Introduction
Modern recommendation systems are increasingly interactive, learning from user feedback in real-time. However, they frequently encounter two significant challenges: the accuracy-diversity dilemma and implicit feedback bias. Standard algorithms often prioritize immediate clicks (accuracy), leading to repetitive recommendations that trap users in "filter bubbles." Furthermore, real-world data is often implicit (e.g., clicks vs. non-clicks), where a lack of interaction doesn't necessarily imply dislike but could result from a lack of exposure.
This paper introduces DC²B (Diversified Contextual Combinatorial Bandit), a framework designed to balance relevance with variety while mathematically accounting for exposure bias in implicit feedback.
Methodology: The DC²B Framework
The DC²B model treats recommendation as a sequential game within a contextual bandit framework. Its core innovation lies in the integration of three key components:
- Determinantal Point Process (DPP): Used to select a "super arm" (a set of items). Unlike greedy selection, DPP considers both the individual quality of items and their mutual dissimilarity, ensuring a diverse slate.
- Thompson Sampling with Variational Inference: To learn model parameters from binary click data, the authors use Thompson Sampling. They address the non-conjugate nature of logistic likelihood using Variational Bayesian inference, providing a closed-form approximation for belief updates.
- Exposure Modeling: DC²B explicitly models the probability that a set of items is exposed to a user. This allows the system to distinguish between a user's lack of interest and a lack of visibility due to redundancy.
A critical feature of the model is the α parameter, which serves as a control knob for the trade-off between precision and diversity.
Experimental Evaluation
The authors evaluated DC²B against several baselines (LogRank, MMR, ε-Greedy, DPP$^{map}$, C²UCB, and EC-Bandit) using three benchmark datasets: MovieLens-100K, MovieLens-1M, and a large, sparse Anime dataset. High ratings were treated as "clicks," and item embeddings were used as contextual features.
Key Results
DC²B consistently demonstrated superior performance across all datasets, particularly in balancing competing metrics.
- Optimal Balance: DC²B achieved the highest F-measure (the harmonic mean of precision and diversity) in nearly all test cases.
- Performance in Sparse Data: On the Anime dataset, DC²B outperformed the C²UCB baseline by 59.35% in Precision@50, highlighting its effectiveness in data-sparse environments.
- Stability: Combinatorial methods like DC²B, which update based on the feedback of the entire displayed set, proved more stable than single-item update methods.
Conclusion and Future Directions
DC²B effectively addresses the accuracy-diversity trade-off in interactive recommendation. While the current implementation uses a greedy approximation for DPP and pre-learned embeddings, future research could explore end-to-end representation learning within the bandit framework and more sophisticated DPP kernels to further refine recommendation nuance.
Summary updated