Statistical and Generative Models with Subtitle Extraction for Next Product Title Generation

Honghee Lee, Youngjae Chang, Kyuri Choi, Jongwuk Lee, Youngjoong Ko

Keywords: Session-based recommendation, next product title generation, Markov model, generative model, subtitle extraction

TL;DR: We propose an effective method based on statistical and generative models for the next product title generation, which achieved 3rd place in KDD Cup Task 3.

Abstract: Session-based recommendation aims to predict the next item from the user's actions in the ongoing session. It mainly suffers from the cold start item problem, referring to the difficulty in providing accurate recommendations for items with little or no previous interactions. The KDD Cup 2023 Task 3 (next product title generation) addressed this challenge to improve session-based recommendation. This paper proposes an effective solution for the next product title generation using statistical and generative models. In this process, we optimize a model combination strategy that selects the optimal prediction model for each session based on predefined conditions. The title of the last product serves as a fallback when the session does not meet any conditions. We also devise subtitle extraction techniques to identify a common element among multiple predicted titles. Consequently, our team, We Bare Bears, has achieved third place in the KDD Cup Task 3 with a BLEU score of 0.26998, demonstrating the effectiveness of our proposed solution.

Submission Number: 11