Improving Library Recommender Systems with a Hybrid Interaction-ALS Approach under Data Sparsity

A large library collection needs to be presented to users according to their needs and preferences to provide time efficiency for library users. An efficient and accurate library collection recommendation system is needed. The development of this system is highly dependent on data availability. Recommendation systems generally require user ratings to be able to predict user preferences for a collection, to then be developed to recommend other collections. On the other hand, the system in the library does not use many user ratings but rather relies more on borrowing interaction records. The problem that often arises is data sparsity because the number of library collections is generally much greater than the number of users, and will be even smaller when compared to the number of active users who interact with a particular collection. This study aims to overcome the problem of data sparsity, which often interferes with the accuracy of recommendation systems, especially extreme data sparsity where the level of data sparsity exceeds 95%. The experimental results show that the hybrid model, by combining User and Item Similarity and the Alternating Least Square (ALS) model, can reduce the problem of data sparsity while still producing an acceptable level of metric evaluation ranking. The fill-out matrix can be improved following the sequential process. The original User-Item (UI) matrix of 0.53% increases to 2.20% when User Similarity is applied, then increases to 19.52% when Item Similarity is applied. A 100% fill-out occurs when the last sequential is applied using ALS. The precision, recall, and f1 obtained in this study are 0.84, 0.86, and 0.80, respectively.