Is It Really Complementary? Revisiting Behavior-based Labels for Complementary Recommendation

Is It Really Complementary? Revisiting Behavior-based Labels for Complementary Recommendation [email protected], Sugahara, Chihiro Yamasaki, Kazushi Okamoto [email protected], [email protected] The University of Electro-Communications, Tokyo, Japan Summary: Behavior-based labels (BBLs) from co-purchase logs are used for complementary recommendation in training and evaluation, but they often deviate from the true definition. Our study found that most BBLs don't reflect true functional relationships, and models trained on them show inaccuracies in terms of annotated function-based labels (FBLs). We also proposed a method for correcting BBLs using a small set of FBLs. Fig.1: General overview of this study Fig.2: Result for RQ1 and RQ2 Background: Complementary Recommendation A type of item-to-item recommendation focusing on "What should be purchased together for an item?" Experiment Fig.3: Result for RQ3 We conducted experiment on ASKUL dataset. (See Fig.1) * https://www.askul.co.jp Labels: BBLs were firstly constructed according to [1], and FBLs were generated by annotating part of with the labels defined in [2]. Metrics: Hit@ and nDCG@ . A model ranks target items for a query item based on complementarity and measures whether the top-k items are actually complementary to the query. Label Construction for Training and Evaluation RQ1 Are BBLs really complementary? Behavior-based Labels (BBLs): Target Method: For each query item in , rank target items Low cost method by the co-purchase frequency of users based on complementarity identified by . (Called as BBL) McAuley et al. [1] first contributed: crawled the recomm. areas Result: The Hit@1 of BBL was below 0.5 pt, indicating that i. "Users who viewed 𝑥 also viewed 𝑦" = complementary pairs identified by BBLs often did not have ii. “Users who viewed 𝑥 eventually bought 𝑦" = complementary relationships from the perspective of FBLs. iii. “Users who bought 𝑥 also bought 𝑦" = iv. “Users frequently bought 𝑥 and 𝑦 together" = RQ2 Can existing models trained on BBLs accurately predict? Risk of deviating from the true definition of complementary Baseline: Sceptre [1] and P-Companion [3]. All of them trained on and rank target items for each query item in . Function-based Labels (FBLs): Result: The accuracy of both models tended to be equal to or Highly Cost, but generally high quality for recommendation lower than that of BBL, hindering their practical application in We previously annotated item pairs with our defined labels in [2]: industrial settings. i. Items and have the same function and usage = RQ3 Can we correct BBLs into trusted ones with fewer FBLs? ii. Item can be replenished with item = iii. Item can be replenished with item = Proposal: LightGBM trained on as a content-based iv. Items and must be combined to be usable = prototype model (LGBM), and BBL re-ranked by the probability of LGBM and label agreement with LGBM (BBL+LGBM). v. When combined with item , item becomes more useful = vi. When combined with item , item becomes more useful = Result: LGBM outperformed BBL in nDCG@ , and BBL+LGBM achieved the best performance. Thus, the use of a contentvii. Combining and makes them more useful = based model trained on small data can be more effective viii. Items and have no relationship. than BBLs. [1] J. McAuley, R. Pandey, and J. Leskovec. 2015. Inferring Networks of Substitutable and Complementary Products. In Proc. 21th ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. 785–794. Substitute Another branded phone Complementary A phone (query item) Its case Subst. Subst. Compl. Compl. Subst. Compl. Compl. Compl. Compl. Compl. Compl. [2] K. Sugahara, C. Yamasaki, Y. Nagi, and K. Okamoto. 2024. A Study of Complementary Recommendation Focused on Functional Aspects. In IEICE Tech. Rep. 17–22. [3] J. Hao, T. Zhao, J. Li, X. L. Dong, C. Faloutsos, Y. Sun, and W. Wang. 2020. P-Companion: A Principled Framework for Diversified Complementary Product Recommendation. In Proc. 29th ACM Int. Conf. Inf. Knowl. Manag. 2517–2524.

• https://www.askul.co.jp/