WSDM'22 Recommendation System Papers

push

recommend

Tie

system

WSDM' 22 has published accepted papers, received 159 papers/submitted 786 papers, and the acceptance rate was 20.23%. For the complete list of accepted papers, see

https://www.wsdm-conference.org/2022/accepted-papers/

@zhang xiaolei, @xiaweimian, etc. have briefly sorted out the papers in [1] [2], and this article is classified and sorted according to the author's ideas. By task, it mainly includes sequence recommendation, session recommendation, CTR estimation, cross-domain recommendation, Debian, path recommendation, collaborative filtering, group recommendation, fairness, E&E, etc.; by technology, it mainly includes comparative learning, adversarial learning, Reinforcement learning, imitation learning, federated learning, ensemble learning, causal inference, meta-learning, graph networks, etc.

Since WSDM'22 has been published very early, there have been many interpretations of the paper, and there are many papers. Due to the length, this paper only sorts out the core contributions of the papers of interest, and then interprets the individual papers of interest.

√

1. Interpretation of the thesis - divided by task

Sequence recommendation

Contrastive Learning for Representation Degeneration Problem in Sequential Recommendation

Paper link: https://arxiv.org/abs/2110.05730

Dissertation institution: University of Queensland

Paper classification: Sequence recommendation, comparative learning

Summary of the paper: This paper studies the representation degradation problem of item embedding matrix in sequence recommendation through empirical observation and theoretical analysis. To address this issue, a novel DuoRec model is proposed that incorporates contrastive regularization while using Dropout-based model-level augmentation and supervised positive sampling to construct contrastive samples. The authors also analyze the properties of this regularizer for the representation degradation problem.

Sequential Modeling with Multiple Attributes for Watchlist Recommendation in E-Commerce

Paper link: https://arxiv.org/abs/2110.11072

Code link: 

https://github.com/urielsinger/Trans2D

Dissertation institutions: eBay Research Institute, Ben-Gurion University

Paper classification: Sequence recommendation

Paper Summary: In this work, we propose a novel watchlist recommendation (WLR) task. The WLR task is a specialized sequence recommendation task that needs to consider a large number of dynamic item attributes during training and prediction. To handle this complex task, we propose Trans2D, an extended Transformer model with a novel self-attention mechanism capable of handling two-dimensional array data (item-attribute) inputs. Trans2D can learn (and retain) complex user preference patterns in a given sequence up to prediction time.

Session recommendation

S-Walk: Accurate and Scalable Session-based Recommendation with Random Walks

Paper link: https://arxiv.org/pdf/2201.01091

Code link: https://github.com/jin530/SWalk

Dissertation institutions: Sungkyunkwan University, Naver Co., Ltd., Google, Seoul National University, etc.

Paper classification: Conversation recommendation

Paper Summary: In this work, we propose a conversational recommendation model using item random walk, called S-Walk. To remedy the shortcomings of existing models, we utilize random walks with restarts to fully capture the internal and interrelationships of sessions, and incorporate an efficient linear item model (transition model and the teleportation model) into the item random walk process.

Heterogeneous Global Graph Neural Networks for Personalized Session-based Recommendation

Paper link: https://arxiv.org/pdf/2107.03813

Code link: https://github.com/0215Arthur/HG-GNN

Dissertation institutions: Tongji University, Jingdong, etc.

Paper Classification: Graph Network, Conversational Recommendation

Paper Summary: In this paper, we propose a heterogeneous global graph neural network for session-based personalized recommendation. Compared with previous methods, we consider the impact of user historical interactions and construct a heterogeneous global graph consisting of historical user-item interactions, item transitions, and global co-occurrence information. Furthermore, we propose a graph-augmented hybrid encoder consisting of a heterogeneous graph neural network and a personalized conversational encoder to comprehensively capture user preference representations.

Learning Multi-granularity Consecutive User Intent Unit for Session-based Recommendation

Paper link:

https://www.atalab.cn/seminar2022Spring/pdf/2022_WSDM_Learning%20Multi-granularity%20Consecutive%20User%20Intent%20Unit%20for%20Session-based%20Recommendation.pdf

Code link: 

https://github.com/SpaceLearner/SessionRec-pytorch

Dissertation institutions: Peking University, Microsoft Asia Research, CMU, etc.

Paper classification: Conversation recommendation

Paper Summary: In this paper, we study the problem of session-based recommendation and propose MSGIFSR, a new model for extracting intent-granularity information across different sessions. It can be observed that different continuous combined intent granularities provide richer user preferences, and our proposed MIHSG successfully captures the relationship between multi-level continuous intent units and long-range dependencies by modeling intra-granularity and inter-granularity intent unit edges. complex preference conversion relationship between. Furthermore, the intent unit encoder mechanism takes into account both order-variant and order-invariant relations for a set of items that are good at expressing the meaning of a schematic unit. Last but not least, ablation studies show that the intent fusion ranking module successfully integrates recommendation results from all intent unit levels.

CTR Estimate

CAN: Feature Co-Action Network for Click-Through Rate Prediction

Paper link: https://arxiv.org/pdf/2011.05625

Code link: 

https://github.com/CAN-Paper/Co-Action-Network

Dissertation institution: Ali, Institute of Automation

Paper Category: CTR Estimation

Paper Summary: In this paper, we highlight the importance of feature interaction modeling, which has not been fully explored by previous work. Inspired by the Cartesian product, we propose a new feature interaction paradigm using a specially designed network, Co-Action Network (CAN). CAN decouples representation learning and feature interaction modeling through a flexible modular collaborative unit. In addition, multi-level augmentation and multi-level independence are introduced in the synergy unit, which further improves the ability of feature interaction modeling.

Triangle Graph Interest Network for Click-through Rate Prediction

Paper link: https://arxiv.org/abs/2202.02698

Code link: https://github.com/alibaba/tgin

Dissertation institutions: Ali, Fudan

Paper classification: graph network, click-through rate estimation

Paper Summary: In this paper, we propose a novel and effective framework, called Triangle Graph Interest Network (TGIN), for the click-through rate prediction task. For each clicked item in the user behavior sequence, we introduce a triangle as a complement in the neighborhood of the item-item graph. TGIN treats these triangles as basic units of user interest, and they provide clues to capture the real motivation of users to click on items. We describe user behavior by aggregating information from several interest units to alleviate the elusive motivation problem. TGIN brings novel and serendipitous items to users by selecting diverse and opposing triangles to break diversity constraints and expand exploration opportunities.

Modeling Users' Contextualized Page-wise Feedback for Click-Through Rate Prediction in E-commerce Search

Paper link: 

https://guyulongcs.github.io/files/WSDM2022_RACP.pdf

Code link: 

https://github.com/racp-submission/racp

Dissertation institutions: Ali, NTU

Paper Category: CTR Estimate, Search

Paper Summary: This paper proposes modeling users' contextual page feedback in e-commerce search ranking for the CTR estimation task. We propose to organize the user's behavior as a series of page feedback and design a RACP model to capture the user's interest. RACP uses the Intra-page Context-aware Interest Layer to extract the user's interest in each page, and uses the Inter-page Interest Back-tracking Layer to capture the user's interest in the session. Dynamic interests, and use the page-level interest aggregation layer (Page-level Interest Aggregation Layer) to aggregate the user's final interest vector.

Cross-domain recommendation

RecGURU: Adversarial Learning of Generalized User Representations for Cross-Domain Recommendation

Paper link: https://arxiv.org/abs/2111.10093

Code link: 

https://github.com/Chain123/RecGURU

Dissertation institutions: University of Alberta, Tencent, Sun Yat-Sen University

Paper Classification: Adversarial Learning, Cross-Domain Recommendation

Personalized Transfer of User Preferences for Cross-domain Recommendation

Paper link: https://arxiv.org/abs/2110.11154

Code link: 

https://github.com/easezyc/WSDM2022-PTUPCDR

Dissertation institutions: Institute of Computing Technology, Chinese Academy of Sciences, Beihang University, WeChat, etc.

Paper classification: Cross-domain recommendation

Multi-Sparse-Domain Collaborative Recommendation via Enhanced Comprehensive Aspect Preference Learning

Paper link: https://arxiv.org/abs/2201.05973

Dissertation institutions: Sichuan University, University of Illinois at Chicago

Paper classification: Cross-domain recommendation

group recommendation

Enumerating Fair Packages for Group Recommendations

Paper link: https://arxiv.org/abs/2105.14423

Code link: https://github.com/joisino/fape

Dissertation institution: Kyoto University, RIKEN AIP

Paper classification: group recommendation, fairness

Paper Summary: In this paper, we study the fair package-to-group recommendation problem. Instead of counting individual packages, we propose to enumerate all fair packages. Although the enumeration problem is computationally challenging, we demonstrate that it is FPT with respect to group size and propose an efficient algorithm based on ZDD. Our proposed algorithm can not only enumerate packages, but also filter items by intersection operation, optimize preferences by linear Boolean programming, and sample packages uniformly and randomly. We experimentally confirm that our proposed method scales to large datasets and can enumerate up to a trillion packages in reasonable time.

Debian

It Is Different When Items Are Older: Debiasing Recommendations When Selection Bias and User Preferences are Dynamic

Paper link: https://arxiv.org/abs/2111.12481

Code link: https://github.com/BetsyHJ/DANCER

Dissertation institutions: University of Amsterdam, University of Nijmegen

Paper classification: Debian

Fighting Mainstream Bias in Recommender Systems via Local Fine Tuning

Paper link: 

http://people.tamu.edu/~zhuziwei/pubs/Ziwei_WSDM_2022.pdf

Code link:

https://github.com/Zziwei/Measuring-Mitigating-Mainstream-Bias

Dissertation institution: Texas A&M University

Paper classification: Debian

Towards Unbiased and Robust Causal Ranking for Recommender Systems

Paper link: 

https://dl.acm.org/doi/abs/10.1145/3488560.3498521

Dissertation institution: Pennsylvania State University

Paper classification: Debian

External Evaluation of Ranking Models under Extreme Position-Bias

Paper link:

https://assets.amazon.science/8b/3a/08b5ec784d2a98f867d506e4c8c1/external-evaluation-of-ranking-models-under-extreme-position-bias.pdf

Paper Institution: Amazon

Paper classification: Debian

Path recommendation

PLdFe-RR: Personalized Long-distance Fuel-efficient Route Recommendation Based On Historical Trajectory

Paper link: 

https://dl.acm.org/doi/abs/10.1145/3488560.3498512

Dissertation institutions: Shandong University, Central South University

Paper classification: Path recommendation

Collaborative filtering

Profiling the Design Space for Graph Neural Networks based Collaborative Filtering

Paper link: http://www.shichuan.org/doc/125.pdf

Code link:

https://github.com/BUPT-GAMMA/Design-Space-for-GNN-based-CF

Dissertation institutions: Beijing Post, Fourth Paradigm

Paper Classification: Graph Network, Collaborative Filtering

Paper Summary: In this work, we present an analysis of the design space of GNN-based CF methods that have been extensively studied in recent years. By overlaying existing GNN-based CF methods in a unified framework, a novel design space is developed and controlled random search is employed to efficiently evaluate the impact of different dimensions on recommendation performance. Furthermore, based on empirical results, design space pruning is performed by excluding some design choices that were shown to be suboptimal in the evaluation. Then empirical studies under different settings demonstrate the high quality and strong generalization ability of the pruned design space. Finally, as a case study, we show that it can quickly obtain the best-performing model architecture on 2 new datasets by randomly searching a pruned design space compared to popular CF methods.

On Sampling Collaborative Filtering Datasets

Paper link: https://arxiv.org/abs/2201.04768

Code link: 

https://github.com/noveens/sampling_cf

Dissertation institutions: University of California, San Diego, Facebook

Paper Classification: Collaborative Filtering

Paper Summary: This paper makes three main contributions: (1) describe the impact of sampling on algorithm performance in terms of algorithm and dataset features (such as sparse features, sequence dynamics, etc.); (2) design SVP-CF, which is a A data-specific sampling strategy designed to preserve the relative performance of the model after sampling, especially for long-tail interactive data; (3) Develop an oracle, Data-genie, that can suggest models that are most likely to preserve a given dataset Sampling scheme for performance.

VAE++: Variational AutoEncoder for Heterogeneous One-Class Collaborative Filtering

Paper link:

https://csse.szu.edu.cn/staff/panwk/publications/Conference-WSDM-22-VAEPlusPlus-Slides.pdf (PPT)

Dissertation institution: Shenzhen University

Paper classification: Collaborative filtering, VAE

Fair recommendation

Enumerating Fair Packages for Group Recommendations

Paper link: https://arxiv.org/abs/2105.14423

Code link: https://github.com/joisino/fape

Dissertation institution: Kyoto University, RIKEN AIP

Paper classification: group recommendation, fair recommendation

Paper Summary: In this paper, we study the fair package-to-group recommendation problem. Instead of counting individual packages, we propose to enumerate all fair packages. Although the enumeration problem is computationally challenging, we demonstrate that it is FPT with respect to group size and propose an efficient algorithm based on ZDD. Our proposed algorithm can not only enumerate packages, but also filter items by intersection operation, optimize preferences by linear Boolean programming, and sample packages uniformly and randomly. We experimentally confirm that our proposed method scales to large datasets and can enumerate up to a trillion packages in reasonable time.

Toward Pareto Efficient Fairness-Utility Trade-off in Recommendation through Reinforcement Learning

Paper link: https://arxiv.org/pdf/2201.00140

Dissertation institution: Rutgers University, Etsy

Paper Classification: Fair Recommendation, Reinforcement Learning

data set

The Datasets Dilemma: How Much Do We Really Know About Recommendation Datasets?

Paper link: 

https://dl.acm.org/doi/abs/10.1145/3488560.3498519

Code link: 

https://github.com/almightyGOSU/TheDatasetsDilemma

Dissertation institution: Nanyang Technological University

Paper classification: Dataset

Paper Summary: Due to the importance and utility of recommender systems, there has been sustained interest in both academia and industry over the years. However, several recent papers have pointed out key issues in the evaluation process of recommender systems. Again, this paper delves into a fundamental but often overlooked aspect of the evaluation process, the dataset itself. To this end, we employ a systematic and comprehensive approach to understanding datasets for implicit feedback-based top-K recommendation. We first check recent papers from top conferences to see how different datasets have been used so far. Next, we look at the characteristics of these datasets to understand their similarities and differences. Finally, we conducted an empirical study to determine whether the choice of dataset used for evaluation affects the observations and/or conclusions obtained. Our findings suggest that more attention needs to be paid to the selection process of datasets used to evaluate recommender systems in order to improve the robustness of the obtained results.

E&E

Show Me the Whole World: Towards Entire Item Space Exploration for Interactive Personalized Recommendations

Paper link:

https://www.microsoft.com/en-us/research/uploads/prod/2021/12/wsdm2022-hcb.pdf

Code link: https://github.com/yusonghust/HCB-pHCB

Dissertation institutions: Huake, Microsoft Asia Research, Meituan

Paper classification: E&E

Paper Summary: In this paper, we propose a general hierarchical bandit machine framework for user interest exploration across the entire space. Specifically, we design two algorithms, namely HCB and pHCB. The HCB algorithm finds the path from the root to the leaf node through a series of decision tasks, while the pHCB gradually expands the receptive field in a top-down manner to explore user interests, which is more flexible and achieves satisfactory results. By fixing the tree structure unchanged, we assume that these items are static in this paper. It would be interesting to extend the proposed framework to a non-static setting, which has not been well studied.

contract advertising

An Adaptive Unified Allocation Framework for Guaranteed Display Advertising

Paper link: 

https://dl.acm.org/doi/abs/10.1145/3488560.3498500

Dissertation institution: Ali, University of Tennessee

Paper Category: Contract Advertising

Abnormal situation consumer demand forecast

Leveraging World Events to Predict E-Commerce Consumer Demand under Anomaly

Paper link: 

https://dl.acm.org/doi/abs/10.1145/3488560.3498452

Dissertation institutions: Israel Institute of Technology, Ben-Gurion University

Paper classification: Prediction of consumer demand in abnormal situations

Scope-aware Re-ranking

Scope-aware Re-ranking with Gated Attention in Feed

Paper link: 

https://dl.acm.org/doi/10.1145/3488560.3498403

Dissertation institution: Ant, University of Science and Technology of China

Paper classification: Scope-aware Re-ranking

√

2. Interpretation of the paper - by technology