programming archaeology, pedantics, and machine intelligence

03 Apr 2013

Book Review: Building Web Reputation Systems

Randy Farmer has done pioneering work as the architect of some of the earliest and most influential virtual communities. His book introduces reputation as a generic decision framework.

The basic insight is that many concepts can be expressed in terms of computing and tracking the reputation of an entity in a context. Interesting examples include

  • fraud detection
  • spam filtering
  • ranking
  • trust metrics

though the book focuses more on traditional interpretations of reputation (as in Xbox achievements).

Content is roughly one third ideas, one third examples, and one third implementation. Implementation covers both reputation models and reputation platforms. A reputation model is a specific semantics of reputation, i.e. a way to translate events into constituents of the reputation score of some entity and an interpretation of that reputation score in one or more contexts. Reputation scores are typically calculated by combining various inputs via simple operations (such as aggregation or weighting) to yield a single output. Reputation platforms are the things on top of which reputation models are executed.

The book’s main contributions are a grammar for reputation models, a detailed explanation of Yahoo! Answers community moderation, and a description of the Yahoo! Reputation Platform, which is in use across Yahoo!’s range of products. Unfortunately, some of the systems described are patent-encumbered.

Open questions include how more complex, related problems such as elections (voting in favor of a ranking rather than a single candidate) and trust management systems (transitive trust metrics) can be fit into the reputation-based view of the world and whether more complex models, e.g. PageRank, can be defined in some extension of the proposed grammar.