14.5 Collaborative vs. Content-Based Filtering: A Summary
Content-based filtering’s major advantage is that it could be used in situations for which there is no previous user data. If a new amusement park opened next year, then its proprietors would have no existing ratings or feedback from users to feed into a recommendation engine; therefore, forming recommendations based on the rides themselves would be a sensible approach to take.
In the real world, collaborative filtering is far more common. While collaborative filtering does require some initial user feedback data in order to get started, it is far easier to employ, once the “cold start” has been overcome.
To see why, let’s imagine that we are running a streaming music service. From our user data, we see that users who listen to Adele tend to also listen to Lionel Richie. What can we do with this knowledge? For starters, we can recommend Adele to listeners who have expressed an interest in Lionel Richie, and vice versa. Doing so requires no knowledge on our part about either artist. In fact, we can make this recommendation without even knowing anything about music at all. If, alternatively, we had to deeply analyze the structure and style of these two artists’ song collections in order to place them together, this would be time-intensive (and therefore cost intensive!) The same can apply towards nearly anything – if Amazon shoppers who buy Product x often go on to buy Product Y, Amazon does not really need to know, or care, why this happens. They simply need to know that it does. The same could be said for any product or service purchases, or any consumption of streaming entertainment.
An effective recommender system can have a considerably positive impact on a business’s bottom line – it can “nudge” consumers towards purchases that they may have already been considering. Many Amazon shoppers half-jokingly say that “Amazon knows what I want to buy before I even know that I want to buy it!” The truth in this humor is that Amazon possesses mountains of data regarding user purchase patterns; based on all of this data, along with data regarding your own recent buys, they really might be able to predict your likely purchases six months from now!
Recommender systems can also directly impact consumers’ feelings about a product or service. Some customers feel a closer sense of connection to a service that can offer tailored, personalized recommendations that suit their needs.
The field of recommender system development is constantly developing, especially as it concerns the nature of the interactions between a service and its users. While the process of setting up a matrix of cosine similarities or jaccard similarities is standard and straightforward, the tougher, more interesting questions come next. How many items should a fast food kiosk ordering system suggest to a customer before checkout? How many times should a streaming entertainment service recommend something to a user before determining that she isn’t interested in it? How might users from different cultural backgrounds react to softer, or more aggressive, sales pitches? The complexity that comes with the human considerations, alongside the data considerations, ensures that recommendation system development will remain a dynamic field for many years to come.