Q&R: A Two-Stage Approach toward Interactive Recommendation

    KDD, pp. 139-148, 2018.

    Cited by: 0|Bibtex|Views37|Links
    EI
    Keywords:
    Long Short Term Memoryrecommender systemMean Average PrecisionRecurrent Neural Networksuser onboardingMore(6+)
    Wei bo:
    We observe that our Q&R Topic Recurrent Neural Networks outperforms the Multiclass-BOW by 8.07%; this demonstrates that using an Recurrent Neural Networks unit to capture the sequential nature of the data allows the model to learn more sophisticated representations compared to us...

    Abstract:

    Recommendation systems, prevalent in many applications, aim to surface to users the right content at the right time. Recently, researchers have aspired to develop conversational systems that offer seamless interactions with users, more effectively eliciting user preferences and offering better recommendations. Taking a step towards this g...More

    Code:

    Data:

    0
    Introduction
    • Recommendation systems play a key role for assisting users to navigate through the vast amount of information available by selecting for them the right item, i.e., product to buy, content to read, video to watch, at the right time [3].

      Recently, recommendation researchers and practitioners have aspired to advance the frontier of recommendation by building conversational recommenders in order to create seamless interactions with the users.
    • Measure of Usefulness # selected topics notification-opens watch-time people give recommendations – they try to quickly understand user preferences by asking a few questions under a certain context, and give a recommendation based on the responses [12]
    • They should predict the user’s potentially evolving, unarticulated interests, while accounting for the fact that users might have a biased view of the world [42].
    • Conversational recommenders, while aiming at recommendation, focus on balancing the exploreexploit tradeoff present in recommender systems [12, 47]
    Highlights
    • Recommendation systems play a key role for assisting users to navigate through the vast amount of information available by selecting for them the right item, i.e., product to buy, content to read, video to watch, at the right time [3].

      Recently, recommendation researchers and practitioners have aspired to advance the frontier of recommendation by building conversational recommenders in order to create seamless interactions with the users
    • Our two-factored video recommendation approach can surface to users more interesting videos to watch, even on top of complex, stateof-the-art Recurrent Neural Networks recommenders, both in YouTube Homepage and in YouTube Notifications (Table 1)
    • Since we study the single round of a conversation between a user and the system, our work can be viewed in the larger context of conversational recommender systems
    • We report Mean Average Precision@20 results, noting that Mean Average Precision@1 results follow similar trends
    • We observe that our Q&R Topic Recurrent Neural Networks outperforms the Multiclass-BOW by 8.07%; this demonstrates that using an Recurrent Neural Networks unit to capture the sequential nature of the data allows the model to learn more sophisticated representations compared to using the bag of word features
    • Users become 18% more likely to complete the User Onboarding experience, and when they do, the numbers of topics they select goes up by 77.7%
    Methods
    • Design Goals

      While designing Q&R, the main goal is the improvement of user experience after the inclusion of the system.
    • Other aspects the authors consider for the system design are: scalability, to scale well with a large pool of items for recommendation, temporal patterns based on user sequence data and freshness of results and conversations to keep track of newly generated content.
    • (1) Question Generation (Section 5.2): a deep sequential network that predicts which topics will interest the user.
    • Based on a triggering mechanism, either the question generation or the item recommendation module is used.
    Results
    • Live Traffic Results on YouTube

      The authors show that Q&R can enhance the user experience in multiple applications within YouTube, highlighting the broad impact of the approach (Section 6).
    • After including the two-factored approach in the pool of nominator models for video recommendations, the authors observed that on average the models result in a 0.07% improvement in the time spent watching videos, compared to the production baseline – a highly optimized baseline, including RNNs, that is hard to compete with.
    • After adding the two-factored approach in the pool of nominators in the experiment, the authors observed that on average the models result in a 1.23% improvement in terms of the number of users who open the recommended video notifications, compared to the production baseline.
    • The authors performed an A/B test, where in A the authors used the Naive Bayes and in B the authors employed the recurrent solution
    Conclusion
    • To the best of the knowledge, this is the first work on learned interactive recommendation demonstrated in a large-scale industrial setting.

      The authors' work brings attention to the often overlooked problem of bootstrapping conversations based on interactions from a traditional system.
    • To the best of the knowledge, this is the first work on learned interactive recommendation demonstrated in a large-scale industrial setting.
    • The authors' work brings attention to the often overlooked problem of bootstrapping conversations based on interactions from a traditional system.
    • In building Q&R, the authors set out to improve the user experience of casual users in YouTube.
    • The authors provide a novel neural-based recommendation approach, which factorizes video recommendation to a two-fold problem: user history-to-topic, and topic& user history-to-video.
    • The authors demonstrate the value of the approach for both the YouTube Homepage and YouTube Notifications
    Summary
    • Introduction:

      Recommendation systems play a key role for assisting users to navigate through the vast amount of information available by selecting for them the right item, i.e., product to buy, content to read, video to watch, at the right time [3].

      Recently, recommendation researchers and practitioners have aspired to advance the frontier of recommendation by building conversational recommenders in order to create seamless interactions with the users.
    • Measure of Usefulness # selected topics notification-opens watch-time people give recommendations – they try to quickly understand user preferences by asking a few questions under a certain context, and give a recommendation based on the responses [12]
    • They should predict the user’s potentially evolving, unarticulated interests, while accounting for the fact that users might have a biased view of the world [42].
    • Conversational recommenders, while aiming at recommendation, focus on balancing the exploreexploit tradeoff present in recommender systems [12, 47]
    • Methods:

      Design Goals

      While designing Q&R, the main goal is the improvement of user experience after the inclusion of the system.
    • Other aspects the authors consider for the system design are: scalability, to scale well with a large pool of items for recommendation, temporal patterns based on user sequence data and freshness of results and conversations to keep track of newly generated content.
    • (1) Question Generation (Section 5.2): a deep sequential network that predicts which topics will interest the user.
    • Based on a triggering mechanism, either the question generation or the item recommendation module is used.
    • Results:

      Live Traffic Results on YouTube

      The authors show that Q&R can enhance the user experience in multiple applications within YouTube, highlighting the broad impact of the approach (Section 6).
    • After including the two-factored approach in the pool of nominator models for video recommendations, the authors observed that on average the models result in a 0.07% improvement in the time spent watching videos, compared to the production baseline – a highly optimized baseline, including RNNs, that is hard to compete with.
    • After adding the two-factored approach in the pool of nominators in the experiment, the authors observed that on average the models result in a 1.23% improvement in terms of the number of users who open the recommended video notifications, compared to the production baseline.
    • The authors performed an A/B test, where in A the authors used the Naive Bayes and in B the authors employed the recurrent solution
    • Conclusion:

      To the best of the knowledge, this is the first work on learned interactive recommendation demonstrated in a large-scale industrial setting.

      The authors' work brings attention to the often overlooked problem of bootstrapping conversations based on interactions from a traditional system.
    • To the best of the knowledge, this is the first work on learned interactive recommendation demonstrated in a large-scale industrial setting.
    • The authors' work brings attention to the often overlooked problem of bootstrapping conversations based on interactions from a traditional system.
    • In building Q&R, the authors set out to improve the user experience of casual users in YouTube.
    • The authors provide a novel neural-based recommendation approach, which factorizes video recommendation to a two-fold problem: user history-to-topic, and topic& user history-to-video.
    • The authors demonstrate the value of the approach for both the YouTube Homepage and YouTube Notifications
    Tables
    • Table1: Q&R leads to a better understanding of user preferences and as a result better user experience in multiple YouTube applications
    • Table2: Relationship with Conversational Recommenders: Q&R is the first system to bootstrap conversations based on largescale user data, demonstrated in live traffic, aimed at new and casual users, and using neural sequence-based models
    • Table3: Live metric % improvement of Q&R Topic RNN vs. topic ranking baseline in User Onboarding
    Download tables as Excel
    Related work
    • Since we study the single round of a conversation between a user and the system, our work can be viewed in the larger context of conversational recommender systems.

      The need to present recommendations in a conversational manner [14, 28] has been studied from many perspectives, including interview-based [41], active-learning [39], entropy [47], picturebased [36], explore-exploit [47], critiquing [10], constraint [16], dialog [8], and utility-based strategies [33]. We refer the reader to [18, 22, 38] for a literature review. Here, we compare a few notable works to our system; a comparison overview is found in Table 2.

      Many conversational recommendation works have focused on balancing the trade-off among exploring the space of user preferences vs exploiting what has been learned thus far [12, 47]; this is a complementary question to our work. Regarding the underlying model, existing works use either latent-factor [12] or regressionbased models [2]. However, such models have been shown to be outperformed by deeper models when large-scale data is available [13]; thus, we build upon deep Recurrent Neural Networks (RNNs). For the space of the questions, most systems ask questions on the end-recommendation items [12, 47]; this is though not possible for domains where the item pool is large and constantly updated, e.g. videos. This is why we ask questions on topics, so as to more effectively propagate feedback among videos sharing the same topic. In terms of the feedback elicited, existing systems typically utilize absolute or relative questions or comparisons among two sets of items in a series of questions [31]. Instead, we use a top-N list setting from which the user selects the topics they are interested in [17]. Also, while most existing conversational systems have as target users cold-start users [12, 47], we show that our system can improve the experience of existing users as well.
    Funding
    • Our results demonstrate that our approach improves upon state-of-the-art recommendation models, including RNNs, and makes these applications more useful, such as a > 1% increase in video notifications opened
    • Casual users become 18% more likely to complete the User Onboarding experience, and when they do, they select 77.7% more topics
    • A by-product of this approach is that not only we make advances to the conversational domain, but we also improve the state-of-the-art in (a) traditional video recommendation, and (b) traditional question (e.g. topic) recommendation
    • We hypothesize that to achieve this, we need improved accuracy in: (1) question ranking quality, and (2) response relevance after user’s feedback and we test this hypothesis with our experiments
    • Users become 18% more likely to complete the User Onboarding experience, and when they do, the numbers of topics they select goes up by 77.7%
    Study subjects and analysis
    people: 2
    For the purposes of this work, we consider a single round of a conversation. We use as inspiration how an actual effective conversation between two people would go: Person A asks a question, or gives a prompt, and then person B replies to this question. Then to complete a full round of conversation, person A is supposed to (1) understand what has been said by person B, relevant to what A

    Reference
    • Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. 2016. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467 (2016).
      Findings
    • Deepak Agarwal and Bee-Chung Chen. 2009. Regression-based latent factor models. In KDD. 19–28.
      Google ScholarLocate open access versionFindings
    • Xavier Amatriain. 2012. Building industrial-scale real-world recommender systems. In RecSys. 7–8.
      Google ScholarLocate open access versionFindings
    • George Anders. 2017. Alexa, Understand Me. https://www.technologyreview.com/s/608571/alexa-understand-me/. Accessed:2017-10-25.
      Findings
    • Ricardo A Baeza-Yates, Carlos A Hurtado, Marcelo Mendoza, et al. 2004. Query Recommendation Using Query Logs in Search Engines. In EDBT workshops, Vol. 3268. 588–596.
      Google ScholarLocate open access versionFindings
    • 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
    • Ágnes Bogárdi-Mészöly, András Rövid, Hiroshi Ishikawa, Shohei Yokoyama, and Zoltán Vámossy. 2013. Tag and topic recommendation systems. Acta Polytechnica Hungarica 10, 6 (2013), 171–191.
      Google ScholarLocate open access versionFindings
    • Derek G Bridge. 2002. Towards Conversational Recommender Systems: A Dialogue Grammar Approach. In ECCBR Workshops. 9–22.
      Google ScholarLocate open access versionFindings
    • Cheng Cao, Hancheng Ge, Haokai Lu, Xia Hu, and James Caverlee. 2017. What Are You Known For?: Learning User Topical Profiles with Implicit and Explicit Footprints. In SIGIR. 743–752.
      Google ScholarLocate open access versionFindings
    • Li Chen and Pearl Pu. 2012. Critiquing-based recommenders: survey and emerging trends. User Modeling and User-Adapted Interaction 22, 1 (2012), 125–150.
      Google ScholarLocate open access versionFindings
    • Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014).
      Findings
    • Konstantina Christakopoulou, Filip Radlinski, and Katja Hofmann. 2016. Towards Conversational Recommender Systems. In KDD. 815–824.
      Google ScholarLocate open access versionFindings
    • Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. In RecSys. 191–198.
      Google ScholarFindings
    • Ingemar J Cox, Matt L Miller, Thomas P Minka, Thomas V Papathomas, and Peter N Yianilos. 2000. The Bayesian image retrieval system, PicHunter: theory, implementation, and psychophysical experiments. IEEE transactions on image processing 9, 1 (2000), 20–37.
      Google ScholarLocate open access versionFindings
    • James Davidson, Benjamin Liebald, Junning Liu, Palash Nandy, Taylor Van Vleet, Ullas Gargi, Sujoy Gupta, Yu He, Mike Lambert, Blake Livingston, et al. 2010. The YouTube video recommendation system. In RecSys. ACM, 293–296.
      Google ScholarFindings
    • Alexander Felfernig, Gerhard Friedrich, Dietmar Jannach, and Markus Zanker. 2011. Developing constraint-based recommenders. In Recommender Systems Handbook. Springer, 187–215.
      Google ScholarLocate open access versionFindings
    • Mark P Graus and Martijn C Willemsen. 2015. Improving the user experience during cold start through choice-based preference elicitation. In RecSys. 273–276.
      Google ScholarFindings
    • Chen He, Denis Parra, and Katrien Verbert. 2016. Interactive recommender systems: A survey of the state of the art and future research challenges and opportunities. Expert Systems with Applications 56 (2016), 9–27.
      Google ScholarLocate open access versionFindings
    • Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015).
      Findings
    • Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation 9, 8 (1997), 1735–1780.
      Google ScholarLocate open access versionFindings
    • How Jing and Alexander J Smola. 2017. Neural survival recommender. In WSDM. 515–524.
      Google ScholarLocate open access versionFindings
    • Michael Jugovac and Dietmar Jannach. 2017. Interacting with recommenders – overview and research directions. TiiS 7, 3, 10.
      Google ScholarLocate open access versionFindings
    • Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Greg Corrado, László Lukács, Marina Ganea, Peter Young, et al. 2016. Smart reply: Automated response suggestion for email. In KDD. 955–964.
      Google ScholarFindings
    • Yehuda Koren. 2010. Collaborative filtering with temporal dynamics. Commun. ACM 53, 4 (2010), 89–97.
      Google ScholarLocate open access versionFindings
    • Ralf Krestel and Peter Fankhauser. 2009. Tag recommendation using probabilistic topic models. ECML PKDD Discovery Challenge 2009 (2009), 131.
      Google ScholarLocate open access versionFindings
    • Huayu Li, Martin Renqiang Min, Yong Ge, and Asim Kadav. 2017. A Contextaware Attention Network for Interactive Question Answering. In KDD. 927–935.
      Google ScholarLocate open access versionFindings
    • Huizhi Liang, Yue Xu, Dian Tjondronegoro, and Peter Christen. 2012. Time-aware topic recommendation based on micro-blogs. In CIKM. 1657–1661.
      Google ScholarFindings
    • Greg Linden, Steve Hanks, and Neal Lesh. 1997. Interactive assessment of user preference models: The automated travel assistant. In User Modeling. Springer, 67–78.
      Google ScholarFindings
    • Dong Liu, Xian-Sheng Hua, Linjun Yang, Meng Wang, and Hong-Jiang Zhang. 2009. Tag ranking. In WWW. 351–360.
      Google ScholarLocate open access versionFindings
    • Jiahui Liu, Peter Dolan, and Elin Rønby Pedersen. 2010. Personalized news recommendation based on click behavior. In IUI. 31–40.
      Google ScholarLocate open access versionFindings
    • Benedikt Loepp, Tim Hussein, and Jüergen Ziegler. 2014. Choice-based preference elicitation for collaborative filtering recommender systems. In CHI. 3085–3094.
      Google ScholarFindings
    • Zhongqi Lu, Zhicheng Dou, Jianxun Lian, Xing Xie, and Qiang Yang. 2015. Content-Based Collaborative Filtering for News Topic Recommendation. In AAAI. 217–223.
      Google ScholarLocate open access versionFindings
    • Tariq Mahmood and Francesco Ricci. 2009. Improving recommender systems with adaptive conversational strategies. In HT. 73–82.
      Google ScholarLocate open access versionFindings
    • Anirban Majumder and Nisheeth Shrivastava. 2013. Know your personalization: learning topic level personalization in online services. In WWW. 873–884.
      Google ScholarLocate open access versionFindings
    • Nicolaas Matthijs and Filip Radlinski. 2011. Personalizing web search using long term browsing history. In WSDM. 25–34.
      Google ScholarLocate open access versionFindings
    • Julia Neidhardt, Rainer Schuster, Leonhard Seyfang, and Hannes Werthner. 2014. Eliciting the users’ unknown preferences. In RecSys. 309–312.
      Google ScholarFindings
    • Sanjay Purushotham, Yan Liu, and C-C Jay Kuo. 2012. Collaborative topic regression with social matrix factorization for recommendation systems. arXiv preprint arXiv:1206.4684 (2012).
      Findings
    • Filip Radlinski and Nick Craswell. 2017. A theoretical framework for conversational search. In CHIIR. 117–126.
      Google ScholarLocate open access versionFindings
    • Neil Rubens, Dain Kaplan, and Masashi Sugiyama. 2015. Active Learning in Recommender Systems. Recommender Systems Handbook, 809–846.
      Google ScholarLocate open access versionFindings
    • Börkur Sigurbjörnsson and Roelof Van Zwol. 2008. Flickr tag recommendation based on collective knowledge. In WWW. 327–336.
      Google ScholarLocate open access versionFindings
    • Mingxuan Sun, Fuxin Li, Joonseok Lee, Ke Zhou, Guy Lebanon, and Hongyuan Zha. 2013. Learning multiple-question decision trees for cold-start recommendation. In WSDM. 445–454.
      Google ScholarLocate open access versionFindings
    • Yu Sun, Nicholas Jing Yuan, Yingzi Wang, Xing Xie, Kieran McDonald, and Rui Zhang. 2016. Contextual intent tracking for personal assistants. In KDD. 273–282.
      Google ScholarLocate open access versionFindings
    • George Toderici, Hrishikesh Aradhye, Marius Pasca, Luciano Sbaiz, and Jay Yagnik. 2010. Finding meaning on youtube: Tag recommendation and category discovery. In CVPR. 3447–3454.
      Google ScholarFindings
    • Oriol Vinyals and Quoc Le. 2015. A neural conversational model. arXiv preprint arXiv:1506.05869 (2015).
      Findings
    • Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J Smola, and How Jing. 2017. Recurrent recommender networks. In WSDM. 495–503.
      Google ScholarLocate open access versionFindings
    • Xiao Yu, Xiang Ren, Yizhou Sun, Quanquan Gu, Bradley Sturt, Urvashi Khandelwal, Brandon Norick, and Jiawei Han. 2014. Personalized entity recommendation: A heterogeneous information network approach. In WSDM. 283–292.
      Google ScholarLocate open access versionFindings
    • Xiaoxue Zhao, Weinan Zhang, and Jun Wang. 2013. Interactive collaborative filtering. In CIKM. 1411–1420.
      Google ScholarFindings
    • Zhe Zhao, Zhiyuan Cheng, Lichan Hong, and Ed H Chi. 2015. Improving user topic interest profiles by behavior factorization. In WWW. 1406–1416.
      Google ScholarFindings
    • Cai-Nicolas Ziegler, Sean M McNee, Joseph A Konstan, and Georg Lausen. 2005. Improving recommendation lists through topic diversification. In WWW. 22–32.
      Google ScholarLocate open access versionFindings
    Your rating :
    0

     

    Tags
    Comments