Kei HARADA (原田 慧)

Complementary recommendations suggest combinations of useful items that play
important roles in e-commerce. However, complementary relationships are often
subjective and vary among individuals, making them difficult to infer from
historical data. Unlike conventional history-based methods that rely on
statistical co-occurrence, we focus on the underlying usage context that
motivates item combinations. We hypothesized that people select complementary
items by imagining specific usage scenarios and identifying the needs in such
situations. Based on this idea, we explored the use of large language models
(LLMs) to generate item usage scenarios as a starting point for constructing
complementary recommendation systems. First, we evaluated the plausibility of
LLM-generated scenarios through manual annotation. The results demonstrated
that approximately 85% of the generated scenarios were determined to be
plausible, suggesting that LLMs can effectively generate realistic item usage
scenarios.

Serendipity in recommender systems (RSs) has attracted increasing attention
as a concept that enhances user satisfaction by presenting unexpected and
useful items. However, evaluating serendipitous performance remains challenging
because its ground truth is generally unobservable. The existing offline
metrics often depend on ambiguous definitions or are tailored to specific
datasets and RSs, thereby limiting their generalizability. To address this
issue, we propose a universally applicable evaluation framework that leverages
large language models (LLMs) known for their extensive knowledge and reasoning
capabilities, as evaluators. First, to improve the evaluation performance of
the proposed framework, we assessed the serendipity prediction accuracy of LLMs
using four different prompt strategies on a dataset containing user-annotated
serendipitous ground truth and found that the chain-of-thought prompt achieved
the highest accuracy. Next, we re-evaluated the serendipitous performance of
both serendipity-oriented and general RSs using the proposed framework on three
commonly used real-world datasets, without the ground truth. The results
indicated that there was no serendipity-oriented RS that consistently
outperformed across all datasets, and even a general RS sometimes achieved
higher performance than the serendipity-oriented RS.

In information recommendation, a session refers to a sequence of user actions
within a specific time frame. Session-based recommender systems aim to capture
short-term preferences and generate relevant recommendations. However, user
interests may shift even within a session, making appropriate segmentation
essential for modeling dynamic behaviors. In this study, we propose a
supervised session segmentation method based on similarity features derived
from action embeddings and attributes. We compute the similarity scores between
items within a fixed-size window around each candidate segmentation point,
using four types of features: item co-occurrence embeddings, text embeddings of
titles and brands, and price. These features are used to train supervised
classifiers (LightGBM, XGBoost, CatBoost, support vector machine, and logistic
regression) to predict the session boundaries. We construct a manually
annotated dataset from real user browsing histories and evaluate the
segmentation performance using F1-score, area under the precision-recall curve
(PR-AUC), and area under the receiver operating characteristic curve. The
LightGBM model achieves the best performance, with an F1-score of 0.806 and a
PR-AUC of 0.831. These results demonstrate the effectiveness of the proposed
method for session segmentation and its potential to capture dynamic user
behaviors.