[3Xin2] Poster session 1

[3Xin2-34] Evaluation of Session Segmentation Methods Using Item Embedding

〇Yongzhi Jin¹, Kazushi Okamoto¹ (1.The University of Electro-Communications)

Keywords:Session Segmentation, Distributed Representation, Clustering, Classifier

In the context of information recommendation, a session refers to a collection of actions or operations performed by a user within a specific time frame when using a website, application, online service, etc. A session-based recommendation system learns user preferences from input session data and provides item recommendations. However, since user preferences change dynamically, a single session may contain items from different contexts. To capture user preference changes, it is important to divide a session by each context and to analyze the segmented sub-sessions. Therefore, accurate session segmentation methods are one of the important challenges. In this study, we investigate segmentation methods using cosine similarity, k-means clustering, and classifiers (LightGBM, SVM, LR). The methods use three distributed representation methods: Item2Vec, Word2Vec, and OpenAI. We evaluate each method based on the annotated session dataset and compare their accuracy in terms of F-measure, PR-AUC, and ROC-AUC. As a result, we found that the combination of cosine similarity and Item2Vec is particularly suitable for session segmentation tasks, as its values of F-measure and PR-AUC are 0.714 and 0.794, respectively.