Considering Collective Intelligence on the Web

A popular post in the 'sphere at the moment regarding Group Think which considers Collective Intelligence (CI) and the web in the context of two specialised conferences relating to the issue (O'Reilly CI foo camp and ETech). The post highlights the current popularity of CI and considers some of the facets of sociology that feed into the field such as crowd theory, human terrain mapping, social simulations, etc. Three examples are reviewed: simulated models of crowd behaviour, models of traffic behaviour with prediction and potential for real-time manipulation, and finally post-hoc models of inter-personal student relationships and spatial-temporal patterns with prediction.

In my field of computational intelligence, Collective Intelligence has a distinct meaning in the context of achieving desirable emergent behaviours (intelligent problem solving) using the aggregate behaviour of lesser intelligent agents. Common examples include swarm intelligence such as particle swarm optimisation and ant colony optimisation. More generally, one may integrate notions of distributed AI, such as the presently vogue multi-agent systems. As such, reading the post mentioned above elicits standard notions of emergence, along with classical works such as those of Reynolds (boids and steering behaviour), as well as ALife. Affecting emergent properties toward solving abstract computational problems (function optimisation and approximation) is our bread and butter.

Broader notions such as the wikipedia definition of computational intelligence focus on the sociology aspects, although integrates the popular perspectives of eliciting emergent behaviour such as cooperation, coordination, and collaboration. Pushing harder in this direction results in related notions such as collaborative intelligence, wisdom of the crowds, and smart mobs, all very popular memes. MIT are doing some interesting projects along these lines, and maintain a wiki for jumping starting their take on things.

The O'Reilly perspective is different again. Harnessing collective intelligence is a central tenant of Tim's Web2.0 definition. Specifically, the principle is highlighted in notions of network effects by default (automatic aggregation) and improvement scaling with user uptake (product is better the more people that use it). O'Reilly have a book out called Programming Collective Intelligence the contents of which provide examples of exploiting machine learning and data mining (think computational intelligence) for sub problems you may have in collective intelligence web sites, such as recommendation, clustering, regression, classification, optimisation, etcetera. Commenting on the book, Tim highlights a number of seminal examples of harnessing the network effects, specifically PageRank for webpages, Flickr's interestingness measure of photos, amazon's recommendation approach, last.fm's similar artist radio, ebays reputation system, and googles AdSense. Tim's examples highlight the effective use algorithms that exploit user provided content in some form or other.

More recently, Linden's comments from the CI foo camp highlight the emergence adage precisely, and in an internet context as: "The network knows what the nodes don't". Two additional interesting points made in the post were the questions over whether internet-based division of labour were examples of CI, and the effects of gaming in winner take all systems like Digg and google. I was interested in the former point because Linden and others point to "Man Reduce" and the turk as questionable examples. They highlight that CI may delineated by cases in which you could not achieve the solution without the network effects of multiple users. For example, GalaxyZoo and colour naming may not be CI, although crowdsourced stock predictions (along with other things humans are good at approximating like numbers in general) and amazon recommendation are examples of CI.

I like the O'Reilly perspective as it resonates with an intuitive computational approach of devising bottom-up algorithms in an effort to elicit desirable top-down (emergent) effects. The sociology perspective motivates a consideration the factors that influence the actors contributing to the system. Specifically, in terms of the agents performing the low-level computation or generating the information on which the bottom-up algorithms operate. In the cases presented by Tim (listed above), it seems that the emergent effect is a service provided to users that operates on the bottom-up aggregation of user-generated content (information). For example search on webpage corpus, location of interesting photos, and product recommendations.

Collective intelligence seems to involves an effective aggregation algorithm (collective) for per-user data presented as an service (intelligent). The so-called emergent cognition here is bounded strongly by the human designed aggregation algorithm and all the quality or monetary-based optimisation that occurs on the transformation process. The emergent problem addressed by bottom-up information is defined by the scope of what the service provides: such as relevant information to a query, books you are likely to buy, etc. I think there is an opportunity in a broader consideration of emergence, such as the robustness that it affords, or the unboundedness (open systems theory) that it may afford when there is no right and wrong (top-down optimisation). I need to think about this.