Collaborative Autoencoders for Top-K Hierarchical Recommendations
Authors: Dora Jambor and Putra Manggala
Upcoming poster presentation at WiML'18.
Abstract from the paper:In real-world recommender systems, items can often be organized into hierarchical taxonomies where different levels of the taxonomies represent different levels of information. In Koenigstein et al. (2011), flattened representations of hierarchical information are used to mitigate cold-start problems, while in He et al. (2016), explicit hierarchical representations are used as recommender system inputs.
In this work, we consider a novel problem of top-K hierarchical recommendations, where we seek to produce top-K recommendations for each level of the taxonomy. This is motivated by real-world use cases, where e-commerce sellers need to produce recommendations at different granularities, for example, a department store website might seek to produce top-K recommendations for clothing brands (Gap vs Nike), for types of clothing (jeans vs shirt) within a brand, for styles (skinny vs boot cut), and for colors (black or blue). In this example, brands, clothing types, styles and colors are different levels of a hierarchical taxonomy. Providing these top-K lists simultaneously allows the website to appeal to different abstraction level of the buyers’ interests via diversity [Castells et al. (2015)].
We propose a variation of autoencoder-based collaborative filtering recommender system that directly incorporates a pre-defined items’ taxonomy structure into the output label space of our model as hierarchical output layers using maximum/average pooling, similar to the work in Xiong and Manggala (2018). This ensures that the reconstruction loss function of the autoencoder is aware of the errors within the hierarchical structure of the output. The hierarchical output layers enable us to obtain predictions for every element at every level of the taxonomy simultaneously, and to output top-K recommendations for every level of the taxonomy that are consistent with respect to the taxonomy. Furthermore, the losses from the output layers are combined using a convex combination, and the combination parameters reflect the importances of the different taxonomy levels. This allows the recommender system designer with a prior information on the taxonomy level importances to customize the loss function via the convex combination parameters.
Experiments were conducted on a dataset that contains top-K recommendations ground truths at different levels of an item taxonomy. Our method’s performance is evaluated against a few baselines using the convex combination of precision@k’s (with taxonomy level importances as parameters) for top-K recommendation across all levels. Our variation is shown to significantly outperform other baselines, and is guaranteed to be consistent with respect to the taxonomy.
Pablo Castells, Neil J Hurley, and Saul Vargas. 2015. Novelty and diversity in recommender systems. In Recommender Systems Handbook. Springer, 881–918.
Ruining He, Chunbin Lin, Jianguo Wang, and Julian McAuley. 2016. Sherlock: sparse hierarchical embeddings for visually-aware one-class collaborative filtering. IJCAI Proceedings of the Twenty- Fifth International Joint Conference on Artificial Intelligence (2016).
Noam Koenigstein, Gideon Dror, and Yehuda Koren. 2011. Yahoo! music recommendations: modeling music ratings with temporal dynamics and item taxonomy. In Proceedings of the fifth ACM conference on Recommender systems. ACM, 165–172.
Tengke Xiong and Putra Manggala. 2018. Hierarchical Classification with Hierarchical Attention Networks. KDD Deep Learning Day (2018).