The framework, developed by Khadijeh Rahimkhani and Kamran Zamanifar, rests on four coordinated components. The first is a recency-aware user state encoder that captures how preferences evolve, weighting recent interactions more heavily than older ones so the model reflects a user’s current tastes rather than a frozen historical profile.

The second component propagates social signals through what the authors call dual-channel signed graph attention. Because real social networks contain both trust and distrust edges, the model treats positive and negative influence as separate channels, and a structural-balance regularizer shapes the learned representation of the signed network during training.

The third element tackles the confounding problem directly. Users tend to befriend people who already share their tastes, so a naive model cannot tell whether a recommendation works because of genuine social influence or simply because similar people cluster together. CT-IDR applies inverse-propensity scoring inside the attention mechanism, reweighting social messages to deconfound this homophily effect and recover a more causal estimate of influence.

Finally, a diversity-aware re-ranking head lets operators tune the trade-off between accuracy and exposure variety. Rather than treating diversity as an afterthought, the head adjusts the final ranked list so that coverage and novelty can be controlled without a separate system.

All four parts are trained end-to-end with a composite objective that couples prediction, balance, and diversity terms, meaning the model optimizes the whole stack simultaneously instead of stitching together independently trained modules.

The evaluation used two benchmark datasets with explicit social signals: Epinions, which contains both trust and distrust links, and Ciao, which has trust links only, with distrust inferred. The authors used temporal leave-last-one-out splits, meaning each user’s most recent interaction was held out for testing in a way that respects time order, and ranking was performed over the full item catalog rather than a filtered candidate set.

CT-IDR set new lows in rating error, with RMSE and MAE of 0.925 and 0.710 on Epinions and 0.955 and 0.730 on Ciao. In top-K ranking, it reached HR@10, HR@20, and MRR values of 0.405, 0.535, and 0.275 on Epinions, and 0.360, 0.507, and 0.250 on Ciao.

Notably, the system also matched or closely tracked a dedicated diversity specialist. On Epinions it achieved Coverage@10 and Coverage@20 of 0.277 and 0.360, with intra-list diversity of 0.58 and 0.63 at the same cutoffs; on Ciao the corresponding figures were 0.24 and 0.34 for coverage and 0.55 and 0.59 for diversity, all while retaining clear margins in hit rate and MRR over the comparison methods.

Ablation experiments, in which individual components are removed to measure their contribution, showed that temporal encoding, signed propagation, and inverse-propensity debiasing drive the majority of the ranking gains. The re-ranking head, by contrast, supplies coverage and novelty without harming accuracy at its chosen operating point.

The authors argue that CT-IDR shifts the accuracy-diversity frontier outward by integrating temporal, signed, and causal reasoning into one controllable recommendation stack. They position the framework as a practical recipe for responsible, user-centric recommendation in signed social environments, suited to large-scale platforms and datasets, though the reported results come from benchmark evaluations on Epinions and Ciao rather than live production deployments.

Subject of Research: Causal Temporal Influence-Diversity Recommendation (CT-IDR) framework for social recommender systems

Article Title: A unified, causal recommender for temporal preferences, signed influence, and diversity

Article References: Rahimkhani, K., & Zamanifar, K. (2026). A unified, causal recommender for temporal preferences, signed influence, and diversity. Cluster Computing, 29(15), Article 880. https://doi.org/10.1007/s10586-026-06671-z

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06671-z

Keywords: recommender systems, causal inference, social networks, machine learning, diversity, graph attention, inverse-propensity scoring, temporal preferences, unified, causal, recommender, temporal