Demographic attributes based, cold-start recommendation of modules in organizational learning

Abstract

Organizational learning is the process of creating, transferring, and retaining knowledge within an organization. It is of high importance due to the highly dynamic nature of the modern employee base. Moreover, new employees perform sub-optimally and get frustrated when such knowledge and expertise are not readily accessible to them. While many organizations use an organization-learning platform to bridge the knowledge gaps, both for new and existing employees, their effectiveness is being questioned due to lack of relevance, incoherent order of modules to be followed, and lack of fit with the learning style of an employee. While recommendation systems could overcome these challenges, it is difficult to provide a fitting set of recommendations for new employs who do not have any history with learning management system (aka., cold start problem). We address the cold-start problem in recommender systems for organizational learning using the demographic information of employees. First, similar employees are grouped together based on their demographic attributes. Second, the modules that they follow are clustered according to their similarity. Then the orders of modules and the employee clusters are linked together in such a way that the number of module orders related to a user cluster is maximized. When a cold-start employee enters in to the system, his closest employee cluster is identified based on the demographic features and recommendations are generated considering the module sequences which have the least dissimilarities to the other module sequences in the linked module order cluster. We then tested the proposed technique using a synthetic dataset generated considering a medium scale organization. The dataset consists of age, gender, department, designation, and the order of learning modules followed by the employees. The proposed recommendation system has good accuracy, e.g., 71% of the module recommendations were more than 90% similar to the actual module orders.