Brand building increasingly depends on the capacity of firms to detect, interpret, and operationalize dynamic consumer preference signals distributed across heterogeneous digital environments. Recent work by Dong introduced a cross-domain temporal preference prediction (CDTPP) model that integrates social-behavior features and e-commerce purchase records through a factorization-machine ranking architecture, reporting strong benchmark performance on a screened overlapping-user dataset. Specifically, the source study reports an average area under the curve (AUC) of 0.953, an average accuracy of 0.984, a fit of 98.87%, and an average brand-preference-index error of 0.11.
Despite this strong predictive baseline, an important managerial limitation remains: the original model is effective as a predictor but comparatively limited as an interpretive decision-support tool. In particular, it does not explicitly identify which semantic themes in social discourse are associated with changes in brand preference, nor does it provide an explanation layer that management teams can use for strategic planning, campaign refinement, or early risk detection. To address this gap, this paper develops ET-CDTPP, an explainable, topic-aware extension of CDTPP that augments the original temporal social feature space with topic-mixture and sentiment representations derived from social text. The proposed framework preserves the original cross-domain and temporal logic while adding semantic interpretability, time-aligned topic aggregation, and a topic-contribution explanation layer.
This article is framed as a conceptual and methodological extension, not as a substitute for a newly re-estimated benchmark study. It therefore uses the original CDTPP study as its empirical benchmark context and does not claim unsupported performance gains for the proposed extension. Instead, it contributes a rigorous model formulation, a reproducible implementation pathway, and a benchmark-anchored validation protocol that can be applied to the same class of overlapping social and e-commerce datasets. This positioning preserves empirical discipline while offering a clearer methodological contribution for management-oriented journal audiences.