How to Retrain a Recommender System?

SIGIR 2020, 2020.

Cited by: 0|Bibtex|Views145|Links
Keywords:
recommendation accuracynew user-new itemModel Retrainingstochastic gradient descentmulti-layer perceptronMore(24+)
Weibo:
We study the model retraining mechanism for recommender systems, a topic of high practical values but has been relatively little explored in the research community

Abstract:

Practical recommender systems need be periodically retrained to refresh the model with new interaction data. To pursue high model fidelity, it is usually desirable to retrain the model on both historical and new data, since it can account for both long-term and short-term user preference. However, a full model retraining could be very tim...More

Code:

Data:

0
Introduction
  • Recommender systems play an increasingly important role in the current Web 2.0 era which faces with serious information overload issues.
  • The key technique in a recommender system is the personalization model, which estimates the preference of a user on items based on the historical user-item interactions [14, 33].
  • Since users keep interacting with the system, new interaction data is collected continuously, providing the latest evidence on user preference.
  • It is important to retrain the model with the new interaction data, so as to provide timely personalization and avoid being stale [36].
Highlights
  • Recommender systems play an increasingly important role in the current Web 2.0 era which faces with serious information overload issues
  • We design a neural network-based transfer component, which transforms the old model to a new model that is tailored for future recommendations
  • We propose a new retraining method with two major considerations: (1) building an expressive component that transfers the knowledge gained in previous training to the training on new interactions, and (2) optimizing the transfer component towards the recommendation performance in the near future
  • To demonstrate how our proposed Sequential Meta-Learning (SML) framework works, we provide an implementation based on matrix factorization (MF), a representative embedding model for recommendation
  • We proposed a sequential meta-learning (SML) approach, which consists of 1) an expressive transfer network that converts the previous model to a new model based on the newly collected data, and 2) a sequential training method that effectively utilizes the next-period data to learn the transfer network
  • Our proposed SML achieves the best performance on both datasets, consistently outperforming Full-retrain and the most competitive baseline Caser
  • We will extend our generic training paradigm to a wide range of recommender models, such as the recently emerging graph neural networks [14, 44] that are more effective for collaborative filtering, and factorization machines [13, 32] that can incorporate various side information and sequential recommender models [38]
Methods
  • Full-retrain Fine-tune SPMF GRU4Rec Caser

    SML Full-retrain Fine-tune SPMF GRU4Rec Caser SML

    Recall@10 of SML, which is attributed to the dedicated design of the transfer network and the sequential training algorithm.
  • The authors will implement SML on Caser to see whether combing their advantages can lead to further improvements
Results
  • Evaluation Protocols

    To simulate the real-world scenario that there are typically some historical data to train an initial model, the authors start model retraining from the 10-th and 20-th period of Yelp and Adressa, respectively, using the previous data to train an initial model.
  • The authors perform evaluation at each testing period and report the average scores.
  • As it is time consuming to rank all non-interacted items, the authors sample 999 non-interacted items of a user as the recommendation candidates.
  • The method outputs a ranking list on the 1 interacted item and 999 non-interacted items.
  • The authors adopt two widely-used evaluation metrics: Recall@K and NDCG@K [15] and K is set to 5,10, and 20.
  • For parameters tuning on validation sets, the authors take Recall@20 as the main referential metric
Conclusion
  • The authors investigated the retraining of recommender models.
  • The authors formulated the task of recommender retraining, which aims to achieve good generalization on next-period data by modeling the newly collected data.
  • The authors proposed a sequential meta-learning (SML) approach, which consists of 1) an expressive transfer network that converts the previous model to a new model based on the newly collected data, and 2) a sequential training method that effectively utilizes the next-period data to learn the transfer network.
  • The authors will develop personalized meta-learning mechanisms that optimize the learning process for different users differently
Summary
  • Introduction:

    Recommender systems play an increasingly important role in the current Web 2.0 era which faces with serious information overload issues.
  • The key technique in a recommender system is the personalization model, which estimates the preference of a user on items based on the historical user-item interactions [14, 33].
  • Since users keep interacting with the system, new interaction data is collected continuously, providing the latest evidence on user preference.
  • It is important to retrain the model with the new interaction data, so as to provide timely personalization and avoid being stale [36].
  • Objectives:

    That is, based on all accessible data at the time of retraining and the model parameters of the previous retraining, the authors aim to get a new set of model parameters that can perform well on the near future data Dt+1.
  • The authors aim to utilize the newly collected data Dt only plus the previous model parameters Wt−1, so as to pursue a good retrained model as evaluated on Dt+1.
  • The authors aim to solve the task τt defined in Equation (2) which leverages only the new data Dt to achieve a comparable or even better performance than the full retraining
  • Methods:

    Full-retrain Fine-tune SPMF GRU4Rec Caser

    SML Full-retrain Fine-tune SPMF GRU4Rec Caser SML

    Recall@10 of SML, which is attributed to the dedicated design of the transfer network and the sequential training algorithm.
  • The authors will implement SML on Caser to see whether combing their advantages can lead to further improvements
  • Results:

    Evaluation Protocols

    To simulate the real-world scenario that there are typically some historical data to train an initial model, the authors start model retraining from the 10-th and 20-th period of Yelp and Adressa, respectively, using the previous data to train an initial model.
  • The authors perform evaluation at each testing period and report the average scores.
  • As it is time consuming to rank all non-interacted items, the authors sample 999 non-interacted items of a user as the recommendation candidates.
  • The method outputs a ranking list on the 1 interacted item and 999 non-interacted items.
  • The authors adopt two widely-used evaluation metrics: Recall@K and NDCG@K [15] and K is set to 5,10, and 20.
  • For parameters tuning on validation sets, the authors take Recall@20 as the main referential metric
  • Conclusion:

    The authors investigated the retraining of recommender models.
  • The authors formulated the task of recommender retraining, which aims to achieve good generalization on next-period data by modeling the newly collected data.
  • The authors proposed a sequential meta-learning (SML) approach, which consists of 1) an expressive transfer network that converts the previous model to a new model based on the newly collected data, and 2) a sequential training method that effectively utilizes the next-period data to learn the transfer network.
  • The authors will develop personalized meta-learning mechanisms that optimize the learning process for different users differently
Tables
  • Table1: Average recommendation performance over online testing periods on Adressa and Yelp. “RI” indicates the relative improvement of SML over the corresponding baseline
  • Table2: Retraining time (seconds) at each testing period on Yelp. SML-S is the variant that disables transfer update
Download tables as Excel
Related work
  • 5.1 Recommendation on Sequential Data

    The user-item interaction data naturally forms a sequence because each interaction is associated with timestamp information. A large body of work has modeled a sequence of interactions to predict the next interaction, called as sequential [29], next-item/basket [20, 47] or session-based recommendation [18, 46]. An early representative method is Factorized Personalized Markov Chain (FPMC) [34], which models the transition between an interacted item and the previously interacted item with matrix factorization. Later work has extended the first-order modeling [12] to high-order modeling [45]. Recently, many neural network models have been developed, wherein recurrent neural network (RNN) is a natural choice for sequence modeling [18]. The latest work [20] points out a limitation of RNN that it fails to learn the personalized item frequency information in next-basket recommendation. In addition, CNN has also been used for sequential recommendation [38, 46, 47], which
Funding
  • ∗This work is supported by the National Natural Science Foundation of China (61972372, U19A2079,61725203)
Reference
  • Alex Beutel, Paul Covington, Sagar Jain, Can Xu, Jia Li, Vince Gatto, and Ed H Chi. 2018. Latent cross: Making use of context in recurrent recommender systems. In WSDM. 46–54.
    Google ScholarLocate open access versionFindings
  • Homanga Bharadhwaj. 2019.
    Google ScholarFindings
  • Shiyu Chang, Yang Zhang, Jiliang Tang, Dawei Yin, Yi Chang, Mark A Hasegawa-Johnson, and Thomas S Huang. 2017. Streaming recommender systems. In WWW.
    Google ScholarFindings
  • Yihong Chen, Bei Chen, Xiangnan He, Chen Gao, Yong Li, Jian-Guang Lou, and Yue Wang. 2019. λOpt: Learn to Regularize Recommender Models in Finer Levels. In SIGKDD. 978–986.
    Google ScholarLocate open access versionFindings
  • Robin Devooght, Nicolas Kourtellis, and Amin Mantrach. 201Dynamic matrix factorization with priors on unknown values. In SIGKDD. 189–198.
    Google ScholarLocate open access versionFindings
  • Ernesto Diaz-Aviles, Lucas Drumond, Lars Schmidt-Thieme, and Wolfgang Nejdl.
    Google ScholarFindings
  • 2012. Real-time top-n recommendation in social streams. In RecSys. 59–66.
    Google ScholarFindings
  • [7] Xiaoyu Du, Xiangnan He, Fajie Yuan, Jinhui Tang, Zhiguang Qin, and Tat-Seng Chua. 2019. Modeling Embedding Dimension Correlations via Convolutional
    Google ScholarFindings
  • Neural Collaborative Filtering. TOIS 37, 4 (2019), 47:1–47:22.
    Google ScholarLocate open access versionFindings
  • [8] Zhengxiao Du, Xiaowei Wang, Hongxia Yang, Jingren Zhou, and Jie Tang.
    Google ScholarFindings
  • 2019. Sequential Scenario-Specific Meta Learner for Online Recommendation. In
    Google ScholarLocate open access versionFindings
  • [9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic metalearning for fast adaptation of deep networks. In ICML, Vol. 70. 1126–1135.
    Google ScholarLocate open access versionFindings
  • [10] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. 2018. Bilevel Programming for Hyperparameter Optimization and Meta-
    Google ScholarFindings
  • Learning. In ICML, Vol. 80. 1563–1572.
    Google ScholarFindings
  • [11] Jon Atle Gulla, Lemei Zhang, Peng Liu, Özlem Özgöbek, and Xiaomeng Su. 2017.
    Google ScholarFindings
  • [12] Ruining He, Wang-Cheng Kang, and Julian McAuley. 2017. Translation-based recommendation. In RecSys. 161–169.
    Google ScholarLocate open access versionFindings
  • [13] Xiangnan He and Tat-Seng Chua. 20Neural factorization machines for sparse predictive analytics. In SIGIR. 355–364.
    Google ScholarLocate open access versionFindings
  • [14] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In SIGIR.
    Google ScholarFindings
  • [15] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In WWW. 173–182.
    Google ScholarLocate open access versionFindings
  • [16] Xiangnan He, Hanwang Zhang, Min-Yen Kan, and Tat-Seng Chua. 2016. Fast
    Google ScholarFindings
  • [17] Dan Hendrycks and Kevin Gimpel. 2016. Bridging Nonlinearities and Stochastic Regularizers with Gaussian Error Linear Units. CoRR abs/1606.08415 (2016).
    Findings
  • [18] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk.
    Google ScholarFindings
  • 2016. Session-based Recommendations with Recurrent Neural Networks. In
    Google ScholarLocate open access versionFindings
  • [19] Kurt Hornik. 1991. Approximation capabilities of multilayer feedforward networks. Neural networks 4, 2 (1991), 251–257.
    Google ScholarFindings
  • [20] Haoji Hu, Xiangnan He, Jinyang Gao, and Zhi-Li Zhang. 2020. Modeling
    Google ScholarFindings
  • [21] Muhammad Abdullah Jamal and Guo-Jun Qi. 2019. Task Agnostic Meta-Learning for Few-Shot Learning. In CVPR. 11719–11727.
    Google ScholarFindings
  • [22] Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. In ICLR.
    Google ScholarFindings
  • [23] James Kirkpatrick, Razvan Pascanu, Neil C Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabskabarwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks. PNAS 114, 13 (2017), 3521–3526.
    Google ScholarLocate open access versionFindings
  • [24] Hoyeop Lee, Jinbae Im, Seongwon Jang, Hyunsouk Cho, and Sehee Chung. 2019. MeLU: Meta-Learned User Preference Estimator for Cold-Start Recommendation. In SIGKDD. 1073–1082.
    Google ScholarFindings
  • [25] Wenqiang Lei, Xiangnan He, Yisong Miao, Qingyun Wu, Richang Hong, Min-Yen Kan, and Tat-Seng Chua. 2020. Estimation-Action-Reflection: Towards Deep Interaction Between Conversational and Recommender Systems. In WSDM. 304– 312.
    Google ScholarLocate open access versionFindings
  • [26] Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature interactions for recommender systems. In SIGKDD. 1754–1763.
    Google ScholarFindings
  • [27] David Lopez-Paz and Marc’Aurelio Ranzato. 2017. Gradient Episodic Memory for Continual Learning. In NeurlPS 2017. 6467–6476.
    Google ScholarLocate open access versionFindings
  • [28] Feiyang Pan, Shuokai Li, Xiang Ao, Pingzhong Tang, and Qing He. 2019. Warm Up Cold-start Advertisements: Improving CTR Predictions via Learning to Learn ID Embeddings. In SIGIR. 695–704.
    Google ScholarLocate open access versionFindings
  • [29] Massimo Quadrana, Paolo Cremonesi, and Dietmar Jannach. 2018. Sequenceaware recommender systems. ACM Computing Surveys (CSUR) 51, 4 (2018), 66:1–66:36.
    Google ScholarLocate open access versionFindings
  • [30] Sachin Ravi and Hugo Larochelle. 2017. Optimization as a Model for Few-Shot Learning. In ICLR.
    Google ScholarFindings
  • [31] Yi Ren, Cuirong Chi, and Zhang Jintao. 2019. A Survey of Personalized Recommendation Algorithm Selection Based on Meta-learning. In CSIA. 1383– 1388.
    Google ScholarFindings
  • [32] Steffen Rendle. 2010. Factorization Machines. In ICDM. 995–1000.
    Google ScholarLocate open access versionFindings
  • [33] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2009. BPR: Bayesian personalized ranking from implicit feedback. In UAI. 452– 461.
    Google ScholarFindings
  • [34] Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factorizing personalized markov chains for next-basket recommendation. In WWW. 811–820.
    Google ScholarLocate open access versionFindings
  • [35] Steffen Rendle and Lars Schmidt-Thieme. 2008. Online-updating regularized kernel matrix factorization models for large-scale recommender systems. In RecSys. 251–258.
    Google ScholarLocate open access versionFindings
  • [36] David Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young, Jean-Francois Crespo, and Dan Dennison. 2015. Hidden technical debt in machine learning systems. In NeurlPS. 2503–2511.
    Google ScholarFindings
  • [37] Karthik Subbian, Charu Aggarwal, and Kshiteesh Hegde. 2016. Recommendations for streaming data. In CIKM. 2185–2190.
    Google ScholarFindings
  • [38] Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In WSDM. 565–573.
    Google ScholarLocate open access versionFindings
  • [39] Manasi Vartak, Arvind Thiagarajan, Conrado Miranda, Jeshua Bratman, and Hugo Larochelle. 2017. A meta-learning perspective on cold-start recommendations for items. In NeurlPS. 6904–6914.
    Google ScholarFindings
  • [40] Jeffrey S Vitter. 1985. Random sampling with a reservoir. TOMS 11, 1, 37–57.
    Google ScholarLocate open access versionFindings
  • [41] Qinyong Wang, Hongzhi Yin, Zhiting Hu, Defu Lian, Hao Wang, and Zi Huang. 2018. Neural memory streaming recommender networks with adversarial training. In SIGKDD. 2467–2475.
    Google ScholarFindings
  • [42] Weiqing Wang, Hongzhi Yin, Zi Huang, Qinyong Wang, Xingzhong Du, and Quoc Viet Hung Nguyen. 2018. Streaming ranking based recommender systems. In SIGIR. 525–534.
    Google ScholarLocate open access versionFindings
  • [43] Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat-Seng Chua. 2019. KGAT: Knowledge Graph Attention Network for Recommendation. In SIGKDD. 950–958.
    Google ScholarLocate open access versionFindings
  • [44] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural Graph Collaborative Filtering. In SIGIR. 165–174.
    Google ScholarLocate open access versionFindings
  • [45] Bin Wu, Xiangnan He, Zhongchuan Sun, Liang Chen, and Yangdong Ye. 2019. ATM: An Attentive Translation Model for Next-Item Recommendation. IEEE Transactions on Industrial Informatics 16, 3 (2019), 1448–1459.
    Google ScholarLocate open access versionFindings
  • [46] Fajie Yuan, Xiangnan He, Haochuan Jiang, Guibing Guo, Jian Xiong, Zhezhao Xu, and Yilin Xiong. 2020. Future Data Helps Training: Modeling Future Contexts for Session-based Recommendation. In WWW. 303–313.
    Google ScholarLocate open access versionFindings
  • [47] Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M Jose, and Xiangnan He. 2019. A Simple Convolutional Generative Network for Next Item Recommendation. In WSDM. 582–590.
    Google ScholarLocate open access versionFindings
Your rating :
0

 

Tags
Comments