Personalisation of d'Hondt's algorithm and its use in recommender ecosystems

TL;DR

This paper addresses personalization in recommender ecosystems where voting-based aggregators like $d'Hondt$ must balance global signals with per-user preferences. It proposes a hierarchical hybrid portfolio combining global and per-user weight models with online multiplicative weights updates, evaluated in sequential online experiments on RetailRocket and SLANTour. Findings show that full per-user personalization increases relevance and minority voices, while hybrid models can outperform non-personalised baselines depending on dataset characteristics (RetailRocket vs SLANTour). The work demonstrates that per-user data-model level personalisation can improve CTR and fairness, and points to extensions to more domains and improvements for cold-start; code is available.

Abstract

In the area of recommender systems, we are dealing with aggregations and potential of personalisation in ecosystems. Personalisation is based on separate aggregation models for each user. This approach reveals differences in user preferences, especially when they are in strict disagreement with global preferences. Hybrid models are based on combination of global and personalised model of weights for d'Hondt's voting algorithm. This paper shows that personalisation combined with hybridisation on case-by-case basis outperforms non-personalised d'Hondt's algorithm on datasets RetailRocket and SLANTour. By taking into account voices of minorities we achieved better click through rate.

Figures (2)

• Figure 1: Hybrid aggregation model. It aggregates personalised and global model of votes based on threshold changing in time.
• Figure 2: Weight development of model of votes.