that is a lot more evaluative than purely analytical thing they used to have to provide. >> Conrad Albrecht-Buehler: Absolutely. And, you know, the most effective dashboards are the ones that sort of highlight what's important in them that draw your attention to the stuff that's happening in that -- that's being shown in the dashboard that might be relevant to you. But in order to do that the dashboard has to have some sense of relevance ahead of time and so the people who develop the dashboard can do a great job of capturing what is going to be relevant to the users of the dashboard by studying those users. But this allows the users themselves to define it. >> Question: My question unless it is outside this particular range this, particular area, never care about (inaudible) unless it's too high. >> Conrad Albrecht-Buehler: Yeah. >> Question: So you bring all this, note, there is trouble. >> Question: Also, following up on that point and it's kind of interesting, can you comment on how important the cropping actually is? Humans are actually relatively good, in fact, amazingly good at looking for ranges and interpreting, you know, given that you've already compressed it to single dimension ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: -- what if you just list it from 0 to 100 and you didn't crop the -- I think you were calling it a Heed region. >> Conrad Albrecht-Buehler: Yeah. You mean compacting it sort of just down to that one line and then just leaving that whole ->> Question: So you do two compact ones as far as I can tell, right? >> Conrad Albrecht-Buehler: Right. >> Question: One is you compress time and you get rid of time in many of these cases. And then the second thing you do is you crop the "ignore" and the top and the bottom. >> Conrad Albrecht-Buehler: Yeah. Exactly. >> Question: And I'm particularly interested in how much you gain by cropping the top and bottom off as opposed to just letting everything go really high or really low because that does provide you with more information. >> Conrad Albrecht-Buehler: It does, absolutely. But when the information that it is preparing you -- I'll walk over here. The information it's providing you is essentially analytical in nature. You're already, if you are being told that it's gotten super high you're apt to start trying to interpret it, trying to understand it. This is actually just trying to let you know that you need to go find out if it's super high or not. That the distinction is about how -- when this sort of activity is taking place. Is the activity first taking place in your -- deciding this is how you're going to spend your time or you are going to go to that important meeting. Then the activity is about analyzing the problem, understanding it, understanding how high it is, is that a serious problem, all that. That involves quite a bit of sort of cognitive effort. You want to be able to sort of distinguish those two and that's where these -- what I'm sort of advocating for is that we need something that helps us make the decision to go do that to start interpreting it. >> Question: So maybe the follow-up question to that then is for each of those domains was there really a threshold at which I would go dip into it or -- because most of the scenarios I would envision would say it's always a trade-off in time whether I go attend to something and there's a continuum of, this thing is like shooting through the roof, I'm going to drop everything I'm doing right now and go attend to it or if it's just above the threshold I can probably wait five or 10 minutes. >> Conrad Albrecht-Buehler: What I would say is that if you've said that it's just above the threshold then that threshold isn't properly set, then it might still need some adjustment. You're basically saying it's not important enough for me to drop everything to do it, that means to me that it's actually below sort of maximum Heed and if approximate you're saying you care when it's really shooting through the roof, then I would actually say that it's not so much about making the decision based on the actual graph of the data, but actually making the decision and setting the thresholds in the first derivative of the data. So basically creating a new censor from other censor values. So the thing is that people actually are really good at -- in my experience actually describing what they want to pay attention to. I didn't show this, but I found that if I asked a person both of these questions, inevitably they could answer one of them. They could just tell me what are you looking at, what are you looking for. By answering what are you looking at, it's basically allows me to sort of help them define the situation that's important to them in a bottom-up fashion. Starting with the censor they are looking at and sort of adding mitigating circumstances to it, mitigating conditions to it. And if they could tell me what they were looking for, they were describing the situation like unbalanced forum. What does that mean? And then they would start describing the conditions that led to that, the conditions that were meaningful to recognize it was an unbalanced forum and so then we could build that situation description in a top-down fashion. >> Question: And kind of this reminds me so much of (inaudible) small multiples and this example the hospital chart that he did where there are many, many censors that he presented on normalized -- normalized each one then ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: -- and actually did display a time series instead of selecting a time series, we would have (inaudible) ->> Conrad Albrecht-Buehler: Well, I have to admit that my recall for Tufty's work in small multiples, I remember the picture with all the t-shirts on it and not the one you're talking about, so I'm afraid I'd have to actually go back and look it up a little bit. The importance of the time series is that in that case it may have been that it allowed -- I'm only speculating here because I don't remember, but it may have allowed the nurse or the doctor to make the interpretation based on how much things were changing. That again this was about including first derivative into the sort of decision-making process. And by -- in this case you can actually further compact it by basically saying that first derivative is just another censor in this case and we can add it to that. But the thing is it's not necessary to compact it. It's actually fine to show it just in a full dashboard with lots of these and then look -- letting a user look over them or even sorting them based on how high the Heed is. So here's a miscellaneous one, many situations, and subject can look over them and just identify, you know what, this one is not only high, but it is increasing really rapidly, I should go look in on that one. It's not necessary to always compact them and maybe just as valuable to sort of set up a sort of mini-dashboard or even a larger dashboard. I don't know if that answers your question sufficiently. Patrick? >> Question: I was wondering how you deal with changing -- in the cases where you have maximum threshold it's obvious, but in this example you showed the "or" searches -- >> Conrad Albrecht-Buehler: Uh-huh. >> Question: -- you were sort of mapping 0 to 1 to the maximum value, but the maximum could be changing from one day to the next. >> Conrad Albrecht-Buehler: Uh, yeah, absolutely. So essentially what you're asking is that those maximums are changing because the situation is changing, the personal understanding of the situation is changing ->> Question: If I understood your picture right, I thought the maximum was like how many users have locked in or are registered in the system or something like that. >> Conrad Albrecht-Buehler: Uh-huh. >> Question: There could be a day where there's 100 new users and your maximum is higher than the previous day. Right? >> Conrad Albrecht-Buehler: Absolutely. So the situation is actually changed. The world has been somewhat changed. So that is actually another problem that we have with alarm systems that in fact there is a great study by some people in the U.K. in Bambory(phonetic), these colleagues. And what they found is that alarm systems fail in a lot of these large industrial plants because they're set for these values that are not actually appropriate to that factory or to that situation as the factory that was moving through more and more product. And so they actually advocate it's very important that you have a system by which people can adjust their alarms. So what you're talking about is maybe on one day it was perfectly fine that, you know, 50 users logging in is plenty and that's not a problem, but the next day there was a newspaper announcement, tons of people joined the site and now he's going to care when les than 100 people log in. And so the thing is the adjusting of the slide allows them to actually adjust those so to continually try to match what his experience is with the system. It says essentially that that's a never-ending process. >> Question: (Inaudible) ->> Conrad Albrecht-Buehler: Well, to some degree, but that would be -- to some degree the whole problem with alarms is they're telling you, go do this as opposed to saying, go look into this. So maybe it would be more about identifying when they need to go look in on maybe adjusting the sliders as opposed to saying, drop everything, go fix them. But some problems don't change just hour to hour. They change day-by-day. They take quite a bit more time to sort of change and the understanding of these it takes longer to change, but absolutely. There is this sort of problem that it's a never-ending adjustment process. But it just happens in their mind already, it's just is there a system that ends up enabling them to reflect that and sort of work as a cognitive artifact that helps remind them. >> Question: So you did talk explicitly about this, but I didn't see any visualizations around it. >> Conrad Albrecht-Buehler: Uh-huh. >> Question: But it seems that at least some of these scenarios you have sort of this bimodal thing where I care about if my log-ins drop below this number ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: And I care if they go above this other number. >> Conrad Albrecht-Buehler: Uh-huh. >> Question: So there's almost an "ignore" band in the middle and pair of Heed regions and then a pair of alert regions. >> Conrad Albrecht-Buehler: Exactly. >> Question: Did you explicitly describe any of that? >> Conrad Albrecht-Buehler: I didn't explicitly describe that, but actually I can definitely tell you about it. So that's actually just an "or" combination of those two, the high and the low saying, you know, I care when things are really high or when things are really low. So just may go an "or" combination, the to and (inaudible) like the higher ones. So when the data's gotten way up high then it will -- so basically it means that in between the two ignore boundaries it will be well for both and as it moves through one of the two attend boundaries it will become higher. >> Question: So when you say it will be low for both, you mean it will be just basically right down on the bottom? >> Conrad Albrecht-Buehler: Yep. Exactly. >> Question: Yeah, that's a good example to get across the difference between the evaluation and analysis. You don't get anything from -- you get very little just from that one data point ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: -- from the evaluation or for the analysis. >> Conrad Albrecht-Buehler: Sort of the big takeaway I wanted to really push is that people have to make -- constantly have to make these tactical decisions about how they're going to use their time, where they're going to put their attention. And we need tools that help guide that, but don't force it. So like we said with alarms that it's not about telling you go do this. It's about saying maybe go look at this. Frank. >> Question: I've already mentioned sort of (inaudible), but one (inaudible) -which direction are moving away? >> Conrad Albrecht-Buehler: Uh-huh. >> Question: For you this is one obviously (inaudible) direction on this and show us where is the (inaudible) very fast or very slow and which direction they are moving, just kind of ->> Conrad Albrecht-Buehler: No, that's actually a fantastic point and actually I purposely avoided adding any other color mappings because I wanted to sort of leave that open for some future work. I think there's actually an opportunity exactly like you said to sort of add additional information about how fast things are moving based -- using a color mapping to maybe identify things as sort of hotter and colder. But -- and which direction they're going. But there's an intention here that they're always oriented the same way so that if you sort of read left to right and up and down, you can see things that are sort of moving this way, which give you a sense of how they're moving. But the question of how fast they're moving is actually a very important one. And something that I've avoided a little bit here. So I didn't tell you how -- what the time distance is between here. There is a dimension of time, nothing sort of said in between here. And some cases that is something hour by hour, day-by-day, minute by minute, second by second, it's not there. Now the actual display ended up actually having some labels here that indicated how much time there was in between them. But the thing is that two different Heed displays next to one another could have different time scales. So one that seems to be growing really rapidly might actually be just growing over the course of a week as opposed to one right next to it that is just changing second by second or may not seem to be growing very quickly. And making that distinction is very important, but it tends to map to what I found is that users end up having a real problem with that. They ended up knowing that the problems they were paying attention to had a time scale, they knew that these things change either very quickly or very slowly and so they -- that didn't bother them. But it's still actually a very important point and SML that will have to be addressed more sort of in future work looking at this and trying to understand how to compare things that really do happen at different time scales and compare them together. >> Question: So to follow-up of that. If you feel that treads are important, then the main thing you're benefiting from in order to get a representation is space compression. >> Conrad Albrecht-Buehler: That's right. >> Question: Could I get you to speculate on perhaps behavioral differences you might have expected. So I don't know if you did, sorry, but (inaudible) person because it looked like in a couple of these cases that they are very rich, but they went from having no representation in the data to your representation of the data. >> Conrad Albrecht-Buehler: Yeah. >> Question: Could you speculate if they got, you know, the traditional -- the timeline representation of the data ->> Conrad Albrecht-Buehler: Yeah. >> Question: -- would you expect behavior would have been different, too, in reacting to the information? >> Conrad Albrecht-Buehler: I -- I can't -- well, I'd like to try to speculate, but I'm certainly not sure TOCHLT some degree it matters if the person is monitoring something very technical and it also matters essentially sort of what their role with regard to the information is. So for example, the server, the CTO is also the systems administrator. He will respond to the graphs very well. Much of his work involves looking at this. He's involve involved. Looking at this, he's very involved in the very detailed knitty-gritty. But the CEO who works at a much higher level, she wouldn't necessarily benefit from the graphs and in fact from my experience of working with her, very few of her reports have graphs. They have some numbers she looks at and she'll sometimes copy them out into Excel and maybe do a couple of comparisons, but not very much. So in her case I don't think that she -necessarily the graph will have -- a graph or more traditional sort of display would be just as meaningful. But I don't know for example with the chief veterinarian, whether he would benefit more from having sort of graphs showing how frequently there were infections and how often that got spread and so on because I don't know -- I assume he has quite a bit of technical training and must use them similarly, but I don't have any way to know. >> Question: It is a very nice compact representation. >> Conrad Albrecht-Buehler: I think it's actually -- what I think is kind of neat you can end up actually mapping this representation to other things. So for example, just sort of off the top of my head, if you imagine a cell phone dashboard that when you bring it up it shows you a bunch of different information, but it shows you it all relatively consistently. What if the battery indicator got bigger as the battery got lower? But I may be a person who just drives between home and work and I've got a charger everywhere, I don't really care until it gets down to 5% or 10% so I might want to sort of shrink it, but other people who travel a lot more, who have to get on planes, they care more about being able to have that battery. And if it is getting low or getting down to 50%, then maybe it should be bigger and sort of though them that more, give them the ability to sort of make those adjustments or changing. Just like having, you know, you have an appointment coming up, maybe that should be much bigger if it's sooner. Yeah? >> Question: Situations where you think that this would not be appropriate? >> Conrad Albrecht-Buehler: Absolutely. Absolutely. I think it's not appropriate for data that changes incredibly rapidly and has to be monitored extremely carefully. I mean, offhand I don't have any particular example to give you. I'd like to -- yeah. Well, actually the stock market is an interesting example. That comes up often. People say, could you use this for the stock market and I'm not sure because I don't -- I'm not an investor and I don't understand how often investors have to look in on a portfolio. I know that you can set stop orders and stuff like that, or stop loss orders and stuff. But it seems like it is an extremely analytical process and so the relationship of investors to that data can be very -- so analytical that maybe this is not meaningful. >> Question: I think there are some opportunities here. I don't study that very closely, but some of my friends do and I've seen them, I think there are Heed -there definitely is Heed areas based on past performance, but if it's involving a certain range over the last five years then somebody enters into this, so I think there is some potential there. >> Question: I wanted to follow-up. Do you think the displayer should increase the situation awareness or in fact decrease situation awareness? >> Conrad Albrecht-Buehler: So what I saw when -- I saw is it seemed there was an increase. And the way in which I knew that was first of all to some degree define situational awareness. I'm sorry, the title sort of pumps me and we have to do that, but I skipped it. Situation awareness is this field of study that's been primarily promoted by Mike Enzoy(phonetic) and situation where she describes sort of these three levels of awareness that the first level of situational awareness is just sort of knowing the numbers, knowing the data. The second is sort of being able to understand sort of what it means and the third is sort of when you can predict and you can see what sort of coming down the pike, the awareness that something's gonna happen soon. So there was definitely -- there seemed to be in some cases there was increase in the first level of situation awareness, like with the form owner that he didn't have any idea what the numbers were, but he did know sort of what his -- he had a level two sort of situational awareness about the -- what was happening on the form because of his interaction, that was sort of his gut. He was actually using it to increase his level one situational awareness. The user who, the CEO and the CTO both actually increased their sort of level two situational awareness. They understood better about how these things were doing. And then the vet actually was using it in some ways to actually look forward. In some ways he had increased his -- his level three situational awareness. The biggest problem with situational awareness at this point is that it's still really hard to study and to know for sure that awareness has been increased. And so you've got to try to sort of look for indicators, but so that is why this is somewhat -- isn't quite so numerically studied. Did that answer your question? >> Question: Yes. Sort of. I was just thinking like in certain critical situations, you know, a person's ability to react to a changing situation is really key and important. >> Conrad Albrecht-Buehler: Uh-huh. >> Question: And study like auto pilot situations, it's very dangerous because pilots actually have to fly somewhere and it's about the status of the plane. >> Conrad Albrecht-Buehler: Yeah. >> Question: So when an alarm goes off, like, gosh, what is going on. >> Conrad Albrecht-Buehler: Yeah. >> Question: They don't really know. >> Conrad Albrecht-Buehler: Yeah. Exactly. >> Question: I imagine the same thing happens with this, once you hit a certain threshold, unless you really understand the mappings underneath it, you don't actually know what's happening in the situation. >> Conrad Albrecht-Buehler: So one of the things that's going on here is that it's actually, I'm not saying get rid of dashboards, get rid of alarms. Use this with them so there still is a need to be able to say, you know what, there's -- pull up the plane, which I always think is so interesting about alarms and airplane cockpits. They don't just go off, you know, especially the collision avoidance stuff actually gives you a command, which I think is really interesting, as opposed to like a smoke alarm. So now I lost my train of thought. Oh, I lost it. I'm sorry. I was answering -- oh, yeah; right. Sorry. Something getting serious enough that you've got to respond. So there's still a need for alarm. You still need an alarm, but hopefully this actually allows you to maybe interact with the alarm system that this ends up helping define those alarm thresholds and it will let you know when something is going to happen or that something is getting close to happening so that you've got some idea beforehand. >> Question: (Inaudible) derivative seems like a crucial aspect of that. >> Conrad Albrecht-Buehler: Uh-huh. >> Question: Or a likely ->> Conrad Albrecht-Buehler: Absolutely. >> Question: Can you expose the underlying algorithm that computes the (inaudible)? >> Conrad Albrecht-Buehler: That computes the ->> Question: Yeah. >> Conrad Albrecht-Buehler: Yeah. It's -- it's going ->> Question: Like number of posts divided by replies or if it's for ->> Conrad Albrecht-Buehler: Oh, it's not even -- well, no, it's just setting thresholds, it's just setting the line and then doing a transformation between. This is just what function you use to transform in between isn't defined. It's up to the user. So for example the gas gauge, it's not as important -- it certainly doesn't decrease linearly that a change between half a tank and one tenth below half a tank is not the same as the change between, you know, the last 10%. So you might need a sort of exponential function instead. But this might look familiar to other people, but to some of you, because this is essentially the same definition as a fuzzy set. So in some ways this is applying fuzzy set theory to interaction. And these -- this transformation function in fuzzy set there is called a hedge. And on the same sort of hedge stuff applies. >> Question: (Inaudible) -- this one (inaudible) ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: (Inaudible) ->> Conrad Albrecht-Buehler: Uh-huh. >> Question: (Inaudible) ->> Conrad Albrecht-Buehler: Yeah, absolutely. It's a wonderful way to sort of classify sort of lots of different things. You can use this sort of as a matter of classification in that way. >> Question: (Inaudible) -- function. >> Conrad Albrecht-Buehler: Uh-huh. (Laughter) Yeah. Well, it really could be anything. One of the things -- one of the things this function can be is it can actually be based on something very abstract, like the words that people use. So that if in a -- you can just make a giant mapping of different words that sort of require more attention than others, so if people are just talking about cameras and gadgets or whatever, it may be low, but if people start talking about bombs or something, maybe it's more important that you go pay attention to that so that you can apply -- you can end up using just sort of any function. But it's -- it's an evaluation. >> Jonathan Grudin: Okay. If there's no more questions I'll thank Conrad again and anybody who wants we're going off to lunch over next door if anybody wants to join us it's fine to join us. (applause) >> Conrad Albrecht-Buehler: Thank you. Thank you. Mapping of different words that sort of require more attention than others, so if people are just talking about cameras and gadgets or whatever, it may be low, but if people start talking about bombs or something, maybe it's more important that you go pay attention to that so that you can apply -- you can end up using just sort of any function. But it's -- it's an evaluation. >> Jonathan Grudin: Okay. If there's no more questions I'll thank Conrad again and anybody who wants we're going off to lunch over next door if anybody wants to join us it's fine to join us. (applause) >> Conrad Albrecht-Buehler: Thank you. Thank you.