AI helps you reading Science

AI generates interpretation videos

AI extracts and analyses the key points of the paper to generate videos automatically


pub
Go Generating

AI Traceability

AI parses the academic lineage of this thesis


Master Reading Tree
Generate MRT

AI Insight

AI extracts a summary of this paper


Weibo:
The promising results from our experiments show that Recurrent Neural Networks allow to take a novel perspective on applications such as Recommender Systems that were originally designed in a time-agnostic manner

Sequential User-based Recurrent Neural Network Recommendations

RecSys, pp.152-160, (2017)

Cited by: 87|Views44
EI

Abstract

Recurrent Neural Networks are powerful tools for modeling sequences. They are flexibly extensible and can incorporate various kinds of information including temporal order. These properties make them well suited for generating sequential recommendations. In this paper, we extend Recurrent Neural Networks by considering unique characterist...More

Code:

Data:

0
Introduction
  • Sequence modeling has overall been less explored in RS research due to the long-lasting focus on time-agnostic models.
  • This gap may be explained by the unique nature of the recommendation problem.
  • Heavily influenced by Machine Learning [41], many modern techniques successfully applied in other domains cannot be transfered to RS.
  • Even if they consider consumption sequences [13, 18, 19, 48, 53, 54], they often have other limitations
Highlights
  • The majority of today’s Recommender Systems (RS) [41] relies on algorithms that are designed under the assumption

    Compared to other areas, sequence modeling has overall been less explored in RS research due to the long-lasting focus on time-agnostic models
  • With the rise of Deep Learning (DL) in the past decade, Recurrent Neural Networks (RNN) have become practical and powerful tools for large-scale supervised learning of sequences. This progress has become most apparent in Natural Language Processing where they have set several new benchmarks outperforming techniques that have been considered state-of-the-art for a long time
  • The promising results from our experiments show that RNNs allow to take a novel perspective on applications such as RS that were originally designed in a time-agnostic manner
  • The way we formulate RNNs enables us to model user behavior and to capture dependencies between consumption events more adequately than with established recommendation techniques that often fail at appropriately representing temporal dynamics in user interests
  • Gated recurrent networks have set records in accuracy on many tasks in recent years. These advances result from novel or extended architectures rather than from fundamentally novel algorithms [31, 45]. This applies to our user-based RNNs: We essentially adopted the original architecture to take the unique characteristics of the recommendation domain into account
  • Including a user-specific layer, the networks seem capable of learning concepts that are unique to a certain user
Methods
  • The authors describe the evaluation metrics and datasets the authors used, as well as the algorithmic setup.

    4.1.1 Metrics.
  • The authors describe the evaluation metrics and datasets the authors used, as well as the algorithmic setup.
  • Based on temporally ordered lists of consumed items, the objective is to correctly predict the item a target user will likely consume.
  • The ground truth at a particular time step is represented by a single user-item tuple.
  • To present the user with adequate recommendations, the target item should be among the first few recommended items.
  • In accordance with recent RS research, the authors use the following evaluation metrics:.
Results
  • On the MovieLens dataset, there is a substantial gain in objective performance compared to the best baseline.
  • There is an improvement of 12 % in MRR@20 and 23 % in Recall@20, respectively, when comparing exponentially decaying item-based -NN to the standard GRU.
  • The integration variant that performs best in terms of MRR@20, i.e. attentional on the MovieLens and rectified linear on the LastMF dataset, leads to an improvement of 32 % or 33 %, respectively.
  • For Recall@20, there is similar gain with results 30 % better on the MovieLens, and 28 % better on the LastFM dataset
Conclusion
  • The results indicate that RNNs clearly outperform other recommending approaches when it comes to generating sequential recommendations.
  • As earlier experiments by others suggested [e.g. 18, 48], even out-of-the-box RNNs achieve superior results with respect to widely used evaluation metrics when compared to state-of-the-art item-based -NN or MF approaches
  • This is true for their derivatives extended to consider temporal effects.
  • With the rise of DL in the past decade, RNNs have become practical and powerful tools for large-scale supervised learning of sequences
  • This progress has become most apparent in Natural Language Processing where they have set several new benchmarks outperforming techniques that have been considered state-of-the-art for a long time.
  • The networks can distinguish between user preferences more accurately
Summary
  • Introduction:

    Sequence modeling has overall been less explored in RS research due to the long-lasting focus on time-agnostic models.
  • This gap may be explained by the unique nature of the recommendation problem.
  • Heavily influenced by Machine Learning [41], many modern techniques successfully applied in other domains cannot be transfered to RS.
  • Even if they consider consumption sequences [13, 18, 19, 48, 53, 54], they often have other limitations
  • Objectives:

    Based on temporally ordered lists of consumed items, the objective is to correctly predict the item a target user will likely consume.
  • Methods:

    The authors describe the evaluation metrics and datasets the authors used, as well as the algorithmic setup.

    4.1.1 Metrics.
  • The authors describe the evaluation metrics and datasets the authors used, as well as the algorithmic setup.
  • Based on temporally ordered lists of consumed items, the objective is to correctly predict the item a target user will likely consume.
  • The ground truth at a particular time step is represented by a single user-item tuple.
  • To present the user with adequate recommendations, the target item should be among the first few recommended items.
  • In accordance with recent RS research, the authors use the following evaluation metrics:.
  • Results:

    On the MovieLens dataset, there is a substantial gain in objective performance compared to the best baseline.
  • There is an improvement of 12 % in MRR@20 and 23 % in Recall@20, respectively, when comparing exponentially decaying item-based -NN to the standard GRU.
  • The integration variant that performs best in terms of MRR@20, i.e. attentional on the MovieLens and rectified linear on the LastMF dataset, leads to an improvement of 32 % or 33 %, respectively.
  • For Recall@20, there is similar gain with results 30 % better on the MovieLens, and 28 % better on the LastFM dataset
  • Conclusion:

    The results indicate that RNNs clearly outperform other recommending approaches when it comes to generating sequential recommendations.
  • As earlier experiments by others suggested [e.g. 18, 48], even out-of-the-box RNNs achieve superior results with respect to widely used evaluation metrics when compared to state-of-the-art item-based -NN or MF approaches
  • This is true for their derivatives extended to consider temporal effects.
  • With the rise of DL in the past decade, RNNs have become practical and powerful tools for large-scale supervised learning of sequences
  • This progress has become most apparent in Natural Language Processing where they have set several new benchmarks outperforming techniques that have been considered state-of-the-art for a long time.
  • The networks can distinguish between user preferences more accurately
Tables
  • Table1: Hyperparameter values for all algorithms used in our experiments on the two datasets
  • Table2: MRR@20 and Recall@20 for all baselines as well as our proposed user-based RNN variants
Download tables as Excel
Related work
  • User representation in RS is often designed as a statistically stationary process where every expressed preference is assumed to be fixed over time [29, 50]. However, relational data in real-world scenarios are often evolving and exhibit strong temporal patterns [55]. Time may therefore be considered an important contextual dimension also in RS [4]. Generating recommendations in a time-independent manner may in contrast result in estimations of preferences based on user-item relations that are no longer valid. This lack of adaptability with respect to one of the most natural properties of user behavior has motivated several approaches that integrate temporal dynamics1. In the following, we discuss the most important ones, and especially those most closely related to ours that also exploit DL techniques as a means to generate sequential recommendations.
Funding
  • For Recall@20, there is similar gain with results 30 % better on the MovieLens, and 28 % better on the LastFM dataset
Study subjects and analysis
users: 71567
We ran our approach as well as all the baselines on the following two real-world datasets:. ∙ MovieLens 10M: The MovieLens 10M dataset5 consists of 10 000 054 ratings assigned to 10 681 movies by 71 567 users. In order to mimic implicit data, we binarized all ratings independent of their value, considering them as positive feedback as it has been done in [e.g. 39]

users: 992
∙ LastFM 1K Users: The LastFM 1K Users dataset6 contains user-timestamp-artist-song tuples collected via the LastFM API. The dataset has a total of 19 150 868 data points for 992 users. Due to computational reasons we performed our evaluation on a 10 % subsample

Reference
  • L. Baltrunas and X. Amatriain. 2009. Towards time-dependant recommendation based on implicit feedback. In CARS ’09.
    Google ScholarFindings
  • Y. Bengio, P. Frasconi, and P. Simard. 1993. The problem of learning long-term dependencies in recurrent networks. In ICNN ’93. IEEE, 1183–1188.
    Google ScholarLocate open access versionFindings
  • G. Bonnin and D. Jannach. 2015. Automated generation of music playlists: Survey and experiments. ACM Comput Surv 47, 2 (2015), 26:1–26:35.
    Google ScholarLocate open access versionFindings
  • P. G. Campos, F. Díez, and I. Cantador. 201Time-aware recommender systems: A comprehensive survey and analysis of existing evaluation protocols. User Model User-Adap 24, 1-2 (2014), 67–119.
    Google ScholarLocate open access versionFindings
  • K. Cho, B. Van Merriënboer, D. Bahdanau, and Y. Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In SSST ’14.
    Google ScholarFindings
  • J. Chung, C. Gulcehre, K. Cho, and Y. Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. In Deep Learning Workshop at NIPS ’14.
    Google ScholarLocate open access versionFindings
  • Y. Ding and X. Li. 2005. Time weight collaborative filtering. In CIKM ’05. ACM, 485–492.
    Google ScholarLocate open access versionFindings
  • S. El Hihi and Y. Bengio. 1995. Hierarchical recurrent neural networks for long-term dependencies. In NIPS ’95. 493–499.
    Google ScholarLocate open access versionFindings
  • F. Gers. 2001. Long short-term memory in recurrent neural networks. Ph.D. Dissertation. University of Hannover.
    Google ScholarFindings
  • I. Goodfellow, Y. Bengio, and A. Courville. 2016. Deep learning. MIT Press.
    Google ScholarFindings
  • A. Graves. 2013. Generating sequences with recurrent neural networks. (2013).
    Google ScholarFindings
  • A. Graves, A. Mohamed, and G. Hinton. 2013. Speech recognition with deep recurrent neural networks. In ICASSP ’13. IEEE, 6645– 6649.
    Google ScholarLocate open access versionFindings
  • A. Greenstein-Messica, L. Rokach, and M. Friedman. 2017. Session-based recommendations using item embedding. In IUI ’17. ACM, 629–633.
    Google ScholarLocate open access versionFindings
  • A. Gunawardana and G. Shani. 2009. A survey of accuracy evaluation metrics of recommendation tasks. J Mach Learn Res 10 (2009), 2935–2962.
    Google ScholarLocate open access versionFindings
  • K. He, X. Zhang, S. Ren, and J. Sun. 20Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV ’15. 1026–1034.
    Google ScholarLocate open access versionFindings
  • K. He, X. Zhang, S. Ren, and J. Sun. 20Deep residual learning for image recognition. In CVPR ’16. IEEE, 770–778.
    Google ScholarLocate open access versionFindings
  • C. Hermann. 2010. Time-based recommendations for lecture materials. In EdMedia ’10. 1028–1033.
    Google ScholarFindings
  • B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk. 2015. Session-based recommendations with recurrent neural networks. In ICLR ’16.
    Google ScholarFindings
  • B. Hidasi, M. Quadrana, A. Karatzoglou, and D. Tikk. 2016. Parallel recurrent neural network architectures for feature-rich session-based recommendations. In RecSys ’16. ACM, 241–248.
    Google ScholarLocate open access versionFindings
  • S. Hochreiter and J. Schmidhuber. 1997. Long short-term memory. Neural Comput 9, 8 (1997), 1735–1780.
    Google ScholarLocate open access versionFindings
  • T. Hofmann. 2001. Unsupervised learning by probabilistic latent semantic analysis. Mach Learn 42, 1-2 (2001), 177–196.
    Google ScholarLocate open access versionFindings
  • T. Hussein, T. Linder, W. Gaulke, and J. Ziegler. 2014. Hybreed: A software framework for developing context-aware hybrid recommender systems. User Model User-Adap 24, 1-2 (2014), 121–174.
    Google ScholarLocate open access versionFindings
  • A. Karatzoglou. 2011. Collaborative temporal order modeling. In RecSys ’11. ACM, 313–316.
    Google ScholarLocate open access versionFindings
  • A. Karatzoglou, X. Amatriain, L. Baltrunas, and N. Oliver. 2010. Multiverse recommendation: N-dimensional tensor factorization for context-aware collaborative filtering. In RecSys ’10. ACM, 79–86.
    Google ScholarLocate open access versionFindings
  • D. Kingma and J. Ba. 2015. Adam: A method for stochastic optimization. In ICLR ’15.
    Google ScholarLocate open access versionFindings
  • R. Klinkenberg and T. Joachims. 2000. Detecting concept drift with support vector machines. In ICML ’00. Morgan Kaufmann, 487–494.
    Google ScholarFindings
  • Y. Koren. 2010. Collaborative filtering with temporal dynamics. Commun ACM 53, 4 (2010), 89–97.
    Google ScholarLocate open access versionFindings
  • Y. Koren and R. Bell. 2015. Recommender systems handbook. Springer, Chapter Advances in collaborative filtering, 77–118.
    Google ScholarLocate open access versionFindings
  • Y. Koren, R. Bell, and C. Volinsky. 2009. Matrix factorization techniques for recommender systems. IEEE Computer 42, 8 (2009), 30–37.
    Google ScholarLocate open access versionFindings
  • A. Krizhevsky, I. Sutskever, and G. E. Hinton. 2012. Imagenet classification with deep convolutional neural networks. In NIPS ’12. 1097–1105.
    Google ScholarLocate open access versionFindings
  • Z. C. Lipton, J. Berkowitz, and C. Elkan. 2015. A critical review of recurrent neural networks for sequence learning. (2015).
    Google ScholarFindings
  • T. Mikolov, K. Chen, G. Corrado, and J. Dean. 2013. Efficient estimation of word representations in vector space. Workshop at ICLR ’13.
    Google ScholarLocate open access versionFindings
  • T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean. 2013. Distributed representations of words and phrases and their compositionality. In NIPS ’13. 3111–3119.
    Google ScholarLocate open access versionFindings
  • T. Mikolov and G. Zweig. 2012. Context dependent recurrent neural network language model. In SLT ’12. 234–239.
    Google ScholarLocate open access versionFindings
  • O. Nasraoui, J. Cerwinske, C. Rojas, and F. Gonzalez. 2007. Performance of recommendation systems in dynamic streaming environments. In SDM ’07. SIAM, 569–574.
    Google ScholarLocate open access versionFindings
  • R. Pascanu, C. Gulcehre, K. Cho, and Y. Bengio. 2014. How to construct deep recurrent neural networks. In ICLR ’14.
    Google ScholarFindings
  • D. M. Powers. 2011. Evaluation: From precision, recall and Fmeasure to ROC, informedness, markedness and correlation. J Mach Learn Tech 2, 1 (2011), 37–63.
    Google ScholarLocate open access versionFindings
  • S. Rallapalli, L. Qiu, Y. Zhang, and Y.-C. Chen. 2010. Exploiting temporal stability and low-rank structure for localization in mobile networks. In MobiCom ’10. ACM, 161–172.
    Google ScholarLocate open access versionFindings
  • S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme. 2009. BPR: Bayesian personalized ranking from implicit feedback. In UAI ’09. AUAI Press, 452–461.
    Google ScholarFindings
  • S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme. 2010. Factorizing personalized Markov chains for next-basket recommendation. In WWW ’10. ACM, 811–820.
    Google ScholarLocate open access versionFindings
  • F. Ricci, L. Rokach, and B. Shapira. 2015. Recommender systems handbook (2nd ed.). Springer.
    Google ScholarFindings
  • N. Sahoo, P. V. Singh, and T. Mukhopadhyay. 2012. A hidden Markov model for collaborative filtering. MIS Quarterly 36, 4 (2012), 1329–1356.
    Google ScholarLocate open access versionFindings
  • R. Salakhutdinov, A. Mnih, and G. E. Hinton. 2007. Restricted Boltzmann machines for collaborative filtering. In ICML ’07. ACM, 791–798.
    Google ScholarLocate open access versionFindings
  • J. Schmidhuber. 1992. Learning complex, extended sequences using the principle of history compression. Neural Comput 4, 2 (1992), 234–242.
    Google ScholarFindings
  • J. Schmidhuber. 2015. Deep learning in neural networks: An overview. Neural Networks 61 (2015), 85–117.
    Google ScholarLocate open access versionFindings
  • H. Soh, S. Sanner, M. White, and G. Jamieson. 2017. Deep sequential recommendation for personalized adaptive user interfaces. In IUI ’17. ACM, 589–593.
    Google ScholarLocate open access versionFindings
  • C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. 2015. Going deeper with convolutions. In CVPR ’15. IEEE, 1–9.
    Google ScholarLocate open access versionFindings
  • Y. K. Tan, X. Xu, and Y. Liu. 2016. Improved recurrent neural networks for session-based recommendations. In DLRS ’16. ACM, 17–22.
    Google ScholarLocate open access versionFindings
  • A. van den Oord, S. Dieleman, and B. Schrauwen. 2013. Deep content-based music recommendation. In NIPS ’13. 2643–2651.
    Google ScholarLocate open access versionFindings
  • J. Vinagre, A. M. Jorge, and J. Gama. 2015. An overview on the exploitation of time in collaborative filtering. Data Min Knowl Disc 5, 5 (2015), 195–215.
    Google ScholarLocate open access versionFindings
  • E. M. Voorhees. 1999. The TREC-8 question answering track report. In TREC ’99. 77–82.
    Google ScholarLocate open access versionFindings
  • H. Wang, N. Wang, and D.-Y. Yeung. 2015. Collaborative deep learning for recommender systems. In KDD ’15. ACM, 1235–1244.
    Google ScholarLocate open access versionFindings
  • C. Wu, J. Wang, J. Liu, and W. Liu. 2016. Recurrent neural network based recommendation for time heterogenous feedback. Knowl-Based Syst 109 (2016), 90–103.
    Google ScholarLocate open access versionFindings
  • C.-Y. Wu, A. Ahmed, A. Beutel, A. J. Smola, and H. Jing. 2017. Recurrent recommender networks. In WSDM ’17. ACM, 495–503.
    Google ScholarLocate open access versionFindings
  • L. Xiong, X. Chen, T.-K. Huang, J. Schneider, and J. Carbonell. 2010. Temporal collaborative filtering with bayesian probabilistic tensor factorization. In SDM ’10. SIAM, 211–222.
    Google ScholarLocate open access versionFindings
  • Z. Xu, C. Chen, T. Lukasiewicz, Y. Miao, and X. Meng. 2016. Tag-aware personalized recommendation using a deep-semantic similarity model with negative sampling. In CIKM ’16. ACM, 1921–1924.
    Google ScholarFindings
  • H.-F. Yu, N. Rao, and I. S. Dhillon. 2015. High-dimensional time series prediction with missing values. (2015).
    Google ScholarFindings
  • W. Zaremba, I. Sutskever, and O. Vinyals. 2014. Recurrent neural network regularization. (2014).
    Google ScholarFindings
Author
Your rating :
0

 

Tags
Comments
小科