摘要:结合领域内知识的个性化推荐算法在近年来受到了广泛关注，许多研究工作尝试将商品之间的关系（如互补关系等）融入到推荐算法中.对于商家而言，了解商品互补的关系能够帮助他们更好地制定定价策略；对于推荐算法而言，结合商品关系的推荐也更有可能生成令人满意的结果，因此，如何挖掘商品间的互补关系是一个很有意义的研究问题.现有方法大多从用户历史中的"共同购买"发掘商品的互补关系，但是由于真实的购买场景非常复杂，得到的很可能仅仅是共现关系而不是互补关系.借鉴经济学的相关研究，提出了商品潜在互补性发现推荐模型（latent complementarity discovery model，简称LCDM），试图从另一角度更准确地刻画商品间关系.首先，基于经济学理论中的需求交叉弹性（cross-price elasticity of demand），提出互补性发现模型（complementarity discovery model，简称CDM）联合商品价格与购买历史来挖掘商品间的互补关系.在用户标注任务中，所提算法较已有方法在用户标注的一致性上提升了10.6%.随后，基于此提出了融合商品互补关系的双重注意力机制推荐模型LCDM.最后，在真实数据集上的对比实验结果表明，提出的LCDM推荐模型能够显著改善推荐的效果，在Recall@5和NDCG@5上分别有54.5%和125.8%的提升，验证了所提方法的有效性.



Abstract:Domain-specific personalized recommendation algorithm is getting more popular nowadays. In particularly, item-item relationship (e.g. complementary good, substitute good) has already been considered in the development of recommendation algorithms. In terms of its potential application for sellers, the ability to notice actual item-item complementarity from data is of paramount importance, as it helps sellers to gain a market competitive advantage via designing better pricing strategies (e.g. bundling or pricing discount). For recommender systems, integrating algorithms with the item-item relationship is also more likely to generate better recommendation results. Therefore, how to mine item-item complementarity is a research problem deserving of study. Even though most existing methodologies leverage on co-occurrence relationship, yet, the recommendation accuracy might easily be adversely affected by noise data due to the complex dynamics in the online shopping environment. In light of the research on economics, the latent complementarity discovery model (LCDM) is proposed in an attempt to more accurately describe the item-item relationship from a different perspective. Specifically, complementarity discovery model (CDM) is firstly proposed based on cross-price elasticity of demand, which jointly utilizes item pricing and purchase history to discover item-item complementarity relationship. Comparing with existing mining methods based on item co-occurrence relationship, the proposed method yields 10.6% increase in user label consistency. Then, LCDM is constructed by integrating dual-item attention with item-item complementarity insight mined from CDM. Lastly, from the comparison experiments conducted on real-world dataset, LCDM has made a significant improvement in recommendation performance, in which there is a 54.4% and 125.8% increase in Recall@5 and NDCG@5 respectively.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5925