Tagsplanations: Explaining Recommendations Using Tags

Jesse Vig, Shilad Sen, John Riedl, University of Minnesota

Summary

Authors explore the design space of explaining recommendations to user.

Details

One of the research areas explored in improving transparency, trust and user satisfaction in recommender systems is explaining them based on item, user or a certain feature. Authors explore the relevance (relation between item and tag) and preference (relation between user’s sentiments and tag) of tags using explanations called tagsplanations.

For their experiment, authors tested four different interfaces based on different combinations of relevance and preference. Interface that showed relevance and preference while sorting using relevance scored highest in this study pointing that users preferred relevance but did not trust the system enough to view it alone. Both these attributes seemed to be of equal effectiveness.  Subjective tags seemed to perform better than factual tags but if depicting the same idea, users preferred factual tags. For future work, authors propose studying trust and scrutability of such a system.

Review

Authors have presented a strong design space exploration of recommendation explanations and make some insightful observations. In the words of author themselves, it would be desirable to have an empirical measurement of how well the explanations worked instead of using self-reporting.

Disclaimer

The work discussed above is an original work presented at IUI 2009 by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.

Searching Large Indexes on Tiny Devices

Guy Shani et al., Microsoft Research

Summary

This paper proposes a novel technique for searching large indexes on devices with limited UI (5 keys and one line display). It employs a distribution based OBST (Optimal BST) extended to an OTST (Optimal Ternary Search Tree) using pinning. With shrinking mobile media devices, UI options become increasingly limited and a technique to allow user entry faster with lesser keystrokes is desirable.

Details

OBST relying on probability distribution of search strings would be faster than lexicographic BSTs. Authors extend this structure to an OTST by adding pinning. So while up and down keys traverse OTST in conventional sense to left and right subtree, and left key retraces route taken, right key traverses OTST to middle subtree which is a subtree with an extra character ‘pinned’ in the prefix string shared amongst all the child nodes. Authors argue that with prefixes often shared amongst artist and album names adding this new subtree enhances the search experience. Authors present quantitative comparison showing OTST to be the best technique amongst various other techniques. ROTST (Restricted OTST) represents a design choice based on user feedback where the middle subtree retains previous root as one of the options (even though users can simply end the search by pressing ‘enter’).

User study shows lowest keystroke count and search times for ternary searches although spelling bases searches had lowest error rate because of being a well familiar concept. This tone was also reflected in user comments where though they preferred ternary searches, they recommended spelling based searches for an average user. Ternary searches register an improvement over binary searches in average time for single keystroke as users need to focus only on the working prefix string.

Review

Perhaps this paper represents the classic struggle between the ‘best thing to do’ and ‘simplest thing to do’. While OTST and ROTST ranked highest in the quantitative study, user preference for average user lied squarely with spelling based linear search. In the words of author, a key question that remains unanswered is what an average user thinks about this technique. Authors admit that their results might be skewed by their sample space which consisted entirely of Microsoft employees. Having said that, the technique is very innovative and consolidates existing search techniques in a neat manner.

Disclaimer

The work discussed above is an original work presented at IUI 2009 by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.

Evaluation of Expert Recommender Systems

Tim Reichling & Volker Wulf, University of Siegen

Summary

Expert recommender systems (ERS) hold a great potential and this paper does an evaluation of some such systems.

Details

Knowledge management (KM) has moved from repository based approaches to social networking based strategies. ERS is a major application for KM systems. Tradeoffs exist between automating profile builder and yet respecting privacy. Existing ERS use text matching algorithms, consider structured data can match different sources of personal data. Authors had developed an ERS called ExpertFinding (EF) for a European national industrial association for which they have performed this evaluation.

EF uses two mechanisms for profile creation. First one creates a keyword list out of documents provided by users for this purpose. These documents are present on user’s local or a shared file system and could include dynamic data such as public email folders to get a tap on their day to day conversations. Another mechanism creates yellow page (YP) style form that is maintained by user themselves in case they wish to shape their expertise profile. A local, client end UI was made to allow people to search for experts by entering keywords.

Based on their evaluation, authors found that while ERS accurately found keywords indicating user’s domain of knowledge, it was inadequate in representing the nature as well as level of their expertise. Various incidents also cropped up of people bloating or expanding their expertise profile by means of YP. Authors present a 4 stage diagram where they envisaged their tool in balancing out workloads and representation of employees.

Review

From the paper it seems authors did a very poor job in building this tool and I describe why below:

• user control over both profile generation mechanism meant that users could bloat their profiles which is exactly what happened.
• choice of testers for tool evaluation was done poorly with all testers knowing each other. This led to a very specific usage pattern where A searched only for people A knew. Sample size of 23 testers also seems to be small for a study of such nature.
• no socialising of profiles was available, an important extension could have been providing ability for people to comment on other’s profile. This could have added some authenticity to YP profiles.
• the tool is too simplistic in nature. For ex. neither level nor nature of expertise was indicated on one’s profile.

Having said above, there were some good things in their study:

• privacy concerns of users were well taken care of.
• plugin based development allowed them to continuously re-deploy their tool easily.
• to their credit, they described their work exactly as they did it without attempting to cover up.

Disclaimer

The work discussed above is an original work presented at CHI 2009 by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.

The Inmates are Running the Asylum I

Alan Cooper

Discussion

In this book, author describes common problems that plague the world of software design. It says programmers belong to a separate species homo logicus instead of homo sapiens. Programmers make really good programs that are meant for other programmers. Their differences in perception with the rest of humanity however can cause problems when the software is not meant for them.

Author points out that programmers are running the show most of the time when it comes to designing how a user interacts with a program. The author shows that this has negative consequences on the finished product because either the programmers are taking shortcuts in UI design or misinterpreting what the users desire.

Disclaimer

The work discussed above is an original work by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.

SmartPlayer: User-Centric Video Fast-Forwarding

Kai-Yin Cheng, et. al., National Taiwan University

Summary

A new video interaction model allows video players to adaptively fast forward through mundane stretched of video while ensuring no areas of interest are missed.

Details

Techniques of still image abstraction and video skimming try to summarise a video but face disadvantages in not being able to present to user finer details of areas they are interested in and ensuring nothing is missed out from the summary. Author’s SmartPlayer performs video skimming but instead of skipping, fast forwards ‘dull patches’. A user study was performed to find following user expectations about playback speed and base the SmartPlayer video flow on:

• constant, when video is ‘interesting’.
• changes gradually only.
• changes based on minutes of viewed footage.

Based on these, SmartPlayer has a motion and semantic layer to achieve video skimming. While motion layer adapts the playback rate, semantic layer detects predefined semantic events in the video. Personalization layer keeps track of user’s video browsing history. Based on user testing, it was found that SmartPlayer was quite useful in long, predictable videos with understandable characteristics and where audio was of secondary importance in nature.

Review

How many times does it happen with you that you have to fast forward through hours long wedding video to go through those ‘special moments’. Chances are a lot many times and this technique promises that those are going to be days in past. The technique proposed in this video sounds quite intuitive and it is only surprising that no one has thought of it before. It almost sounds like MPEG equivalent in terms of time spaced in viewing a video just like an MPEG would save on space used when a video frame does not change a lot over time.

Disclaimer

The work discussed above is an original work presented at CHI 2009 by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.

Performance Model of Selection Techniques for P300-Based Brain-Computer Interfaces (BCI)

Jean-Baptiste Sauvan, et. al., INRIA/IRISA

Summary

A model based on Markov theory has been proposed to predict performance of selection techniques of a P300 based BCI.

Details

Typical BCIs have been based on positive 300ms EEG signals. On being shown a display with one of the objects flashing, user starts to count which results in P300 being detected 300ms later. The interaction technique has been represented as static Markov chains. This allow authors to compute time required to perform an action and corresponding number of flashes needed. Three different techniques were proposed and validated against model:

• Global : where any object can flash alternatively (& hence directly selected).
• N-chotomic: where user selects one of N sub regions before selecting single target within that sub-region.
• Relative: where user ‘move’ his selection from currently targeted object by moving to its neighbours.

 

Review

While I did not understand the mathematics behind Markov theory but it was interesting to read on BCI.

Disclaimer

The work discussed above is an original work presented at CHI 2009 by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.

Autism Online

A. Taylor Newton, et. al., Univ. of Denver, Univ. of Oregon

Summary

This is an interesting comparison study on word usage of bloggers with and without Autism Spectrum Disorders (ASD) to show that ASD might be due to social contexts and vanish in ‘computer-mediated’ communications.

Details

A set of syndromes collectively referred to as ASD is marked by deficits in responding promptly to socio-emotional cues such as smiling back or maintaining eye contact. Authors have tried to find out how people with ASD fare in socially distal contexts in their communication skills. By taking Internet blogging as such an example, authors found 57 blogs from self-proclaimed ASDs and compared their word usage (using LIWC dictionaries) on a five-factor structure used previously in a study on blog linguistics of neuro-typical (NT) bloggers. Based on their results, they found that ASD word usage did not differ by more than 14% of a SD although there was 4 times as much variance in ‘Sociability’ factor.

Review

This study seems to validate what many in autistic community have claimed that Internet provides them a medium to ‘speak for themselves’. While loopholes can be pointed out in the sample size, uniform representation and increased variance in ‘Sociability’ words, to be fair to authors, they do mention performing an invasive research and comparing distal and proximal studies in future.

Disclaimer

The work discussed above is an original work presented at CHI 2009 by the authors/affiliations indicated at the starting of this post. This post in itself was created as part of course requirement of CPSC 436.