Sliding Spectrum Decomposition (SSD) for Diversified Recommendation in Re‑ranking

Introduction

The paper "Sliding Spectrum Decomposition for Diversified Recommendation" proposes the SSD model for the re‑ranking stage of feed‑flow recommendation, aiming to balance recommendation quality and diversity by considering both the current sliding window and items outside the window.

Challenges and Innovations

Two main shortcomings of existing methods are identified: (1) traditional DPP models only account for diversity within the current window, ignoring historical content, and (2) embedding vectors derived from sparse collaborative‑filtering data perform poorly for long‑tail items. SSD addresses the first issue by treating the recommendation sequence as a time‑series and applying sliding‑window analysis, while CB2CF tackles the second by learning item embeddings that combine content‑based (CB) and collaborative‑filtering (CF) signals.

SSD Model

SSD formulates the recommendation problem as selecting items that jointly maximize quality and diversity. It constructs a trajectory tensor from the sliding window, applies Singular Spectrum Analysis (SSA) to decompose it, and uses the singular values as orthogonal components representing diversity. The overall objective combines a quality score (e.g., CTR) with a diversity term weighted by a hyper‑parameter.

Greedy Inference

Because exact maximization is NP‑hard, the authors design an efficient greedy inference algorithm that updates orthogonalization results incrementally using a modified Gram‑Schmidt process and a circular queue to handle sliding windows with O(k) time and O(k) space complexity.

CB2CF Framework

CB2CF learns item embeddings by training a siamese network that fuses textual features (via BERT) and visual features (via Inception‑V3). Positive samples are constructed from user‑interacted seed items and their high‑exposure candidates, while negative samples are random items from ItemCF recommendations. Cosine similarity is used to normalize embeddings, and an extra dimension of value 1 is added to align cosine distance with volume‑based diversity measures.

Experiments

Offline experiments compare CF and CB2CF embeddings on long‑tail items, showing superior similarity retrieval for CB2CF. Online A/B tests on the re‑ranking stage compare DPP (control) with SSD/SSD* (treatment) and report significant improvements in both diversity metrics (ILAD, MRT) and quality metrics.

Conclusion

SSD introduces a novel way to incorporate sliding‑window diversity into recommendation re‑ranking, and CB2CF provides robust embeddings for long‑tail items; together they achieve higher diversity and quality than traditional DPP‑based methods.