Administer > Manage WebSphere Commerce features > Marketing > Collaborative filtering > WebSphere Commerce Recommendation Engine powered by LikeMinds > Background LikeMinds utilities
How the sifter selects mentors
The sifter finds mentors for users, using the information from the rating or transaction data for the LikeMinds server engines.
How the Sifter Works
How the mentor selection process works
In order to fully understand the sifter, it is important to have a clear idea of how the mentor selection process works, and how to set configuration parameters that increase the accuracy of the recommendations. You should understand the following terms before proceeding:
- Collaborative Filtering
- Collaborative filtering (CF) is a technology that calculates the similarity between users. It uses the behaviors of those who most closely resemble any given user as a functional basis for making predictions and recommendations for that user. Given that definition, the process by which "...the entire population of users is analyzed, their fitness as mentors is calculated, and they are assigned as mentors to individual users..." is critically important to the ultimate recommendations that come from the collaborative filtering approach.
The data used to decide on the levels of similarity is based upon clickstream events that comprise the sequence, duration, and outcome of a Web-surfer's session -- all can form the necessary basis for making similarity calculations. From those similarity measures, this data can result in the measured recommendations from those users deemed most similar to any given user.
- Mentor
- A like-minded user that is used as the basis for recommendations for new users. Every user is assigned mentors by the sifter program, whose stored behaviors are judged to be like-minded to the new user.
- Mentor Pool
- While the purpose of mentors is to form the basis for recommendations for those users deemed most similar to them, in its most basic form the mentor pool should reflect a representative sample of users in the transaction set for which the recommendations are desired. And despite the clearly required emphasis on similarity, no recommendation process can make lucid suggestions without a concomitant space of dissimilarity. We might base our final recommendations on the similar, shared tastes discovered in our analysis of the users being considered for entry to the mentor pool, but it is truly the confluence of similarity in purchases with some difference in the items purchased that make the LikeMinds server collaborative filtering-based recommendations possible.
- Sifter
- Creates mentors by analyzing stored user transactional data. The sifter runs in the background when you run the LikeMinds server.
- Coverage
- The volume or number of items rated or transactions performed.
Mentor selection and assignment
Several factors determine the fitness of any user as a mentor. Similarity to any other user is the final arbiter of any mentor's fitness to make recommendations for a specific user. Yet it is coverage (the volume of number of items rated or transactions performed) that is most important when forming the mentor pool. (The mentor pool is a superset of the final mentors chosen to make the recommendations.) Although a case could be made for using the extent of the unshared purchase space as another dimension of dissimilarity, the shared space is where we find the most data (and the most predictive data) for making the required similarity distinctions within the user population.
In a nutshell, while similarity to the user is important, it is the ability of a mentor to contribute items outside of any user's typical purchasing space into the final pool of possible recommendations that qualifies the user as a possible mentor.
For this purpose, the first step the sifter uses in mentor assignment is to periodically create a new mentor pool in an effort to collect a representative sample of experienced users who will then be considered as potential mentors for any user who requires recommendations.
The second step in the mentor selection process is to assign mentors from the mentor pool to be mentors for specific users. The challenge is to create a fair balance between the similarity and the coverage of users being considered as mentors. You can configure the sifter to emphasize similarity, coverage, or to automatically determine which dynamic to emphasize for each user in the final assignment of mentors to users.
Related concepts
WebSphere Commerce Recommendation Engine powered by LikeMinds