Librarians talk off and on about the need for us to offer Netflix / Amazon-style automated recommendations for our patrons. It seems almost self-evident that this is something patrons have come to expect. But there’s a self-evident question about this that we must ask:
Have patrons actually told us that they want this type of service from a library?
Or do we just assume that they want this?
A library doesn’t fulfill the same role in people’s lives that Netflix does, or that Amazon does. Our patrons don’t necessarily expect the same service models from us. We may be holding ourselves accountable to a false comparison here. This is a prime example of the need for us to base decisions on verifiable user data.
Some years ago, I read a blog post written by one of the developers at Amazon who helped to build their automated recommendation service. He expounded on some of the lessons learned in the process.
When Amazon first launched their recommendation service, it actually didn’t work very well. It was clunky and the recommendations weren’t that good. It didn’t meet the standards that Amazon had set for it. The developer identified two main reasons for this:
- The algorithm was rudimentary to start. They knew that they needed real-world user data and feedback to tell them how to improve it. Improving the algorithm had to be an iterative process—by its nature, they had to launch it in an imperfect state.
- They didn’t have enough user data to start. These sorts of data mining functions improve as the available pool of data gets bigger—the more data you can analyze, the more accurate and nuanced the results will be. But these functions can’t work at all without a certain minimum amount of data—too little data and you get nothing. Amazon discovered that the critical mass of data necessary for this service to function turned out to be a lot higher than they had predicted.
The main reason why Netflix and Amazon do so well with their automated recommendations is because they have such massive pools of data to analyze.
Setting aside the significant concerns I have regarding patron privacy, the biggest obstacle to libraries instituting this sort of automated recommendation service is that we don’t have anything like enough data for these algorithms to work optimally. All the data in your local ILS probably isn’t enough.
Right now, there are two service models that libraries have available to try and make automated recommendations work:
- Subscription services like NoveList and BookLetters function based on professionally curated lists. They don’t depend on any user data at all. Professional curation is labor-intensive, so the sticker price for top-of-the-line professionally curated recommendation services can be fairly high. In addition, when a library subscribes to such a service, it cedes local control over the process to an outside third party.
- BiblioCommons offers an automated recommendation service via the catalog. To offer this functionality, they combine data from multiple library systems in order to generate a large enough pool to make algorithmic analysis possible. The cost of this, though, is that local library data loses its ability to meaningfully affect the outcome. Local data gets diluted. The recommendations offered by BiblioCommons therefore can’t reflect local trends or local cultural factors in any significant way.
One of the great values of public libraries is that they’re embedded in local culture and society. They serve local needs with locally relevant resources and services. Netflix / Amazon / BiblioCommons-style automated recommendation services come at the cost of this locality. *
The alternative to these automated recommendation systems is to rely on local, personal, non-automated RA service. It’s my opinion that personal RA service is the single greatest advantage that libraries have over Netflix and Amazon in any case. No automated recommendation service can yet duplicate the value of a recommendation from someone who understands you as a unique individual, and not merely how you compare to a statistical model. No automated recommendation service can yet duplicate the value of person-to-person contact embedded in the local cultural milieu.
This is not to say that these costs might not be worth it for the benefits of a well-functioning automated recommendation service. But in order for us to judge if the cost is worth it—both in terms of sticker price and also in terms of what we’re required to concede in order to get it (patron privacy notwithstanding)—we need to know if this is a service that our patrons actually want from us.
Thus, I circle back to the most important question in this discussion:
Have patrons actually told us that they want this type of service from the Library?
Then again, as Steve Jobs once said—no one knew they wanted an iPod until he showed it to them…