Summary

  • LibraryThing launched UnSuggester in November 2006, using unexpectedly sparse co-ownership to produce deliberately contrary book lists.
  • In May 2008, readers found Gaudy Night among their anti-recommendations. Tim Spalding attributed the result to its catalogue identity changing during compilation.

Being told to avoid a favourite book is an efficient way to get a reader's attention. LibraryThing managed it with Gaudy Night. In May 2008, members reported finding Dorothy L. Sayers's novel in their UnSuggester lists despite owning it; some volunteered considerable affection for it too.

The computer appeared to have developed literary objections. Tim Spalding's explanation was less glamorous: the book's catalogue identity had changed while the list was being calculated.

A recommendation engine with its jacket on backwards

LibraryThing introduced UnSuggester on 12 November 2006. Start with a book, find the people who had catalogued it, then examine their other books. Titles appearing less often than expected supplied the contrary suggestions. The site turned a familiar recommendation mechanism into a parlour game.

Its records supplied the raw material. Each newly entered library helped shape which books appeared to keep company and which seemed to avoid one another. The joke depended on those absences being meaningful.

There was a clock attached to the calculation. In a December 2006 post, Spalding explained that the expensive results were cached for a while. Someone entering books and immediately trying the service could encounter a result calculated before the new arrivals.

The mystery was in the catalogue

The 26 May 2008 announcement expanded the offer to personal lists, including up to 500 entries in Your Unsuggester, and explanations of what produced a recommendation. Readers now had more scope to inspect what the system thought of their collections.

Gaudy Night became an awkward exhibit. In the official discussion, Spalding first suggested that combining catalogue entries while the computation ran could leave it looking at an old work reference. Later on 27 May, he stated that the work codes had changed during compilation: the calculation expected a quantity and kept finding zero.

The novel had acquired an impressive statistical absence while remaining comfortably present on readers' shelves. A change in recordkeeping was wearing the costume of a reading judgment.

Declining a suggestion had its own meaning

The discussion also asked what the new “no thanks” action did. On 28 May, Spalding said it affected the individual user's list. Readers could have quite ordinary reasons for dismissing a title: they might already have read a borrowed copy, for instance.

That distinction mattered alongside the catalogue problem. A missing entry, a delayed calculation and a personal dismissal each told a different story. A single unfriendly-looking list could conceal any of them.

UnSuggester made automated recommendations entertaining enough for readers to argue with. Their objections brought its bookkeeping into view. Gaudy Night had supplied the mystery; the catalogue supplied the suspect.