Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
162 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Sliding Spectrum Decomposition for Diversified Recommendation (2107.05204v1)

Published 12 Jul 2021 in cs.IR and cs.LG

Abstract: Content feed, a type of product that recommends a sequence of items for users to browse and engage with, has gained tremendous popularity among social media platforms. In this paper, we propose to study the diversity problem in such a scenario from an item sequence perspective using time series analysis techniques. We derive a method called sliding spectrum decomposition (SSD) that captures users' perception of diversity in browsing a long item sequence. We also share our experiences in designing and implementing a suitable item embedding method for accurate similarity measurement under long tail effect. Combined together, they are now fully implemented and deployed in Xiaohongshu App's production recommender system that serves the main Explore Feed product for tens of millions of users every day. We demonstrate the effectiveness and efficiency of the method through theoretical analysis, offline experiments and online A/B tests.

Citations (32)

Summary

  • The paper introduces SSD, a novel method that models item sequences as time series to capture comprehensive diversity in recommendations.
  • The approach employs singular value decomposition on trajectory tensors to integrate out-of-window items into diversity assessment.
  • Empirical tests on social media data confirm that SSD improves engagement metrics and diversity measurements over traditional methods.

Sliding Spectrum Decomposition for Diversified Recommendation

The paper "Sliding Spectrum Decomposition for Diversified Recommendation" proposes a novel method aimed at enhancing the diversity in recommendation systems, specifically within the context of content feeds commonly found on social media platforms. This research addresses the challenge of maintaining diverse recommendations by introducing a technique called Sliding Spectrum Decomposition (SSD) which leverages time series analysis to handle long sequences of recommended items.

In the domain of recommender systems, balancing relevance with diversity is crucial to enhance user satisfaction and engagement. Traditional approaches often apply sliding windows for achieving diversity in recommender systems. However, such methods typically overlook out-of-window items, thereby failing to capture the full perception of diversity from a user's perspective. This paper suggests approaching this problem from an item sequence perspective by treating the sequence as a time series. By deriving SSD, the authors stack multiple sliding windows into a trajectory tensor, enabling the decomposition of the tensor through singular value decomposition (SVD) to assess diversity based on the entire item sequence.

The SSD method defines the diversity of a sequence of items using the volume of the tensors, calculated as the product of singular values derived from the SVD of the trajectory tensor. This perspective is novel as it integrates the influence of out-of-window items into the diversity assessment, aligning more closely with the user experience on feed platforms.

To complement SSD, this research underscores the significance of appropriate item embeddings for similarity measurement, especially given the challenges posed by the long tail effect prevalent in social media contexts. They propose a Content-Based to Collaborative Filtering (CB2CF) strategy using a siamese network that combines both textual and visual features to strengthen the similarity signals and alleviate issues related to sparse user engagements.

For empirical validation, theoretical analysis is supplemented by offline experiments and extensive online A/B testing conducted on the Xiaohongshu app. The results demonstrate that SSD, and its refined variant SSD*, provide measurable improvements over existing methods like DPP (Determinantal Point Processes) in metrics such as time spent on the platform, user engagements, and diversity measurements like ILAD (Intra-List Average Distance) and MRT (Mean Read Taxonomies).

From a practical standpoint, the research enhances the capabilities of recommender systems to deliver diverse recommendations without compromising on computational efficiency—an essential requirement for real-time systems handling large-scale data. Theoretically, this work contributes to the understanding of diversity in recommendation systems by bridging the gap between short-term sequence diversity and long-term user content interaction.

In anticipating future research developments, this paper suggests further exploration into more sophisticated tensor decomposition methods that could offer deeper insights into complex user behaviors and preferences. Additionally, expanding the application of these techniques to other domains beyond social media feeds could reveal their utility in a broader range of recommendation scenarios. This work paves the way for extending similar methodologies into more generalized frameworks for personalized and contextually aware recommendation systems in future AI developments.