Hierarchical Portfolios in Recommender Ecosystems: Multi-level aggregations of heterogeneous ensembles integrating long-term and short-term
Abstract
This paper focuses on the recommendation problem from the perspective of hierarchical heterogeneous portfolios. These portfolios consist of several base recommenders.
Each of them processes different subset of available data and achieves the best performance under different circumstances. By running base recommenders in parallel and employing a suitable aggregation of their results (i.e. ensemble approach) considerable performance gains can be achieved.
The main contribution of this paper is the proposal of a hierarchical ensemble approach to the recommendation problem and its utilization in the case of repeated recommendations. We extend flat portfolios to hierarchical ones with two levels of aggregation.
For the aggregation of base recommenders, we experimented with Thompson sampling multi-armed bandits algorithm and a modified version of D'Hondt's mandates allocation algorithm. As for the hierarchical aggregation, we implemented a modified version of D21-Janecek mandate allocation algorithm, which allows us to incorporate implicit negative feedback as well.
Experiments were performed on real-world data from the domain of e-commerce. Hierarchical portfolios outperformed both flat portfolios as well as individual base recommenders w.r.t. click-through rate.