Temporally and Distributionally Robust Optimization for Cold-Start Recommendation

TL;DR

This work tackles cold-start item recommendation under temporal feature shifts by identifying limitations of standard DRO when distributions evolve over time. It introduces Temporal DRO (TDRO), which combines a worst-case factor to raise the lower bound on performance with a shifting-factor to track temporal trends in item features, using a gradient-based approximation for efficient group selection across $K$ item groups and $E$ time periods. Gradient smoothing stabilizes training via a group-weighting scheme and a streaming loss update, and TDRO is instantiated on strong backbones such as CLCRec and GAR, demonstrating improved generalization for cold-start items across three real-world datasets. The approach yields robust performance under varying shift strengths, and the authors provide open-source code to facilitate adoption and further research in temporally robust recommender systems.

Abstract

Collaborative Filtering (CF) recommender models highly depend on user-item interactions to learn CF representations, thus falling short of recommending cold-start items. To address this issue, prior studies mainly introduce item features (e.g., thumbnails) for cold-start item recommendation. They learn a feature extractor on warm-start items to align feature representations with interactions, and then leverage the feature extractor to extract the feature representations of cold-start items for interaction prediction. Unfortunately, the features of cold-start items, especially the popular ones, tend to diverge from those of warm-start ones due to temporal feature shifts, preventing the feature extractor from accurately learning feature representations of cold-start items. To alleviate the impact of temporal feature shifts, we consider using Distributionally Robust Optimization (DRO) to enhance the generation ability of the feature extractor. Nonetheless, existing DRO methods face an inconsistency issue: the worse-case warm-start items emphasized during DRO training might not align well with the cold-start item distribution. To capture the temporal feature shifts and combat this inconsistency issue, we propose a novel temporal DRO with new optimization objectives, namely, 1) to integrate a worst-case factor to improve the worst-case performance, and 2) to devise a shifting factor to capture the shifting trend of item features and enhance the optimization of the potentially popular groups in cold-start items. Substantial experiments on three real-world datasets validate the superiority of our temporal DRO in enhancing the generalization ability of cold-start recommender models. The code is available at https://github.com/Linxyhaha/TDRO/.

Figures (5)

• Figure 1: (a) An example of item category feature shifts towards sanitary products. (b) T-SNE visualization of visual features of item thumbnails in three time periods on a Micro-video dataset. The stars represent the average item features in each time period. (c) An example of the shifting trend of three item groups over time. (d) Illustration of the inconsistency issue of DRO.
• Figure 2: Illustration of the shifting factor with three groups and three time periods (i.e., $i\in\{1,2,3\}$ and $e\in\{1,2,3\}$). (a) depicts the three steps of obtaining the weighted period gradient in each time period. And then, by summing up the weighted period gradient, we can obtain the shifting trend as shown in (b). Finally, the shifting factor for each group is obtained by calculating the similarity between the group gradient and the shifting trend as presented in (c).
• Figure 3: Effect of period importance steepness control factor $p$ for period importance and period number $E$.
• Figure 4: Effect of the strength of shifting factor $\lambda$.
• Figure 5: Effect of group number $K$.