Paper Reading #32

Taking advice from intelligent systems: the double-edged sword of explanations

Authors:	Kate Ehrlich	IBM, Cambridge, MA, USA
	Susanna E. Kirk	IBM, Cambridge, MA, USA
	John Patterson	IBM, Cambridge, MA, USA
	Jamie C. Rasmussen	IBM, Cambridge, MA, USA
	Steven I. Ross	IBM, Cambridge, MA, USA
	Daniel M. Gruen	IBM, Cambridge, MA, USA

Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

 

Summary

• Hypothesis - The researchers point out that intelligent systems can both help and hurt users by giving explanations. The explanations were hypothesized to help users when a correct answer is available, but harm when one is not available. Their goal was to determine the extent of this situation.

• Content - Intelligent systems have the capability of giving suggestions and explanations on problems. If the situation is critical, users must be able to trust the system for correct responses and understand that the explanations are correct. However, even if the system is correct, the researchers claim that correct explanations can lead to incorrect results from ambiguity.
• Methods - The researchers designed a prototype machine learning program (NIMBLE) that give recommendations and explanations in regard to network security issues. 19 participants were selected to perform a 2 hour session consisting of 24 timed trials of network security tasks. The task was to rank a set of 11 threats by importance, either low or medium. To help the users, the system gave recommendations in 3 levels: no recommendation, recommendation only, and recommendation along with a justification. The recommendations also could either have only one or no correct choices with the 3 given recommendation conditions. The helpfulness of the system was then rated in a Likert scale.
• Results - Suggestions and justifications were found to significantly help accuracy when a correct choice was available. However, when it is not available, there was an insignificant drop in accuracy. In all, the researchers found that the participants used the suggestions widely, depending on the level of skill they have.  

Discussion
The closest I can relate to this system is back when I used the spelling and grammar check of earlier Microsoft Word products. Back then, the explanations for some grammar corrections were vague, and some didn't even bother giving explanations. Granted, rationally parsing all of incorrect English is not an easy task, but I remember being unsatisfied with some of the results. For future work, these principles should be applied in areas where fault tolerance is necessary.

Paper Reading #31

Identifying emotional states using keystroke dynamics

 

Authors:	Clayton Epp	University of Saskatchewan, Saskatoon, Saskatchewan, Canada
	Michael Lippold	University of Saskatchewan, Saskatoon, Saskatchewan, Canada
	Regan L. Mandryk	University of Saskatchewan, Saskatoon, Saskatchewan, Canada

 

Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

 

Summary

• Hypothesis - Traditionally, emotion detection requires sensors to directly measure physical data of a participant. The researchers believe that their system can reliably detect participant emotions based on the dynamics of how they type.

• Content - This paper avoids other types of machine learning processes and data collection to determine user emotional state.  Here, the only data collected is solely they keystroke dynamics and feedback for the machine learning via questionnaires. Previously, non-fixed typing data collection was unable to accurately determine user emotional state, so fixed-length typing sets were giving for users for data analysis. Free text was too variable according to the researchers to accurately accomplish the researchers goals.
• Methods - Instead of gathering keystroke data in a laboratory setting, the data was collected at home in a daily basis. 26 participants were asked to fill out a questionnaire about how they felt emotionally and then have them write a fixed length piece of text. The data was collected regularly for 4 weeks. The information collected was in 4 categories:

1. Keystroke features - keystroke duration and latency.
2. Content features - which characters are being typed including capital letters and numbers.
3. Emotional state classes - data collected from the questionnaires that included 15 questions about their emotional state.
4. Additional data points - the active process name for each collected keystroke.

Also since people often took pauses during the tests, further computation was performed on the keystroke data. If a pause was longer than the standard deviation of the data, the data point was thrown away, which took out 0.07% of all the data.

• Results - Models were created from the data that detected confidence, hesitance, nervousness, relaxation, sadness, and tired states. Some states were temporary, such as excitement and anger, and were not classified properly within the fixed length text. However, only a few features were necessary to classify emotions accurately.

Discussion
The researchers accomplished their goal adequately but I think their system should have been designed to allow more flexibility, especially over time. To require that a user types a fixed length text is prohibitive to deploying their models to any free-form typing. However, since the content of the words typed are not analyzed itself, this may have simply been impossible. Otherwise, this type of machine learning could be applied to operating systems to act differently around users that give different emotional states through how they type. Also another data point they left out for data analysis is mouse movement. This data could further model generation and possibly make it more accurate.

 

 

 

Paper Reading #30

Life "modes" in social media

Authors:	Fatih Kursat Ozenc	Carnegie Mellon University, Pittsburgh, Pennsylvania, USA
	Shelly D. Farnham	Yahoo!, Sunnyvale, California, USA

Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

 

Summary

• Hypothesis - The researchers suggest that people interact in different ways based on their social contexts of their life. These life facets can include work, school, and home environments. They also believe if that online interactions could be oriented against these facets, then the interactions will be more effective.

• Content -

1. Modes of the self - different life facets afford different "modes" people interact within.
2. Focused sharing - Broadcasting information is generally disliked compared to the idea of focused sharing, which allows information to be shared with only a few people instead of everyone.
3. Mobile Me - A theme that many users explain their online interactions as being based on mobile interfaces.

• Methods - 16 participants were asked to perform a two hour interview to help them understand the facets of their life. There were 3 specific goals:

1. Life mapping activity - Participants were asked to draw out parts of their lives and tell how they handle transitions from different states online, as well as how they communicate.
2. Need validation session - Theoretical situations were presented to the participants and they were asked to identify possible problems and how they would handle them.
3. Visual metaphors ranking activity - Users were asked to rank how the best way to visualize their life facets.

• Results - For the life mapping activity, many of the participants drew a social meme map or a chronological map to visualize their lives. For the social meme maps, most of them start from the idea of "me" and spread outwards to family, work, and social facets. These categories were then further split up. Chronological maps were able to more accurately show transition states but lacked the detail that social meme maps could separate. For the color coding activity of marking communication practices, the closer the participant was to a person, the more likely it was that the participant would mark the possible communication channels in multiple ways. However, others were much more specific on how they communicated, avoiding crossover. For working environments, it was found that most people segmented their work and social life. For example many participants had an email specifically for work, personal, and junk email. Crossing the segmentations of their life facets were usually done by physically moving from one location to another, or by the passage of time.

Discussion
Many people simply ignore their life facets when communicating online. Historically this can have disastrous consequences, such as and employee badmouthing their boss on Facebook after they have become friends. Social networking sites now are implementing functionality that allows emulation of personal life facets in ways more normal for humans, such as Google and Facebook, but use may or may not catch on due to the different environment online interaction provides at the time. I think the researchers bring up a very good idea that is currently changing how people interact online. However, since different interactions with different people in different life facets is a generalization of the concept of privacy, and with more and more technology, there seems to be less privacy available. I do not think that a single mode of interaction is possible with humans, but it would simplify the process considerably. 

Paper Reading #29

Usable gestures for blind people: understanding preference and performance

 

Authors:	Shaun K. Kane	University of Washington, Seattle, Washington, USA
	Jacob O. Wobbrock	University of Washington, Seattle, Washington, USA
	Richard E. Ladner	University of Washington, Seattle, Washington, USA

Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

Summary

• Hypothesis -The researchers believe blind and sighted people use touch interfaces differently. Using the differences in design practices can help different users with better gestures.

• Content - Four design guidelines were set for observation:

1. Avoid symbols in print writing - Blind people may not learn print as a form of input.
2. Favor edges, corners, and other landmarks - spatial orientation without sight can be replaced by relative positions, increasing gesture effectiveness.
3. Limit time-based gestures - Blind people take a longer time to perform gestures, so this could limit what they can do.
4. Reproduce traditional layouts when possible - familiar layouts such as QWERTY let blind people immediately know how to use an interface.

• Methods - 10 sighted and 10 blind people were asked to perform gestures in order to accomplish a task described by a moderator. After making two gestures, they were asked how effective they thought their gestures were on a Likert scale. The study was repeated but the participants were asked to perform the gestures they previously made and rate them again on a Likert scale.
• Results - Because of the lack of visual feedback, blind people found their gestures more fitting than sighted people based on their Likert responses. Also the gestures from blind people were considerably more complex, having more edges and taking a longer time to complete. They were also more likely to use multi-touch compared to sighted people. The gestures from blind people were also physically larger. Overall, blind people did not find their gestures easier, but in different tasks there were more variance. This means that some tasks are easier than others. 

Discussion
The researchers achieved their goal of understanding interaction with blind people reasonably well. Unfortunately the technologies created today have been built up for so long with visual feedback being an essential backbone, it is difficult to tell if blind people will ever have the same utility from technology. In the future, this issue could be solved by giving artificial sight, however that is far down the road.

Paper Reading #28

Experimental analysis of touch-screen gesture designs in mobile environments

Authors:	Andrew Bragdon	Brown University, Providence, Rhode Island, USA
	Eugene Nelson	Brown University, Providence, Rhode Island, USA
	Yang Li	Google Research, Mountain View, California, USA
	Ken Hinckley	Microsoft Research, Redmond, Washington, USA

Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems

 

Summary

• Hypothesis - The researchers believe that bezel initiated gestures are superior to soft buttons, especially in non-ideal environments where focus can not be held on the interface. 
• Content - Bezel gestures start from the edge of touch interfaces. The other types of gestures tested were soft and hard buttons. A soft button is simply a GUI representation of a button where the hard button is an actual mechanical button. The second factor tested was whether or not the gestures were aligned to an axis. Those that were aligned to an axis are mark-based gestures, where the others are termed free-form. 
• Methods - 15 participants were asked to test the effectiveness of the different types of gestures in different environments. Participants were tested when sitting, standing, walking, and introduced 3 levels of distraction to these motor states. The least distracting environment allowed constant visual contact with the interface, where the most distracting environment did not allow the user to look at the screen. At the end of the study, participants took a questionnaire to determine other factors. 
• Results - More time was taken for gestures in more distracting environments, regardless of the gesture type. However, in this environment, it was found that bezel initiated gestures were superior to hard and soft buttons. There was no difference from gestures performed during sitting and standing. The accuracy of free-form gestures was lower than mark-based gestures. When distractions were not an issue, soft buttons performed better. The fastest and most preferred environment was sitting with no distractions and soft buttons.

Discussion
The bezel initiated gesture built into Android to display notifications is one of my favorite things about the user interface of the system. I am able to perform the action in distracting environments relatively easily. The only issue for future use is that since it is used to display the notification menu, there is a degree of dimensionality lost for future applications. However, in distracting environments such as driving, the best thing to do is to remove the distraction entirely, either by pulling over, or not messing with a phone.

Paper Reading #27

Sensing cognitive multitasking for a brain-based adaptive user interface

 

Authors:	Erin Treacy Solovey	Tufts University, Medford, Massachusetts, USA
	Francine Lalooses	Tufts University, Medford, Massachusetts, USA
	Krysta Chauncey	Tufts University, Medford, Massachusetts, USA
	Douglas Weaver	Tufts University, Medford, Massachusetts, USA
	Margarita Parasi	Tufts University, Medford, Massachusetts, USA
	Matthias Scheutz	Tufts University, Medford, Massachusetts, USA
	Angelo Sassaroli	Tufts University, Medford, Massachusetts, USA
	Sergio Fantini	Tufts University, Medford, Massachusetts, USA
	Paul Schermerhorn	Indiana University, Bloomington, Indiana, USA
	Audrey Girouard	Queen's University, Kingston, Ontario, Canada
	Robert J.K. Jacob	Tufts University, Medford, Massachusetts, USA

 

Proceeding  

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

 

 

 Summary

• Hypothesis - The researchers want to prove that fNIRS is a reliable way to measure brain states and to test the amount of stress different types of multitasking has on participants. 
• Content - FNIRS stands for functional near-infrared spectroscopy. It directly measures the oxygenation of the blood in the brain to determine how much stress user's brain is exhibiting. Multitasking is broken up into 3 possible scenarios, each with a different amount of stress on the brain. 

1. Branching - Changing focus to another task requires keeping a state memory of the tasks. 
2.  Dual Task - Changing focus to another task does not require keeping a state memory. 
3.  Delay - The interfering task can be ignored until the first task is done. 

• Methods - The researchers tested the three types of multitasking with different scenarios of problem solving involving directing a robot explorer. Delay multitasking was tested by asking the participant whether two successive rock classifications follow in immediate consecutive order. Getting a message about distance can simply be ignored. Dual Task was handled by duplicating the circumstances of delay, but with focus being on both distance and rock classification. Only an immediate change in the current type of message required attention. Branching was tested by requiring the participant to remember the previous states of each message. 12 participants were selected. Each participant practiced the system before hand to get used to the multitasking environment before using the fNIRS equipment. 
• Results - Delay tasks were the most accurate and done the quickest. The slowest and least accurate was the branching test. The stresses caused by multitasking were easily seen by the fNIRS data, showing the greatest amount of oxygen depletion with branch-type multitasking.

Discussion
I find myself multitasking all the time. Using these principles if I created scenarios where I worked with delay-based multitasking primarily, I could theoretically become 26% more efficient according to the timing data collected. However, I'm fairly certain with the relative newness of computation in society in regards to evolutionary time, there is a very likely chance that a distinct evolutionary pressure is being put on people who can multitask more efficiently. In the mean time, technology such as fNIRS can be used to minimize "harmful" types of multitasking that our brains can not fully handle yet.

Paper Reading #26

Embodiment in brain-computer interaction

 

Authors:	Kenton O'Hara	Microsoft Research, Cambridge, United Kingdom
	Abigail Sellen	Microsoft Research, Cambridge, United Kingdom
	Richard Harper	Microsoft Research, Cambridge, United Kingdom

 

 Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human

 

 

Summary

• Hypothesis - The researchers believed that people would act differently when operating a brain-computer device.

• Content - The MindFlex game board is a system that consists of a hand-actuated circular obstruction game path, and a EEG headset to measure brain activity. The more the user concentrates, the more a fan levitates a styrofoam ball. After this, the user rotates the game board to direct the ball past the obstacles.
• Methods -  The researchers gave 4 different groups of people a MindFlex game and instructed a leader to gather people to play the group and record the reactions during playing. The video footage was later collected and analyzed for patterns.
• Results - Participants were found to change their orientation in an attempt to change their level of concentration. Some people held their breath or came in closer to the ball as well. Other people lowered their concentration by not looking at the game. Another way for changing the level of concentration participants used was to imagine the ball moving up. The action of imagination caused more mental activity and caused the ball to move up.

Discussion
Considering the brain has the most entropy of data of all possible systems on the body for collecting data, I find it depressing that the current level of technology only allows such little interaction in BCI. For the future, it would be amazing to have the level of technology rise to the point where interfaces not just simply tell a difference in activity in localized places of the brain, but to concretely separate out different ideas being processed in the brain in real time. Obviously this wont come in the form of a $70 toy in any point in the near future, but I think it is possible.

Paper Reading #25

Twitinfo: aggregating and visualizing microblogs for event exploration

 

Authors:	Adam Marcus	Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
	Michael S. Bernstein	Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
	Osama Badar	Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
	David R. Karger	Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
	Samuel Madden	Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
	Robert C. Miller	Massachusetts Institute of Technology, Cambridge, Massachusetts, USA

 

Proceeding  

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

 

Summary

• Hypothesis - The Researchers believe that events in twitter can be determined using timelines from the Twitter API and algorithms in a system called TwitInfo.

• Methods/Content - TwitInfo takes a keyword and gathers tweets about the keyword for a period of time. After collection, the tweets are organized by the internal algorithms. From there, peaks are calculated and classified into events. The system can also determine whether the sentiment of individual tweets are positive or negative using machine learning. To test the system, 12 participants were asked to use it to find events on certain subjects.
• Results - Most events were found properly using the system. For soccer games, there were peaks during the individual goals, making it easier to visualize the information. The participants were able to recreate the details of an event usually within 5 minutes.

Discussion
There is an enormous amount of data on Twitter considering so little text is in each individual tweet. What makes Twitter such a good medium for communicating data is the fact that information is shared directly when an event happens. A strange example of this is when a person in Pakistan tweeted about hearing the (stealth) helicopters used to kill Osama bin Laden. Using TwitInfo to analyze this event would be very interesting. I'm imagining this system could be used by politicians to determine how well their constituency agrees with their policies, since the System already calculates positive or negative emotions.

Paper Reading #24

Gesture avatar: a technique for operating mobile user interfaces using gestures

Authors:	Hao Lü	University of Washington, Seattle, Washington, USA
	Yang Li	Google Research, Mountain View, California, USA

Proceeding  

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems

Summary

• Hypothesis - The researchers believed that their system can be used to easily select small objects on a touch screen better than a previous system called Shift.
• Methods - In order to compare the usability of the Gesture Avatar to Shift in a non-biased way, 20 participants were asked to try each using different sized targets, different number of targets, and testing while either walking or sitting. 10 participants used the Gesture Avatar first, while the other 10 used Shift first. 
• Content - The point in this research is to resolve the "fat finger" and occlusion problems with touch gestures. The fat finger problem is the fact that a finger may be significantly larger than the object to be selected. The occlusion problem states that the user can not specifically see what they are selecting because the finger occludes the screen while interacting with it. Gesture Avatar works by drawing a letter to select a link on a page. If the link's first letter is the letter drawn, it is highlighted. To select the link, the user taps on the bounding box of the gesture they created. If they want to select another link that satisfies the letter, the user drags outside the bounding box to go to the next link. To undo the gesture, the user taps outside the bounding box.
• Results - GA was faster at selecting larger items, but slower at smaller items compared to Shift. GA had the upper hand to shift when walking was involved since walking did not effect the error rate for GA, while shift had a much higher error rate. Because of this, the majority of the participants preferred the Gesture Avatar.

Discussion
This is a pretty clever way of solving the fat finger and occlusion problems. However, I believe most of the issues this solves for web browsing have mostly been partially solved by having mobile versions of pages. Also, I'm not convinced that this is faster than simply zooming into the page to make the links larger. For future work, this application could be integrated into android as an option.

Paper Reading #23

User-defined motion gestures for mobile interaction

Authors:	Jaime Ruiz	University of Waterloo, Waterloo, Ontario, Canada
	Yang Li	Google Research, Mountain View, California, USA
	Edward Lank	University of Waterloo, Waterloo, Ontario, Canada

Proceeding 
CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems

Summary

• Hypothesis - The Researchers believe that some types of gestures will be more difficult than others
• Methods - Twenty participants were asked to make gestures on how to do simple tasks with mobile devices. Since learning about new concepts in smart phones may make the users less able to perform the tasks, only people who were previously used to smart phones were tested. The gestures to be made were supposed to be created to be as simple to the user as possible. Since the hardware in the mobile devices are not able to completely recognize all gestures perfectly, the users were told to treat them as "magic bricks" that could understand any gesture. Also the users were asked to describe why they chose such gestures while they were performing them in order to understand their reasoning. After the users created their gestures, they graded their own gestures on a Likert scale.
• Content - The tasks performed were as follows: Answer call, hang-up call, ignore call, voice search, place call, act on selection, home screen, app switch next, app switch previous, next (vertical), previous (vertical), next (horizontal), previous (horizontal), pan left, pan right, pan up, pan down, zoom in, and zoom out. The ways gestures are described are with the nature of the gesture (symbolic or abstract), whether it has a context, whether the gesture causes an action after or during the movement, how much impulse is applied to the gesture, the dimensionality of the gesture, and it's complexity.
• Results - The more direct mapping to real life examples for manipulation the gestures were, the more widely the participants liked the gestures. For example, the gesture for answering a call that most participants agreed upon was a motion that was similar to the motion that would be to answer a call by putting the phone to the ear. Hanging up the phone was done in another similar mapping to old-style phones by turning the screen around and parallel to the ground. A slightly unnatural mapping was found for moving the screen. On touch gestures, people "drag" the objects in the screen with a movement that is in the same direction as the motion of the object. However, for motion gestures, the participants moved the window in the opposite direction of the movement of the objects in the screen.

Discussion
It would be nice if every gesture had a very easy and simple direct mapping to normal movements in the physical world. The problem is that with the added complexity and functions that computer systems can offer, there sometimes is no other way to perform a task other than using a computational device. In this case it is probable that an unnatural gesture would have to be made. Then again, this theoretical situation would make the gesture the only way to perform the action in such a world, and in a sense, would be the "normal" action to do anyways. One can only hope that adequate gestures are mapped to actions when mobile device complexity explodes.

Paper Reading #22

Mid-air pan-and-zoom on wall-sized displays

Authors:
Mathieu Nancel - Univ. Paris-Sud & CNRS; INRIA, Orsay, France
Julie Wagner - INRIA; Univ. Paris-Sud & CNRS, Orsay, France
Emmanuel Pietriga - INRIA; Univ. Paris-Sud & CNRS, Orsay, France
Olivier Chapuis - Univ. Paris-Sud & CNRS; INRIA, Orsay, France
Wendy Mackay - INRIA; Univ. Paris-Sud & CNRS, Orsay, France

Proceeding CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems

Summary
Extremely large data sets are difficult to visualize. For example, the Spitzer's 4.7 billion pixel image of our galaxy can be visualized on multiple displays in different ways, but some are better than others. A user can sit at a desk in front of the wall-sized display and use physical input such as a mouse to manipulate the images, or use mid-air gestures in 1,2, or 3 dimensions.

Hypothesis
H1: Two handed gestures should be faster than one handed gestures.
H2: Two handed gestures should be easier and more accurate.
H3: Linear gestures for zooming should be better for zooming but eventually will be slower than circular gestures due to repositioning.
H4: Users will prefer gestures that do not require repositioning.
H5: Gestures using small muscles will be faster than larger ones.
H6: 1D-path gesures should be the fastest.
H7: 3D-free gestures will be more tiring.

Methods
There are several design factors to be determined. The user can use one or both hands for input. Using one hand allows the other hand to perform other actions, but using two hands can allow panning and zooming to be done at the same time. Gestures can be linear or circular. Linear gestures (moving a slider) are naturally mapped, but may require clutching. Circular gestures do not require clutching, but is not natural. Different degrees of freedom have different advantages as well. 1D path gestures have strong haptic feedback but are limited in dimension. 2D surface has more freedom but has less feedback. 3D free hand gestures have the most freedom, and does not require any device. The only issue is that there is no feedback at all.



Results
Two handed gestures were found to be faster than one hand gestures. One dimensional path gestures were the fastest of all the gestures as well. Even though for circular gestures, the participants did not have to move back to reposition, the linear gestures were faster because of less overshoot. The participants favored the faster gestures because they were the easiest.

Conclusion
There exists 3D stylus devices for 3D drawing programs. I can imagine these devices could be modified to allow better haptic feedback with servos and strain sensors to allow much better feedback. A more interesting task would be to allow for 3D-free manipulation of virtually projected 3D images. There exist several engineering problems to solve this issue however. Back to the display of large data sets, I don't think too many people will benefit directly from being able to see 100 million pixels at home.

Paper Reading #21

Human model evaluation in interactive supervised learning


Authors:

Rebecca Fiebrink Princeton University, Princeton, New Jersey, USA
Perry R. Cook Princeton University, Princeton, New Jersey, USA
Dan Trueman Princeton University, Princeton, New Jersey, USA


Proceeding
CHI '11 Proceedings of the 2011 annual conference on Human


Summary
Machine learning allows for algorithms to determine output from an input based on a trained model for the data. Not every input is specifically represented in the model, so the model needs to make generalizations. Interactive machine learning (IML) uses model evaluation so that users can assess how well a certain model is performing. Without IML, understanding these models can become a difficult debugging process.

Hypothesis /Motivation

The first goal of the paper is to learn more about how interactive machine learning based on user evaluation. The second goal is to determine how these evaluations effect what users do in utilizing IML, specifically what algorithms they use. The last goal is to describe the utility of supervised machine learning.

Methods
A tool called the Wekinator was designed to encompass several standard machine learning algorithms. It also allows for the training, modifying, and visualization of real time data used to create the model. Three studies were performed using the Wekinator.

• Study 'A' involved 6 PhD students and one faculty member at Princeton's Music Composition department. For 10 weeks, they met and discussed how they were using the tools in their work, and offered suggestions for the software. At the end, they completed a questionnaire about how they felt about the software.
• Study 'B' involved 21 undergraduate students. They were asked to design systems of input and train models using neural networks to create different sounds. The inputs needed to be continuously controlled.
• Study 'C' was a case study that involved a cellist for the purpose of building a gesture recognizing cello bow, called the "K-Bow." This had sensors for acceleration, tilt, and relative position. Also grip pressure is measured. The goal was to create a model to process the sensor data to create physically appropriate music.

Results

All 3 study groups used direct evaluation to determine the system is working properly in processing input data. However, in studies 'B' and 'C', the researchers noted that the participants used cross validation todetermine how well the models evaluated new data. Subjective evaluation was used in study C to find errors in the model that were not found using cross or direct evaluation.

Discussion
Machine learning is very interesting to me, primarily since I'm enrolled in CSCE 420. Creating good models is very good, but I personally would like to see a fully automated process for model creation and evaluation. This isn't to say that intelligence depends on having best model of the world, but it is a good start. 

Paper Reading #20

The aligned rank transform for nonparametric factorial analyses using only anova procedures

Authors:

Jacob O. Wobbrock - University of Washington, Seattle, Washington, USA

Leah Findlater - University of Washington, Seattle, Washington, USA

Darren Gergle - Northwestern University, Chicago, Illinois, USA

James J. Higgins - Kansas State University, Manhattan, Kansas, USA

Proceeding
CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems

Summary

ANOVA is short for ANalysis Of VAriance. Nonparametric data in experiments with multiple factors occur in the field of HCI. Analyzing interaction effects with this type of data is impossible with traditional analyses. The Aligned Rand Transform is able to analyze this type of data.

Hypothesis

The researchers propose an algorithm and tools for analyzing multi-factor non-parametric data.

Methods

ARTool and ARTweb are programs that utilize the ART algorithm. The algorithm starts by computing each cell's residual value. Next the estimated n-way effects are computed using alternating positive and negative summations. Then the residual is added to the estimated effects. Last the average ranks are assigned and a full factorial ANOVA is performed on the ranks.

Results

The ARTool was used to evaluate a satisfaction survey for how satisfied they were on a test interface. The data was analyzed with ANOVA and with ART. The ANOVA test did not show any interaction but the ART did. The researchers hypothesized the interaction would exist and the ARTool was able to prove it for them.

Conclusion

I have yet to take statistics, so much of this is quite confusing. I do know that determining correlations in data is important for finding unknown relations. The best example I know about this is the statistics Blogger and YouTube offer (both being from Google) for individual uploads. The paper notes that this algorithm is accessible to anyone "familiar with the F-test," which I am not. After skimming through what it is, it seems there is quite a bit of statistics I need to understand before fully understanding the ART algorithm.

Paper Reading #19

Reflexivity in digital anthropology

Author: Jennifer A. Rode - Drexel University, Philadelphia


Proceeding: CHI'11 Proceedings of the 2011 annual conference on Human factors in computing systems

Summary
There are three types of anthropological writings, consisting of realist, confessional, and impressionistic writing styles. Even though realist writing is most common, all have some good qualities. These types of writing have specific values that can affect the HCI design process.


Hypothesis
This paper’s purpose is to establish and characterize the contributions of anthropological ethnography.

Methods
Realist writing consists of creating a sense of authority from the author because of the fact they have collected data directly in the environment they are studying. This prolonged exposure to the environment gives enough data to determine an unambiguous result. The goal is to minimize the effect of the observer in the situation from the precision of the data.


Confessional writing instead directly uses the opinions of the observer. Since the observer now has a noted effect on how the perceived data is collected, the authority may be diminished, but sometimes allows for demystifying the data collecting process. Since ethnographers tend to avoid adding their own opinions in the process, this type of writing is not as common as realist writing.


Impressionist writing is done similarly to a narrative. Sometimes writers try to startle the audience through dramatic recall instead of dry facts.


One task the ethnographer should try to accomplish is to establish and maintain rapport with the subjects they are studying. This is to say that the subjects have a mutual understanding of how and why the ethnographer is doing a study. This allows access to the data being collected. However, lack of rapport sometimes allows understanding of the differences of cultures from the ethnographer and the subjects.


Participant observation is the core to reflexive studies. It is described as "deep hanging out" and is the key to cultural anthropology. 


If anthropology is purely experimental, it is essential for the opinions of the author to be left out. However, the relationships with the experimenter always causes actions that can not be properly explained without considering the writer's existence. This is the key to a formative design process.

Three different forms of ethnography can be used to HCI practices.

1. Formative ethnographies determine how technologies are currently being used for the purpose of improving the technology.
2. Summative ethnographies details how a certain group uses a technology, but does not help towards the HCI design process since it done after the fact.
3. Iteratively evaluated ethnographies use prototypes in practice and determine how well they are used. This data is used to improve another prototype design.

Results
There was no testing


Conclusion
It makes perfect sense to use ethnographic data for the HCI design process. If a design or product isn't working for some reason, it can either be redesigned to be "better" in the designer's eye, or the customer's mind. The designer can become disconnected from what the end-user wants, but the end-user can't always know what they want since they don't understand the technology in detail.