by Trevor Legwinski
-
February 2, 2010 at 12:33 pm
· Filed under Strands Recommender
People’s tastes and desires are transient, ephemeral and even chaotic sometimes. Some people have well-defined tastes and they are guided by the trends, that most of society is following. Others take some ideas from their families, friends and the media (which is not necessarily the same that they will want tomorrow). This invariability in trend following results in a great deal of unpredictability. This variability and unpredictability of consumers brings headaches to recommender systems.
As most of you may or may not know: When delivering recommendations to customers different methods can be used.
On one hand we have personal recommendations, that is, those recommendations based on the user’s profile. Which are built from the user’s interaction history on the site and explicitly mentioned personal preferences. On the other hand we have collaborative filtering (CF) methods, which make recommendations based on the behavior of other users whose preferences are similar to those of the target user. However, these methods do not consider how the customers’ purchase behavior may vary over time.
Recommending for the Mainstream:
The mainstream audience of trendsetters and their followers mimic similar behaviors
Let’s use a customer who follows the fashion trends or certain patterns driven by the trendsetters. The behavior shown over successive visits to a site will be very similar to other people. The recommendations based on past behavior of these customers can work, but the recommender is not going to be able to react in time to a trend change if we rely only on this data. Only when one begins to search for items according to the latest fashion trends, will they start receiving new recommendations tailored to their preferences. The collaborative filtering, in a way, could anticipate that moment, because there’s always someone who has already begun to follow the new trend, and has already made purchases or have visited product pages that the target user will want soon. The store owner has an important role here too. A good recommendation engine is aware of emerging trends amongst products, and automatically adjusts business rules so that each person receives recommendations based on the new lines of products without having to wait for the users profile to “react” to the trends.
Outliers Can Put A Wrench In The Recommendation Equation:
What would you recommend for her?
However, recommender systems can sometimes fail for people whose tastes fall out of the mainstream, or have a very narrow or too wide range of tastes. Changing tastes and preferences are usually not sudden but morph over time. The recommendations for these users based on their profile and their interactions are inherently harder to create. In these cases current data should be given greater weight than historical data, focusing on the changing preferences of these users who follow their own way.
Collaborative filtering is more complicated. Because there is little correlation between these outsiders and the rest of the community and there is fewer users of that type lacking matches between them. This situation makes it difficult for the recommender to predict which will be the customers next preference. But these kinds of customers are not a hopeless case. The CF could work if the recommendation engine would only consider a few customers that may be similar to our target, and they are usually his/her friends. If the recommendation engine can automatically select smaller groups of people to learn from and to trust, they will receive more accurate recommendations that are tailored to their tastes and their possible evolution. This is what we call Not Anonymous Collaborative Filtering (NACF).
What’s Best? Both Methods or User Choice?
We usually receive different types of recommendations at the same site, depending on the location, based on different data sets or based on a mix of them. Some of them are going to work better for you, than others. The ideal case for both situations previously shown is that the user could choose what sort of recommendation they want, based on what other users bought, only showing related items, similar or complementary, not showing more one type of recommendation or only what my friends recommend me. This way, the customer can choose what to use and thus enhances both the learning of the recommender according to your changing preferences, and the results obtained by the user and the store.
Is this then a failure of the recommender? Should they satisfy everyone equally and there is a need to refine the algorithms for finding the right recommendations for everyone? or do we have to admit that it doesn’t always work? Would it improve if the recommendations based on smaller groups of users, and consistent with the customer? Or is the point that the customer’s should choose what kind of recommendation they wish to receive? Post a comment below or tweet at us on Twitter @strandsrecs.
