Kakao Enhances Webtoon Related Recommendations with Content-Based Filtering

Kakao introduced content-based filtering technology, which analyzes the content of items consumed by users (such as category, text, and images) to recommend similar items. Unlike collaborative filtering, which leverages consumption history from other users, this approach is based on the item's own information, making it effective for recommending items with little consumption history or new items.

One-hot encoding and embedding techniques are used to calculate item similarity. Text is vectorized using pre-trained models such as Word2Vec or BERT, while images use features from layers prior to the classification layer of models pre-trained on ImageNet as embeddings. Specifically, in the webtoon domain, an art style extraction model developed by Kakao Brain is applied to capture visual similarity.

In KakaoWebtoon's related recommendation area, results from content-based filtering and collaborative filtering are combined. After merging candidate groups extracted via text, image, and art style embeddings with collaborative filtering results, a Multi-armed bandit algorithm is applied to arrange the recommendation list in an order that maximizes the user's actual viewing probability. This allows users to receive recommendations that reflect both current popularity and content-based similarity of works.