QLectives: evolving software to support quality Nigel Gilbert and the QLectives team This work was partly supported by the Future and Emerging Technologies Programme (FP7-COSI-ICT) of the European Commission through the project QLectives (grant no.:231200). 1 Friday, June 8, 2012
QLectives COMMON INTERESTS POOLED RESOURCES QUALITY CONTENT • The vision A European project to connect people with common interests and • The theory deliver quality content to them • The practice • The future Social P2P Peer networking infrastructure production Smart quality recommendations Quality Collectives QLectives 8 Jun 2012 www.qlectives.eu 2 Friday, June 8, 2012
The vision www.qlectives.eu 3 8 Jun 2012 Friday, June 8, 2012
The basic idea • If we give people the right tools they will self-organise into communities that support their needs • A common need is to find high quality content for entertainment or professional activities • We believe communities can be grown around the creation, distribution and recommendation of quality content in given domains • We aim to implement tools that are fully distributed, requiring no central control or authority 8 Jun 2012 www.qlectives.eu 4 Friday, June 8, 2012
Partners University of Surrey, UK • – social modelling Technical University of Delft, Netherlands • – P2P design and deployment engineering ETH Zurich (Swiss Federal Institute of Technology), Switzerland • – social modeling with Physics connotation University of Szeged, Hungary • – P2P and distributed systems algorithm design University of Fribourg, Switzerland • – social modelling with an EconoPhysics approach University of Warsaw, Poland • – social complexity modelling with a Psychological approach The Centre National de la Recherche Scientifique, France • – network analysis Institut für Rundfunktechnik GmbH in Munich, Germany • – metadata, exploitation 2009 - 2013 • 8 Jun 2012 www.qlectives.eu 5 Friday, June 8, 2012
People 8 Jun 2012 www.qlectives.eu 6 Friday, June 8, 2012
Project structure www.qlectives.eu 7 8 Jun 2012 Friday, June 8, 2012
The theory www.qlectives.eu 8 8 Jun 2012 Friday, June 8, 2012
www.qlectives.eu 9 8 Jun 2012 Friday, June 8, 2012
Quality, trust and reputation www.qlectives.eu 10 8 Jun 2012 Friday, June 8, 2012
quality ¡is ¡assessed ¡ through ¡the ¡opinions ¡of ¡ trusted ¡others ¡and ¡is ¡ the ¡basis ¡of ¡one’s ¡ reputation ¡among ¡ peers ¡and ¡the ¡wider ¡ community. 8 Jun 2012 www.qlectives.eu 11 Friday, June 8, 2012
Quality is social • Quality is not an inherent property of an object – so not like its colour, for instance • Quality is ‘constructed’ through interaction with others – so there is a process of achieving consensus about the quality of an object (or failing to do so) – the same object may have di fg erent quality assessments simultaneously • di fg erent evaluating groups • di fg erent metrics • di fg erent objectives 8 Jun 2012 www.qlectives.eu 12 Friday, June 8, 2012
Multiple quality assessments • The same object may have di fg erent quality assessments simultaneously – di fg erent evaluating groups • e.g. wine bu fg s, alcoholics – di fg erent metrics • e.g. the ‘nose’, percentage alcohol – di fg erent objectives • a pleasurable experience, to get drunk 8 Jun 2012 www.qlectives.eu 13 Friday, June 8, 2012
Consequences • Judgements of quality serve to validate own and others’ opinions • Evaluation is a way of achieving belonging and a ffj liation with others • The ‘others’ are those with similar characteristics – so evaluation is a group-reinforcing process: •I belong to the group that thinks like me about quality •The group consists of those who think similarly about quality • The evaluations of quality stand as symbols for the group •e.g. persian cat fanciers 8 Jun 2012 www.qlectives.eu 14 Friday, June 8, 2012
Trust • Trust increases with: – number of interactions – number of successful interactions – number of altruistic interactions – higher status / reputation of trusted person – degree of homophily • e.g. same culture, background, discipline, experience – .... • Trusted others are those whose opinions are valued – No trust for those not known 8 Jun 2012 www.qlectives.eu 15 Friday, June 8, 2012
Reputation • Ascribed by others • Assessed on the basis of – quality of ego’s work – trust in ego – activity • Context dependent 8 Jun 2012 www.qlectives.eu 16 Friday, June 8, 2012
How they inter-relate • Ego’s reputation is a function of – the aggregation of the quality scores of ego’s objects – the aggregation of the trust placed in ego by others • An object’s quality score is a function of – the aggregation of the quality ratings of everyone who has rated it – adjusted by •ego’s degree of trust in the rater •age of rating • The degree of trust awarded to a person by ego is a function of – the distance from ego measured by e.g. •the number of links; •the number and frequency of interactions – that person’s reputation 8 Jun 2012 www.qlectives.eu 17 Friday, June 8, 2012
Help, please! Scientists have to understand and evaluate a wide range of information from a variety of different sources. Indeed, in recent years this amount of information has increased rapidly. In this survey, we are interested in better understanding how those working in science make decisions about information relating to authors and publications. By taking part in this survey you are helping us to build a better knowledge of how scientists rely upon various forms of information. This can help the wider scientific community, for example, by developing literature search tools which produce more useful results. The survey can be accessed via the following URL and should take no longer than 5 minutes: http://bit.ly/KC5Mkt 8 Jun 2012 www.qlectives.eu 18 Friday, June 8, 2012
Algorithms www.qlectives.eu 19 8 Jun 2012 Friday, June 8, 2012
www.qlectives.eu 20 8 Jun 2012 Friday, June 8, 2012
Gossip learning framework • Given a network of autonomous nodes – possibly many millions of them, with fragile links • How can one implement – reputation mechanisms – spam filtering – recommender systems – distribution of global data through the network • While preserving – scalability – privacy 8 Jun 2012 www.qlectives.eu 21 Friday, June 8, 2012
Gossip learning classification • These examples reduce ultimately to solving a classification problem, by developing a ‘model’ that, given a case, x , classifies it as a y – model f() such that f(x i ) ≈ y i with minimum error • Most methods for finding f() require access to the complete database – but we need an algorithm in which each node has a very partial database and each node can work in parallel with the others 8 Jun 2012 www.qlectives.eu 22 Friday, June 8, 2012
Gossip learning Ormándi, R., Heged ű s, I. and Jelasity, M. (2012), Gossip learning with linear models on ‘merge’ can be as simple as averaging the two models fully distributed data. Concurrency Computat.: Pract. Exper.. doi: 10.1002/cpe.2858 www.qlectives.eu 23 8 Jun 2012 Friday, June 8, 2012
The practice www.qlectives.eu 24 8 Jun 2012 Friday, June 8, 2012
QMedia www.qlectives.eu 25 8 Jun 2012 Friday, June 8, 2012
www.qlectives.eu 26 8 Jun 2012 Friday, June 8, 2012
www.qlectives.eu 27 8 Jun 2012 Friday, June 8, 2012
8 Jun 2012 www.qlectives.eu 28 Friday, June 8, 2012
www.qlectives.eu 29 8 Jun 2012 Friday, June 8, 2012
www.qlectives.eu 30 8 Jun 2012 Friday, June 8, 2012
www.qlectives.eu 31 8 Jun 2012 Friday, June 8, 2012
QMedia • Completely decentralised – No central web site like piratebay.org • Quality assessed through self-organising personalised collectives – called ‘channels’ • Users can – publish and consume media files – create and modify metadata – post comments 8 Jun 2012 www.qlectives.eu 32 Friday, June 8, 2012
Challenges Peer discovery • – puncture NATs (network allocation table, i.e. a router) – random walk algorithm Free riding bandwidth • Anti-spam measures • – user comments Security • – Dispersy distributed permission system Scalability • – works with 10,000s nodes Extensibility • – user widgets Performance • – carefully tuned Portability • – written in Python, with MySQL database at each node Availability • – open source 8 Jun 2012 www.qlectives.eu 33 Friday, June 8, 2012
www.qlectives.eu 34 8 Jun 2012 Friday, June 8, 2012
Number of page downloads by month from Tribler.org 200000 150000 100000 50000 Jan 2011 Apr 2011 Jul 2011 Oct 2011 0 Jan 2012 www.qlectives.eu 35 8 Jun 2012 Friday, June 8, 2012
QScience With thanks to Stephano Balietti for many of the slides in this section www.qlectives.eu 36 8 Jun 2012 Friday, June 8, 2012
Stefano Balietti sbalietti@ethz.ch 30.03.2012 www.qlectives.eu www.qlectives.eu 37 8 Jun 2012 Friday, June 8, 2012
Recommend
More recommend
