Tagsplanations: Explaining Recommendations Using Tags

Related comments:

Summary:

In this paper, Jesse Vig, Shilad Sen, and John Riedl from University of Minnesota's Grouplens Research present Tagsplanations, a system of explaining system recommendations based on community tags.  Popular media services like Netflix or iTunes often have recommender systems that suggest similar items based on the user's previous selections or perceived "tastes."  The researchers first discuss the most common method of recommendation - establishing an "intermediary entity" between the user and the item to be recommended.  This is the typical model of an intermediary entity:

 Tagsplanations uses tags to establish this intermediary.  For example, “We recommend the movie Fargo because it is tagged with quirky and you have enjoyed other movies tagged with quirky.”  They aim to provide the user more information about why a user might like a recommendation.  It uses several features like tag preference, relevance, and filtering to increase the overall quality of the tags.  

The researchers conducted a user study in which participants tested Tagsplanations in four different interfaces and filled out a survey about their experience.  80% of the subjects found that the system helped them understand recommendations better and make a selection decision.

Discussion:

From my experience with Netflix, I know that tagging and recommendation systems can help a user discover new media that they otherwise might not have come across.  I think Tagsplanations is a great way to bolster current recommendation systems and help users make decisions about what they want to see.  The specific tags are a lot better than a vague "you might also like this" statement.

Detecting and Correcting User Activity Switches: Algorithms and Interfaces

Related comments:

Summary:
In this paper from Oregon State University’s School of EECS, authors Jianqing Shen, et al. present an updated version of an interface they had previously designed to detect and catalog switches in user activity, called TaskTracer.  TaskTracer applies machine learning methods to associate sets of resources with a particular user activity and make those resources more readily available to the user.  In this context, the term resources include documents, folders, email messages, contacts, and web pages.  TaskTracer configures the desktop in several ways to make these resources easy to get to: the task explorer presents a unified view of all resources associated with the current activity; the folder predictor modifies the Windows Open/Save dialogs by defaulting to folders associated with the current activity and adding shortcuts to them; the system has a time reporting feature that allows the user to show how much time was spent on each activity in a given period; finally, TaskTracer automatically tags incoming and outgoing emails associated with the activity.  The first version of TaskTracer had several problems including incorrect associations, unnecessary interruption of users with dialog boxes, and very slow learning algorithms.  TaskTracer2 fixes these issues by implementing an improved association engine, a desktop state estimator, a more intelligent switch detector, a notification controller that minimizes user interruption cost, and a clearer two-panel UI.

The researchers conducted a study on two people, a “power user” who recorded 4 months of data, and a second user who used the system for 6 days.  Overall, the participants found TaskTracer2 to rarely make an incorrect prediction, though it didn’t always make the exact correct one.

Discussion:
I think the user study for this application speaks the most clearly in regards to how I feel about TaskTracer: the “power user” found it very useful and accurate and was even described by the researchers as “fairly careful about declaring switches.”  I think associating resources with particular tasks is a great idea (like a more powerful, dynamic version of Windows’ “Recent Documents” feature), but having to explicitly declare when I’m switching activities would annoy me very quickly, especially if the system started pestering me about it with dialog boxes.  I think TaskTracer is best suited to the sort of conscientious power users that the researchers studied; the average user probably wouldn’t reap enough benefit for all the explicit task-switch declaration to be worth it.

Annotating Gigapixel Images

Related comments:


Summary:
Qing Luan, et al. present their developments in the annotation of very large (billions of pixels or gigapixel) images in this paper.  Their aim was to augment the pan-and-zoom interface used by applications like Google Earth and Virtual Earth with visual and audio annotations, driven by an intelligent rendering system that takes into account the viewer's perceptual distance from the objects being annotated.  They mentioned related work in areas like zoomable UIs, map labeling, human psychophysics, and augmented reality systems.  The system developed by the researchers runs on HD View and consists of text labels, audio loops, or narrative audio.

The system gauges viewer perspective, depth, and field of view relative to each annotation; it then assigns strengths to the various annotations based on these elements.  Thus, farther-off text labels will be smaller or farther-off audio annotations will be played at lower volume.


Discussion:
I haven't played with Google Earth very much, but according to the authors this system is a lot like it, so I think I have a pretty good idea of how the interface works.  I think it's a great idea to render annotations dynamically based on viewer position -- if all available annotations were rendered simultaneously on a map of the United States it would be completely unreadable.  Hiding or showing labels as the user pans and zooms encourages the user to explore the image to see what annotations can be uncovered.  This would be great if applied to an educational setting where children could browse a map of, say, Philadelphia and read excerpts from historical documents or hear audio clips about important locations.

Inky: A Sloppy Command Line for the Web with Rich Visual Feedback

Related comments:


Summary:
In this paper, Robert C. Miller, et al. (researchers at MIT CSAIL and the University of Southampton) present Inky (short for Internet Keywords), which is basically a command line interface that provides shortcuts to common web browser tasks.  It functions as a hybrid between a command line and GUI interface, giving dynamic visual feedback to the user as they type and utilizing sloppy syntax which frees the user from learning any new command syntax.  Keywords can be provided in any order, replaced with synonyms, or entered in a variety of ways; the Inky interpreter attempts to match the user's input against available commands.  Inky is a Firefox extension built using HTML, CSS, JavaScript, Java, and XML.  The various construction, command, and content-highlighting specifics are discussed in the rest of the paper, covering details about the keyword interpreter, functions and usage.

The authors conducted a small user study of seven participants and found that 95 out of 131 user commands were correctly interpreted, and that Inky was fairly easily learnable based on user response.  Most of the commands that were not correctly interpreted were simply attempts by users to invoke commands that did not exist (by entering website names, for example).  The researchers used their findings about Inky to begin development on a second prototype called Pinky (Personal Information Keywords) which focuses on lightweight data capture for personal information management.


Discussion:
My first thought about this application was that it seemed an awful lot like Quicksilver for Mac -- an application launcher/wizard's tool that interprets user input to execute commands and launch applications.  This paper even mentions Quicksilver, in fact, when talking about how Inky is invoked.  It also shares many similarities with Mozilla's own Ubiquity.  Like, a lot of similarities.  As in, these tools are nearly identical in function and aim.  Having used Quicksilver and Ubiquity pretty extensively, I think a command-line-style interface for quick command execution is a great idea that can really streamline a user's experience (if they are willing to take on a small learning curve).

Understanding the Intent Behind Mobile Information Needs

Related comments:


Summary:
In this paper, Karen Church and Barry Smyth (of Telefonica Research and University College Dublin respectively) lay out their study of user's mobile content needs and mobile contexts.  The explosion of mobile device popularity across the world (3.5 billion subscribers in 2007) has led to new patterns of information retrieval and new needs for users.  The authors analyze previous work dealing with what mobile users search for and why; the research indicated that 50% of the top mobile search queries were related to adult content (though only 8% of all users in the study engaged in any kind of search at all) and that intent for search could usually be attributed to a need for awareness or status-checking behavior.


The authors conducted a four-week study of 20 participants who were asked to keep a diary of all their information needs while they were at home, work, or on the move.  Participants logged the date and time, location, and information need, as well as any additional comments they had.  The study generated 405 diary entries (approx. 20.3 per person), the majority of which concerned on-the-go conditions for mobile information need.  The researchers created three categories for user intent: informational (how-to's, advice, showtimes), geographical (location-based or directions), and PIM or personal information management (PIN codes, friend requests, to-do lists).  30% of entries were geographical in nature, and 42% of entries were non-informational.  PIM entries represented 11% of the entries.  


Overall, the study indicated that mobile users look for considerably different information than standard Web users, seeming to have a greater need for geographical-based information such as directions or service locations while they were on-the-move; user information needs also seemed to revolve heavily around social interaction (friend requests, questions asked in conversation, status updates etc).


Discussion:
The research presented in this paper is very relevant to current trends in technology - users are gravitating more and more to mobile content, and the advent of user-friendly mobile browsing devices like the iPhone and the upcoming iPad will only heighten this trend.  I found the categorizations made by the researchers to be pretty accurate - in the brief period that I owned a web-enabled phone, I used it primarily for getting directions, checking Facebook or Twitter, and managing email.  This falls right in line with the results of the study.  However, in the future I think user content needs will shift more and more toward entertainment in a variety of contexts and locations as more content is pushed to the mobile space and bandwidth increases.