Yinqi Zhang, Xintian Han, Haolong Li, Kedi Chen, Shaohui Lin. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers). 2025.
We examine counterfactual explanations for explaining the decisions made by model-based AI systems. The counterfactual approach we consider defines an explanation as a set of the system's data inputs that causally drives the decision (i.e., changing the inputs in the set changes the decision) and is irreducible (i.e., changing any subset of the inputs does not change the decision). We (1) demonstrate how this framework may be used to provide explanations for decisions made by general, data-driven AI systems that may incorporate features with arbitrary data types and multiple predictive models, and (2) propose a heuristic procedure to find the most useful explanations depending on the context. We then contrast counterfactual explanations with methods that explain model predictions by weighting features according to their importance (e.g., SHAP, LIME) and present two fundamental reasons why we should carefully consider whether importance-weight explanations are well-suited to explain system decisions. Specifically, we show that (i) features that have a large importance weight for a model prediction may not affect the corresponding decision, and (ii) importance weights are insufficient to communicate whether and how features influence decisions. We demonstrate this with several concise examples and three detailed case studies that compare the counterfactual approach with SHAP to illustrate various conditions under which counterfactual explanations explain data-driven decisions better than importance weights.
While scaling laws for recommendation models have gained significant traction, existing architectures such as Wukong, HiFormer, and DHEN often struggle with sub-optimal designs and hardware under-utilization, limiting their practical scalability. Our previous TokenMixer architecture (introduced in the RankMixer paper) balanced effectiveness and efficiency by replacing self-attention with a lightweight token mixing operator; however, it encountered critical bottlenecks in deeper configurations, including sub-optimal residual paths, vanishing gradients, incomplete MoE sparsification, and constrained scalability. In this paper, we propose TokenMixer-Large, a systematically evolved architecture designed for extreme-scale recommendation. By introducing a mixing-and-reverting operation, inter-layer residuals, and an auxiliary loss, we ensure stable gradient propagation as model depth increases. Furthermore, we incorporate a Sparse Per-token MoE to enable efficient parameter expansion. TokenMixer-Large successfully scales to 7 billion parameters for online traffic and 15 billion in offline experiments. Currently deployed in multiple scenarios at ByteDance, TokenMixer-Large has achieved substantial performance gains: delivering a 1.66% increase in orders and a 2.98% boost in per-capita preview payment GMV for e-commerce; improving ADSS by 2.0% in advertising; and achieving 1.4% revenue growth in live streaming.