New Pareto-DQN framework breaks filter bubbles with multi-objective AI
Recommender systems can now boost diversity and fairness without sacrificing engagement.
Recommender systems today often trap users in filter bubbles by optimizing only for immediate engagement. A new paper from Cláudio Lúcio Do Val Lopes and colleagues proposes a multi-objective reinforcement learning framework called Semantic Pareto-DQN to solve this. It formalizes recommendation as a semantic multi-objective Markov decision process, using high-fidelity semantic embeddings alongside a Pareto-DQN agent. Instead of blending engagement, diversity, and fairness into one scalar reward, it treats them as distinct signals, avoiding the pitfalls of static reward scalarization. The agent selects actions based on hypervolume contribution, effectively mapping the Pareto frontier.
Testing on the MovieLens small dataset shows the approach sustains high state-trajectory variance, breaking the feedback loops that cause semantic collapse. The model achieves substantial improvements in auxiliary societal objectives like information diversity and provider fairness, while only marginally impacting engagement metrics. This work, accepted at IEEE International Conference on Responsible Artificial Intelligence (IRAI) 2026, offers a concrete path toward intrinsically aligned recommendation systems that balance business retention with broader societal values.
- Treats engagement, diversity, and fairness as separate non-aggregable reward signals using a Pareto-DQN agent.
- Uses semantic embeddings and hypervolume-based action selection to explore the Pareto frontier on MovieLens.
- Achieves significant gains in diversity and fairness with only marginal impact on user engagement.
Why It Matters
Provides a practical framework for building recommendation systems that balance profit with social responsibility.