Session Recommendation ▪ Deezer Research ▪ ACT-R [Anderson+, 1997](心理学分野の人間の認知プロセスモデル) のDeclative Memory(宣言的記憶モジュール)をTransformerで定式化 ▪ 音楽推薦で重要な「リピートする」行動を自然に定式化 RecSys2024: Best paper candidate ① [Tran+] Declarative Memory Procedural Memory Working Memory Storage Match Retrieval Execution Sensory Register Encoding Information [Anderson+, 1997] Tran, Viet-Anh, et al. "Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
▪ 繰り返し曲だけでなく、新規曲の推薦でも良好な結果 RecSys2024: Best paper candidate ① [Tran+] Tran, Viet-Anh, et al. "Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
et al. "Putting Popularity Bias Mitigation to the Test: A User-Centric Evaluation in Music Recommenders." Proceedings of the 18th ACM Conference on Recommender Systems. 2024. ▪ Putting Popularity Bias Mitigation to the Test: A User-Centric Evaluation in Music Recommenders ▪ Delft University of Technology, Utrecht University ▪ 各種人気バイアス軽減手法のユーザ満足度に対する影響を観察 1 Artist: A Song: A1 2 Artist: A Song: A2 3 Artist: B Song: B1 4 Artist: B Song: B2 5 Artist: C Song: C1 # Artist Song Popularity 1 Artist: A Song: A1 2 Artist: C Song: C1 3 Artist: B Song: B2 4 Artist: A Song: A2 5 Artist: B Song: B1 # Artist Song Popularity
Agent, 東京大学 ▪ マッチングPF等で行われる相互推薦問題において、Envy-freenessの 概念を用いて被推薦機会の公平性を定義 ▪ ナッシュ社会福祉関数を応用し、より公平な相互推薦手法を提案 16 RecSys2024: Best paper candidate ③ [Tomita+] Tomita, Yoji, and Tomohiko Yokoyama. "Fair Reciprocal Recommendation in Matching Markets." Proceedings of the 18th ACM Conference on Recommender Systems. 2024. Unfair !
and Tomohiko Yokoyama. "Fair Reciprocal Recommendation in Matching Markets." Proceedings of the 18th ACM Conference on Recommender Systems. 2024. ▪ ナッシュ社会福祉(NSW) ▪ Frank-Wolfe法を用いて2つのNSWを交互に最大化 ▪ Initialize A, B ▪ Until convergence, do ▪ Update for increasing direction of ▪ ▪ Update for increasing direction of ▪ ▪ Return A, B s.t. constraints to X s.t. constraints to Y
③ [Tomita+] Tomita, Yoji, and Tomohiko Yokoyama. "Fair Reciprocal Recommendation in Matching Markets." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
place (Tom3TK) ▪ 多様な観点の情報を特徴量化したLGBMアンサンブル手法 2nd, 3rd ソリューション概要 Iwai, Tomomu, et al. "Harnessing Temporal Dynamics and Content: An Ensemble of Gradient Boosting Machines for News Recommendation." Proceedings of the Recommender Systems Challenge 2024. 2024. 37-41. Xue, Taofeng, et al. "Large Scale Hierarchical User Interest Modeling for Click-through Rate Prediction." Proceedings of the Recommender Systems Challenge 2024. 2024. 53-57.
Collaborative representations by Large Language Models for Next-Point-of-Interest Recommendations ▪ Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models ▪ LARR: Large Language Model Aided Real-time Scene Recommendation with Semantic Understanding ▪ Playlist Search Reinvented: LLMs Behind the Curtain ▪ A Hybrid Multi-Agent Conversational Recommender System with LLM and Search Engine in E-commerce ▪ Large Language Models as Evaluators for Recommendation Explanation 28 Large Language Models sessions Large Language Modesl session2 ▪ Unleashing the Retrieval Potential of Large Language Models in Conversational Recommender Systems ▪ The Elephant in the Room: Rethinking the Usage of Pre-trained Language Model in Sequential Recommendation ▪ ReLand: Integrating Large Language Models’ Insights into Industrial Recommenders via a Controllable Reasoning Pool ▪ Bayesian Optimization with LLM-Based Acquisition Functions for Natural Language Preference Elicitation ▪ Towards Empathetic Conversational Recommender Systems (Best paper) ▪ FLIP: Fine-grained Alignment between ID-based Models and Pretrained Language Models for CTR Prediction ▪ A Comparative Analysis of Text-Based Explainable Recommender Systems ▪ Reproducibility of LLM-based Recommender Systems: the case study of P5 paradigm
▪ LLM出力文を利用したPOI推薦の課題 ▪ ハルシネーション ▪ 存在しないPOIを推薦してしまう ▪ 曖昧な場所の推薦 ▪ クイーンズ郡を推薦してしまう SeCor: Aligning Semantic and Collaborative Representations by Large Language Models for Next-Point-of-Interest Recommendations Wang, Shirui et al. "SeCor: Aligning Semantic and Collaborative Representations by Large Language Models for Next-Point-of-Interest Recommendations" Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
Language Models for Next-Point-of-Interest Recommendations Wang, Shirui et al. "SeCor: Aligning Semantic and Collaborative Representations by Large Language Models for Next-Point-of-Interest Recommendations" Proceedings of the 18th ACM Conference on Recommender Systems. 2024. ▪ 協調表現と意味表現をLLMで統合して扱うSeCorを提案 ▪ CF埋め込み(LightGCN, DirectAU)をLLMの入力に利用 ▪ 協調表現と意味表現のハイブリッド表現(h_u, h_p)の内積が推薦スコア ▪ LLM出力テキストを利用しないことでハルシネーションを回避
Language Models for Next-Point-of-Interest Recommendations Wang, Shirui et al. "SeCor: Aligning Semantic and Collaborative Representations by Large Language Models for Next-Point-of-Interest Recommendations" Proceedings of the 18th ACM Conference on Recommender Systems. 2024. ▪ 複数のベンチマークで精度向上を確認 ▪ CFやLLMでの既存手法より、それらを統合した提案手法のほうが強い ▪ POIの説明テキストを除く(Fig. 6の w/ UI-token)と精度低下 ▪ SeCorでは、POIの意味的な関係を捉えられていることを示唆
Systems RedialデータセットとCRSの例 ※Redial:映画推薦会話のデータセット ・会話ログ ・推薦アイテム ・推薦アイテムに対する評価 Zhang, Xiaoyu, et al. "Towards empathetic conversational recommender systems." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
Emotion-aware item recommendation 37 global entitiesに対する処理 : local entityに紐づく entity集合 : local entity : entityの共起確率 : global entity表現 Zhang, Xiaoyu, et al. "Towards empathetic conversational recommender systems." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
最適化目標として直接モデル化するのが難しい ▪ ラベルのスパースさ ▪ 消費と比べ、コンテンツ作成に使えるラベルは非常にスパース ▪ 正例の不足は効果的な推薦モデルを訓練する上で大きな障害になる 【YouTube】コンテンツ消費だけではなく作成も考えた推薦:課題 Shao, Yuan, et al. "Optimizing for Participation in Recommender System." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
based Sequential Recommendation model (SASRec) の 損失関数、ネガティブアイテムのサンプリング方法を変更 ▪ 結果 ▪ Popularity Bias (人気の偏り) を減少 ▪ バニラ版と比較してクリック数と視聴量を増加 【ZDF】サンプリングを改善したNext Video推薦:概要 Koneru, Venkata Harshit, et al. "Enhancing Recommendation Quality of the SASRec Model by Mitigating Popularity Bias." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
Entropy (gBCE) loss ▪ logistic sigmoid function を利用。正例アイテムをスケールして popularity bias を改 善する (去年の RecSys の Best paper の内容) [Aleksandr Vladimirovich Petrov and Craig Macdonald. 2023.] ▪ ネガティブアイテムのサンプリング:top-k negative sampling ▪ ネガティブサンプリングした中から、バックプロパゲーションではスコアが高かった top-k だけを利用する [Timo Wilm, Philipp Normann, Sophie Baumeister, and Paul-Vincent Kobow. 2023.] ▪ variant 3:SASRec gBCE neg ▪ variant2 にネガティブアイテムのサンプリングをする際に工夫を加える ▪ batchwise and in-batch sessionwise negative sampling [Timo Wilm, Philipp Normann, Sophie Baumeister, and Paul-Vincent Kobow. 2023.] 【ZDF】サンプリングを改善したNext Video推薦:実験設定 Koneru, Venkata Harshit, et al. "Enhancing Recommendation Quality of the SASRec Model by Mitigating Popularity Bias." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
▪ Palumbo, Enrico, et al. "Graph learning for exploratory query suggestions in an instant search system." Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2023. ▪ ユーザー自身の最近の検索や個人的なアイテム ▪ メタデータや「[アーティスト名] + covers」のような展開ルール ▪ LLM(大規模言語モデル)を活用 ▪ Penha, Gustavo, et al. "Improving content retrievability in search with controllable query generation." Proceedings of the ACM Web Conference 2023. 2023. 【Spotify】クエリの推薦による探索行動の促進:提案クエリの生成 Lindstrom, Henrik, et al. "Encouraging Exploration in Spotify Search through Query Recommendations." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
ユーザーレベルの特徴、消費パターンなど ▪ ラベル:検索結果ページでストリーミング、保存、プレイリスト追加な どの成功アクションにつながったか ▪ ranker を削除すると推奨クエリのクリック率が20%減少 【Spotify】クエリの推薦による探索行動の促進:提案クエリの生成 Lindstrom, Henrik, et al. "Encouraging Exploration in Spotify Search through Query Recommendations." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
and Sudarshan Lamkhede. "Joint Modeling of Search and Recommendations Via an Unified Contextual Recommender (UniCoRn)." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
▪ 性能向上 ▪ 各タスクを同時に考慮することで相互に恩恵を受けられる場面がある (クロスドメインレコメンド) ⇒ 一つのモデル・パイプラインで複数の検索・推薦タスクを解きたい 【Netflix】複数タスクを扱う統一的な推薦モデル:モチベーション Bhattacharya, Moumita, Vito Ostuni, and Sudarshan Lamkhede. "Joint Modeling of Search and Recommendations Via an Unified Contextual Recommender (UniCoRn)." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
e2e で学習 ⇒ オフライン指標で検索タスクと推薦タスクそれぞれ7%と10%の向上 【Netflix】複数タスクを扱う統一的な推薦モデル:段階的な工夫 Bhattacharya, Moumita, Vito Ostuni, and Sudarshan Lamkhede. "Joint Modeling of Search and Recommendations Via an Unified Contextual Recommender (UniCoRn)." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.