Our collaborative filtering system was in use at Xerox's Palo
Alto Research Center(PARC) from approximately August 1993 till
December 1993. During this time, the number of votes cast by users of
the system were recorded, and are depicted in figure 
.
Each bar represents the number of votes cast during a week long period
beginning on a Sunday. Over the period of this study, 44 users voted
at least once. 24 of the voters used the nn news reader and
the other 20 voters were xrn users. There are approximately
40 users of nn at PARC, so roughly half of the
nn users voted.
Figure: Number of votes cast
at Xerox PARC during the period from August 1993 till December 1993.
Each bar represents the number of votes cast during the week beginning
on the listed Sunday.
Over the course of the study, the fewest votes were cast the week of August 29th. Coming directly before the holiday of Labor Day weekend, it is possible that this lull in voting was caused by many Net News users taking vacation time. The failure of the voting level to return to its previous level after the holiday weeks may be attributable to a cause identified by a group at Bellcore researching ephemeral interest groups.[3] This group found that adding features unobtrusively to commonly used programs resulted in users frequently forgetting about the new features. One possible explanation for the dip in our voting levels was that people went away on vacation, and then completely forgot about the voting options when they returned to work. Perhaps our reminders to vote were too subtle.
The upswing in voting the week of 9/26 coincides with the release of a new version of the news readers and the sending of an email message to a site-wide mailing list reminding people of the system's existence and how to vote. After this resurgence, the voting levels appear to again drop slowly over time.
Examination the voting records indicates that the collaborative filtering system did not achieve anywhere near a critical mass of users. In almost every newsgroup in which any votes were cast, all the votes came from a single user. Had several users contributed votes on the same article, we would have taken this as a sign that users were paying extra attention to articles their peers had read and voted on. In examining the effect of the votes which were cast, it appears that there is so little overlap in which newsgroups are read by the users of nn and xrn at PARC that it is likely few users had articles pointed out them by the collaborative filtering system.