I was browsing old posts of MobBlog the other day, and ran into Daniele’s post on (double) blind review. I have to say I’ve yet to see a conference in my field which genuinely benefited from double blind review, and despite the anecdotes I’ve heard of double blind being statistically better for female authors than regular review, I’m starting to think we possibly need to move for the exact opposite soon. (more…)
Archive for the ‘reputation’ Category
Reputation systems on mobile phones build “reputation scores” based on personal experiences and on other phones’ recommendations. One poorly-explored research question is when to use personal experiences and when to resort to recommendations. This paper in Biology Letter may suggest few future research directions:
A lot of interesting papers have been discussed during RecSys conference. Here are a few list of interesting ones.
I’ve started reading this book by Daniel Solove (full text available for free). It discusses the growth of the Internet – and begins with a “horror” story depicting the potential the Internet has as a defamatory broadcasting medium on people: the lines between freedom of speech and privacy are once again put to question (like in similar stories in the news).
Has anyone else read the book?
The IFIP trust management conference, this year joined for the second time with the Privacy, Security and Trust (PST) conference, was held in June 18th – 20th in Trondheim, Norway. The conference has also been previously known as iTrust. Next year, PST and IFIPTM split again, as PST returns to its roots as a local event in Canada; next year, IFIPTM is organized in the US, and in Japan after that. We’ve summarized IFIPTM workshops on Monday and Tuesday in earlier posts, and now give a quick run-through of what this year’s conference program held.
The Monday workshop sessions of IFIPTM 2008 were a combination of the second workshop on Context-awareness and trust (CAT) and first workshop on Web 2.0 trust (W2Trust). See the W2Trust website for the full list of papers. In this post, we summarize what we saw.
Yesterday was the first International Workshop on Trust in Mobile Environments (TIME 2008), co-located with the IFIPTM 08 conference in Trondheim, Norway. The workshop merged with the Workshop on Sustaining Privacy in Autonomous Collaborative Environments (SPACE), and consisted of three sessions. Here is a brief summary on what we saw: (more…)
First International Conference on Reputation
What: Work on Reputation from a multidisciplinary standpoint
Where: Tuscany, Italy
When: March 2009
The CINCO (Collaborative and Interoperable Computing) research group at University of Helsinki aims to automate some of the routine tasks in inter-enterprise collaboration management. The vision is that one day, enterprises can trust an automatic system to 1) figure out which service provider to use for a task, e.g. a logistics service to deliver a set of goods, 2) ensure that the collaborating services are interoperable, and 3) gather and share experience on how the collaboration went. And all this should be achievable without first spending a few years to get to know (and integrate your systems with) every single service provider whose offers you might wish to choose between.
Experience sharing for this kind of a system has special needs. A major difference to e.g. eBay and the various recommender systems for consumers is that the information should be possible to both understand and evaluate for credibility automatically. While the average concerned web user can google around for hoaxes, or browse through the profiles and activities of the users behind eBay ratings until convinced, our automatic decision-maker has to have an explicit model of “suspicious” or “sensible” for reputation information in order to determine the credibility of the information available. When a decision to commit real-world resources is made automatically, we’ll need to be able to measure the certainty behind the reasoning.
A few of the interesting research questions we’re working with are how to represent the different factors of trust for these decisions and to combine them into a decision, how to model the shared experiences or reputation, how to evaluate the credibility of information and its sources, and how to make different reputation systems interoperate. See the group’s selected reading for more information and three surveys.
I’m glad to say that the TRECK track of SAC went quite well and did not suffer from some of the things I mentioned in my previous rant. The track was organized by Dr. Jean-Marc Seigneur of the University of Geneva, and the two sessions were chaired by Dr. Virgilio Almeida of the Federal University of Minas Gerais (who I had an interesting discussion with after the track), and was broadly divided into two themes: trust and recommender systems. The trust session had an overall focus on peer-to-peer systems, here are some quick samples:
- Francesco Santini presented the idea of multitrust, which aims at computing trust in a dynamically created group of trustees who all have different subjective trust values ["Propagating Multitrust Within Trust Networks, " Bistarelli/Santini].
- Asmaa Adnane presented the application of trust to detecting misbehaviour in link-state routing algorithms. I always wonder how well these cool ideas will work in practice; if information is lost or delayed they will deduce that another node is untrustworthy! ["Autonomic Trust Reasoning Enables Misbehavior Detection in OLSR," Adnane/Timoteo de Sousa/Bidan/Me']
- The Surework Framework extended the current operation of trust in p2p networks to include the idea of super-peers; nodes with very high reputation can, in fact, become reputation servers. ["Surework: A Super-peer Reputation Framework for p2p Networks," Rodriguez-Perez/Esparza/Munoz]
- The CAT Model was introduced and explained- it is a model of open and dynamic systems that considers services as contexts.. The 15 minute time-limit was a bit constraining and I’ll have to read the full paper! ["CAT: A Context-Aware Trust Model for Open and Dynamic Systems" Uddin/Zulkernine/Ahamed]
- Rowan Martin-Hughes applied a game-theoretic analysis to understand why people would defect in a large-scale open system, like eBay. The analysis was based on a modified version of the Prisoner’s dilemma, which was very interesting; the only question that arises is, as Daniele mentioned, is this appropriate when users may very well behave irrationally? ["Examining the Motivations of Defection in Large-Scale Open Systems," Martin-Hughes/Renz]
The second session focused on recommender systems:
- Karen Tso-Sutter presented her work on combining user-item tags into the collaborative filtering process. Interestingly, tags did not improve accuracy until the algorithm was already boosted by using both user- and item- based algorithms. ["Tag-Aware Recommender Systems by Fusion of Collaborative Filtering Algorithms," Tso-Sutter/Marinho/Schmidt-Thieme]
- My work! Looking at the similarity distribution over a graph generated by a nearest-neighbour algorithm. ["The Effect of Correlation Coefficients on Communities of Recommenders," Lathia/Hailes/Capra].
- Patricia Victor‘s paper discussed an extension to Paolo Massa’s work on trust-aware recommender systems, which concluded that the cold-start problem in recommender systems can be avoided by having users express trust values in other users, which can then be propagated. The problem is: which users should they connect to? The paper has an interesting analysis of the different kind of users in the epinions dataset. ["Whom Should I Trust? The Impact of Key Figures on Cold-Start Recommendations," Victor/Cornelis/Teredesai/De Cock].
- The last paper veered away from collaborative filtering to look at the role of keywords and taxonomies in content-based recommender systems. The taxonomy vs. folksonomy war continues! ["Comparing Keywords and Taxonomies in the Representation of Users Profiles in a Content-Based Recommender System" Loh/Lorenzi/Simoes/Wives/Oliveira]
The full list of abstracts can be read on the trustcomp-treck web site. If any of the attendees or authors are reading this post: we welcome your thoughts and comments, and officially invite you to contribute to this blog! To write a guest-post about your research, please get in touch! (n.lathia @ cs.ucl.ac.uk)
It relies on this fact
One can group receivers by the amount of spam they receive on a daily basis. Say we consider 5 groups. “People in Group 1 receive, on average, 90% spam. Group 2 receives 70% spam, Group 3 receives 50% spam, Group 4 receives 30% spam, and Group 5 receives 10% spam.”
How it works
Messages are classified whether they are spam or not depending on the receiver of the message, “rather than where the message is FROM or what it CONTAINS“. “Essentially, if the message is sent to users who typically receive a high percentage of spam, the message is more likely to be spam. However, if the message is sent to users who typically receive a low percentage of spam, the message is more likely to be legitimate. Combining the reputations of all recipients of a particular message, therefore, is equivalent to combining those users’ rating power to estimate the legitimacy of the sender and the message”
What about new users?
“The system can be bootstrapped from an empty database with just 2 users (someone who gets a lot of spam and someone who gets a lot of ham). … The system was initially seeded with just two users: a person who receives virtually all spam and a person who receives virtually all legitimate mail. The statistics of a third user was then approximated using the ratings established by the first two users. The fourth user was added with that user’s statistics approximated by the first three users, etc.”
“The amazing thing is no human is required to read or rate any email; the system gets smarter on it’s own without any human intervention”
In Italy, hackers have introduced erroneous messages into the traffic signal sent to GPS devices (article). This exemplifies the need of security for vehicular networks. Part of the needed security mechanims may be offered by reputation (trust) models as two recent papers show: