People are often confronted with situations where they need to assess the potential value of something that they have never experienced before. We need to find a new book to read, buy a new beauty product, choose a doctor, and know which business is credible. When confronted with these situations, we often turn to experts or peers for recommendations.
Such situations present a great opportunity to marketing professionals for widening their consumer base. Increasingly, marketing professionals rely on automated tools called recommender systems for this purpose. For example, people browsing a book at Amazon are presented with a list of related books of potential interest. The related books (listed under the “People who bought this book also bought” header) are identified by a recommender system that relies upon the historical purchasing patterns of Amazon customers.
More generally, recommender systems suggest items of potential interest to individuals who do not have sufficient personal experience with the items. Using statistical techniques recommender systems can compute personalized recommendations. Broadly speaking, recommender systems can be classified into two types: collaborative and content-based. Collaborative recommender systems make recommendations based on the prior experience of other users, while content-based systems make recommendations based on features or descriptions of the items themselves. Content-based systems are useful for music, movies, books etc, but unlikely to be useful for mass-market commodities.
Collaborative recommender systems draw on the historical experience of some users to make recommendations to other users.
Figure 1 Schematic Representation of a Recommender System
To make recommendations, a collaborative recommender system must perform the following tasks:
Elicit Preferences: Recommender system needs to learn about the users’ preferences and store them in user profiles. This can be done either explicitly asking the user to rate certain items; or by using implicit measures such as purchase history, search history etc.
Compute Predictions: Based on the users’ elicited preferences, recommender systems predict how well a user would like an item she has not used before. Several alternative algorithms are used to make predictions and thus recommendations. Recommender algorithms can also be classified based on whether they use correlations between users or between items.
In the user-user approach, correlation between two users is computed based on the ratings of items that are rated (or used if the recommender system uses implicit measures) by both the users. The recommender system computes correlations between all such user pairs, which can be used in a variety of ways.
One of the popular approaches is to use these correlations as weights when making predictions by taking a weighted average over the opinions of other users who have rated an item. Another approach is to use the user-user correlations to divide the user population into clusters of users, where users in the same cluster are considered to have similar preferences. Predictions for a user’s hitherto unrated item are made by averaging the opinions of the other users in her cluster.
In systems that use the item-item approach (e.g. Amazon.com), correlations between item pairs are computed instead of correlations between user pairs. Correlation between two items is computed using ratings of all the users who have rated both items. The system recommends items that are highly correlated with the items that are highly rated (or used) by the user.
Make Recommendations: The objective of a recommender system is to present each user with items she is most likely to enjoy. Marketing professionals need to make several important design decisions in this step. Besides the obvious design decisions about how to present the recommendations, they also need to determine the maximum permissible error in the recommendations. The error could be an error of commission (incorrectly recommending an item) or an error of omission (not recommending an item that should be recommended).
The margin of permissible error for either type depends on the benefit of a correct recommendation and the cost of an incorrect recommendation. Consider a hypothetical recommender system which prescribes medical treatment. Here a good recommendation not made, or an incorrectly made recommendation could be a matter of life or death for the user. For a system recommending luxury products, a recommendation is unlikely to have such important ramifications for the user, but it is still an important decision for the service provider. The marketing professional has an incentive to provide more recommendations if it leads to increased sale. At the same time the reduced usability due to poor recommendations may result in the user losing faith in the recommender system altogether.
- Riedl J. and Konstan J."Word of Mouse: The Marketing Power of Collaborative Filtering". New York: Warner Books, 2002.
- Schafer, J.B., Konstan, J., and Riedl, J., Recommender Systems in E-Commerce. Proceedings of the ACM Conference on Electronic Commerce, November 3-5, 1999.