Review representation learning and profiling for Recommendation System
Nurkhairizan Khairudin, Nurfadhlina Mohd Sharef, Azilawati Azizan · 2021
The utility function of users and items in a traditional recommender system is based on rating. However, data sparsity is a fundamental issue in which the user rates a small number of items in comparison to the available items. This rating alone is sometimes insufficient for precisely understanding users' behaviours. A user's general preferences on an item can be shown through the overall ratings, yet they may be dissatisfied with a specific feature of the item. Since most commercial website nowadays, allows user to express their opinion through the review text, then there is an opportunity to precisely understand the user preferences via this element. However, it is essential to represent the review texts in a suitable representation to be used as one of the recommendation components. Therefore, this research aim is to create a relevant review profiling framework based on the data analytic assessment and some observation of the user behaviours reviewing pattern. The review texts are classified into groups of word count, and the consistency of the relevant category is identified. This selected category is used to generate a representation of reviews as one of the components of recommendation processes and resulted in less computational complexity and processing time. The experimental results show that review text count distribution fall in the range of 10 to 30 with a range from 20–30 word count has the highest number of users for almost all datasets.