OrDA Purifies Marketing Recommendations by Removing Habit Bias

Homepage marketing block recommendations are often influenced by a combination of a user's genuine content interest and their ingrained access habits, such as clicking on items in prominent positions. This can lead to "pseudo-positives," where items are clicked due to their placement rather than their quality, biasing the recommendation system. To address this, researchers propose OrDA, the Orthogonal Disentanglement of Access habits framework. OrDA employs a dual-tower neural network architecture with a gated allocation layer, designed to adaptively route features and minimize interference between interest and habit signals. A key innovation is the use of orthogonal regularization, which ensures that the latent representations of interest and habit are geometrically perpendicular, enforcing a rigorous separation. During inference, OrDA applies causal intervention (do-calculus) to rank items based solely on these purified interest scores, effectively removing the influence of access habits. Empirical online evaluations on large-scale datasets confirm that OrDA successfully eliminates access-habit bias, outperforming state-of-the-art methods in predictive accuracy. An online A/B test showed a 5.64% improvement in user click-through rates on a real-world marketing block, demonstrating its practical impact.