Recommendations Based on User - Machine Learning
The below information outlines the algorithmic techniques that are used to calculate recommendations for the following features:
- Top Picks for User
- Inspired by Your Subjects
- Subject Suggestions
These features provide training recommendations that are generated based on a specific user. These recommendations are generated through a hybrid of many algorithms. Cornerstone MLP uses a hybrid of multiple types of algorithms and mathematical techniques to generate recommendations. This includes:
- Interaction over time
- Collaborative Filtering
- Career Mobility
- Birds of a Feather
- Best Sellers
- Time Trends in Specific groups
- Boot-strapping
- Algorithms & Mathematical Techniques
- Collaborative Filtering
See below for descriptions of these techniques:
Collaborative Filtering
The Collaborative Filtering technique analyzes the preferences of many different users, such as the training that the users registered for, and uses this information to make training recommendations in which the user is likely to be interested. There are various techniques used, one of which is a modified version of the Latent-Factors Matrix Factorization algorithm, which runs on the big-data cluster within the Spark execution layer.
Example: Collaborative Filtering analyzes three users and the training that they have taken. It determines that User 1, User 2, and User 3 all took Training A. User 1 and User 2 both also took Training B. Since User 3 has similarities with Users 1 and 2, since they all took Training A, the MLP might recommend Training B to User 3, because Users 1 and 2 both took it.
User Position and Context-Based Recommendations
The location, department, and position that a user belongs to defines a user’s contextually relevant neighborhood of similar users, which provides a basis for pertinent recommendation. This is particularly helpful in cases where there are new users who have been added to the organization, for whom there does not necessarily exist a history of past preferences. Having this neighborhood of users is crucial so that the algorithms can still provide training recommendations based on the most popular training.
Career Mobility
The Career Mobility technique analyzes the career path of users and the training they took while in each role. This is dependent on the portal utilizing the Position OU. The MLP is able to map the history of an employee across various positions within the organization. This history of transitions allows the MLP to recommend training which users that were previously in the position registered for, training that users whose career transition you are following are registering for and completing, etc. This results in training recommendations for a user from a career mobility perspective.
Example: Career Mobility considers two users and the training that they have taken. User 1 and User 2 were both in Position A and then moved to Position B. User 1 is currently in Position B and User 2 has moved to Position C. By analyzing the hierarchy, Career Mobility considers Position C to be a higher position than Position B. Career Mobility also considers that User 2 has registered for and completed Training X while in Position C, but User 1 has not. Since User 1 appears to be following a similar career path to User 2, the MLP might recommend Training A to User 1.
Birds of a Feather
The MLP incorporates specific notions of similarity between users to derive recommendations. Amongst its recommendations are courses inferred from what other similar users have taken in the past. These recommendations are derived from the juxtaposition of users' tastes and training traits. The Athena recommendation engine provides a “bootstrapping” recommendation for users based on the notion of similarity arising from the triplet of <location, department, position> they belong to. Furthermore, clustering based on user similarity influences some of the recommendations. MLP uses registrations of training in the past three months in the minimal OU triplet in order to create bootstrapped recommendations.
Latest Version
While a training may have been prevalent in the past, there is no guarantee that the training will meet recommendation criteria when users access the training carousels. Often, there are latest iterations or versions of the training, representing enhancements or alignment to new goals. As a result, there is a lineage chain weaving through the various learning objects, each representing the next iteration in the chain. It is always preferable to recommend the latest avatar of a course-lineage. The Athena recommender performs this lineage discovery, and makes recommendations appropriately.
Training Equivalencies
In CES (client portal databases), data on equivalent courses and courses mapping with subjects and equivalent subjects is stored. These provide for a rich notion of similarity between learning objects, and this data is used in Insight applications and the Athena recommenders.
Popularity / Best Sellers
Popularity-based recommendations are similar to the bootstrapping recommendations and context based recommendations mentioned above. They operate on the principle of identifying a neighborhood of users and recommending training that is most popular in the neighborhood to users within the neighborhood.
Temporal Trends in a Neighborhood
Temporal Trends in a Neighborhood are popularity-based recommendations in which the neighborhood is the position to which the employee belongs. The most popular courses for this position are the recommendations for the employees in this position. When this cohort has less than a threshold number of employees, the Athena recommender moves hierarchically up to superset positions that contain this position as a subset, until it reaches a cohort with more than the threshold number. Then, popularity and current trends are collectively considered, prioritizing recent learning engagements.