Emotion Aware Recommender Systems
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Recommender Systems help users to overcome information overload by making predictions and recommendations that meet users’ tastes and preferences. A user’s mood influences his/her decision-making in choosing from a list of top-N recommended items. However, Recommenders do not track users’ moods when making top-N recommendations to users. Thus, users often found stale recommendations in the top-N list. I proposed to enhance Recommender Systems by tracking users’ moods and make top-N recommendations based on the updated users’ and items’ emotion profiles. In recognition of several limitations: (1) emotion-labeled attributes are not readily available in datasets, (2) lack of standard definition for emotions and procedure to collect and label emotion metadata, (3) not all objects have a face for facial emotion detection and recognition despite facial micro-expression detection and recognition of basic human emotions are popular methodology to label a person’s primary facial emotional expressions, I developed a text-based Tweets Affective Classifier model capable of emotion detection and recognition based on Ekman’s six basic human emotions and neutral emotion. This model is then used to extrapolate basic human emotions from the subjective text of objects such as movie overview or product descriptions. Furthermore, I developed an innovative Affective Aware Pseudo Association Method (AAPAM) to pseudo connect disjoint objects in datasets within the same or different information domains. This research has shown that an Emotion Aware Recommender could track users’ moods in making subsequent top-N recommendations contain serendipitous items, thus overcoming the cold-start and staleness issues confronted in the field. Using the Affective Index Indicator (AII) to pseudo connect disjoint users or items for making recommendations in Collaborative Filtering is more efficient than the traditional Collaborative Filtering computing through rating matrix. I further extended the APPAM to support decision-making strategies in a multi-user group. Finally, I found that applying users’ and items’ emotion profiles in a system simulcast group can improve the throughput of top-N recommendations.