Performance improvement by adjusting ratings´ mid-point value in the neural network based recommendation models

To understand customer preferences and provide attractive new services, companies have been developing analytical techniques involving customer information. They collect customer evaluations regarding the quality of content, provided service, or movie, and use the data for the subsequent recommendation of new products and services. When a customer evaluates a specific object with the rating ranging from 1 to 5, it is easy to interpret the extreme values like 1 or 5; however, when the earned rating is 3 - the mid-point value - it is difficult to understand clearly whether it is positive or negative. As the rating 3 refers to the average value as well as the median value, we considered that it is more effective to use the average value instead of the rating 3 because it reflects all ratings evaluated by customers. In this study, combined with the prediction system using the neural network, we obtain improvement in the analysis of commercial data by using the average value compared with unadjusted ambiguous ratings.