Daily-Aware Personalized Recommendation based on Feature-Level Time Series Analysis

WWW, pp. 1373-1383, 2015.

Cited by: 50|Bibtex|Views16|Links
EI
Keywords:
Moving Averagee commercelevel modelingproduct level modelingFourier-assisted ARIMAMore(16+)
Weibo:
In order to overcome the problem of data insufficiency in previous product-level modeling approaches, we propose to extract product features automatically from user reviews, and to degrade product-level modeling to feature-level modeling

Abstract:

The frequently changing user preferences and/or item profiles have put essential importance on the dynamic modeling of users and items in personalized recommender systems. However, due to the insufficiency of per user/item records when splitting the already sparse data across time dimension, previous methods have to restrict the drifting ...More

Code:

Data:

0
Introduction
  • The vast amount of online products in various e-commerce websites make it an essential task to develop reliable Personalized Recommender Systems (PRS) [24].
  • One of the most extensively investigated approaches for personalized recommendation is to model user preferences according to his/her historical choices, as well as those of the others’, leading to the success of Collaborative Filtering (CF) techniques [27]
  • Both user preferences and item profiles may change dynamically over time [16, 34, 11], treating the historical decisions of a user or the received comments of an item as static, fixed, or long-term influential information sources can be infeasible in real-world applications.
  • An item can receive new reviews from users continuously in practical systems, which makes its profile drift dynamically over time, as a result, it would be unwise for the system to recommend items that no longer fit the user needs any more
Highlights
  • The vast amount of online products in various e-commerce websites make it an essential task to develop reliable Personalized Recommender Systems (PRS) [24]
  • We develop the Fourier-assisted Auto-Regressive Integrated Moving Average (FARIMA) process that leverages Fourier terms to model the seasonal patterns and short-term Auto-Regressive Integrated Moving Average process to model the error
  • We briefly introduce the primary time series analysis (ARIMA) model first, and integrate Fourier series into Auto-Regressive Integrated Moving Average to derive our Fourier-assisted Auto-Regressive Integrated Moving Average model
  • We propose to leverage direct time series analysis for dynamic daily-aware recommendation
  • In order to overcome the problem of data insufficiency in previous product-level modeling approaches, we propose to extract product features automatically from user reviews, and to degrade product-level modeling to feature-level modeling
  • To make it computationally feasible for time series analysis on daily resolution, we develop the Fourier-assisted AutoRegressive Integrated Moving Average (FARIMA) approach for percentage time series fitting and prediction, and we further adapt the conditional opportunity model for personalized rating prediction and recommendation
Methods
  • The authors conduct extensive experiments to evaluate the performance of product feature extraction, dailyaware time series prediction, dynamic personalized recommendation, and the performance of the framework in handling with cold-start users.
Results
  • The authors find that a simple one-order fixed-period Fourier is able to preserve more than 70% of the energy, and 87.9% of the features even got 80% or more energy preserved.
  • The authors select the result with the highest F1-score primarily (Trial #3), which contains 58 features
Conclusion
  • The authors propose to leverage direct time series analysis for dynamic daily-aware recommendation.
  • The observation that a large amount of features exhibit clear seasonal and cyclic patterns may be inspiring for the construction of intelligent automatic marketing tools
  • Experimental results on both rating prediction and top-K recommendation, as well as cold-start performance verifies the superiority of the feature-level time series analysis approach
Summary
  • Introduction:

    The vast amount of online products in various e-commerce websites make it an essential task to develop reliable Personalized Recommender Systems (PRS) [24].
  • One of the most extensively investigated approaches for personalized recommendation is to model user preferences according to his/her historical choices, as well as those of the others’, leading to the success of Collaborative Filtering (CF) techniques [27]
  • Both user preferences and item profiles may change dynamically over time [16, 34, 11], treating the historical decisions of a user or the received comments of an item as static, fixed, or long-term influential information sources can be infeasible in real-world applications.
  • An item can receive new reviews from users continuously in practical systems, which makes its profile drift dynamically over time, as a result, it would be unwise for the system to recommend items that no longer fit the user needs any more
  • Methods:

    The authors conduct extensive experiments to evaluate the performance of product feature extraction, dailyaware time series prediction, dynamic personalized recommendation, and the performance of the framework in handling with cold-start users.
  • Results:

    The authors find that a simple one-order fixed-period Fourier is able to preserve more than 70% of the energy, and 87.9% of the features even got 80% or more energy preserved.
  • The authors select the result with the highest F1-score primarily (Trial #3), which contains 58 features
  • Conclusion:

    The authors propose to leverage direct time series analysis for dynamic daily-aware recommendation.
  • The observation that a large amount of features exhibit clear seasonal and cyclic patterns may be inspiring for the construction of intelligent automatic marketing tools
  • Experimental results on both rating prediction and top-K recommendation, as well as cold-start performance verifies the superiority of the feature-level time series analysis approach
Tables
  • Table1: Statistical distribution of pf of the features
  • Table2: A summary of the notations used
  • Table3: Cumulative statistical information in each quarter, where the statistics for a quarter represent the number of users, items and reviews in the system by the end of that quarter (included)
  • Table4: Four time-dependent datasets constructed for model training and testing
  • Table5: Evaluation of automatic feature extraction when trading off between precision and recall
  • Table6: Evaluation of time series prediction
  • Table7: The best RMSE achieved on four datasets and their averaged value for each method, standard deviations ≤ 0.015 for each experimental run
  • Table8: The number of selected users and best
  • Table9: The number of users for evaluation, and the
Download tables as Excel
Related work
  • With the ability to help discover items of potential interests, Personalized Recommender Systems (PRS) [24] have been widely integrated into many online applications, such as e-commerce, social networks, and online review services. Early systems for personalized recommendation rely on contentbased approaches [21], which make recommendations by analyzing the item features or user demographics. Recently, the Collaborative Filtering (CF) [27] based approaches have gained great popularity due to their free from human efforts, Nutri9ous An9-­‐UV superior performance, and the ability to take advantage of the wisdom of crowds.

    In real-world systems, however, it is usually observed that both user preferences and item profiles may dynamically drift over time [5], thus a static modeling of users/items that treats the historical purchases or reviews as fixed and long-term effective may be inappropriate to track users’ interests. For example, a user may be less likely to buy an anti-UV suncream in winter, although she may well have bought one in summer; and a user who has purchased an SLR camera may not buy another one in a reasonably long period. As a result, the ability of time-aware modeling is of essential importance to practical systems [5].
Funding
  • This work was supported by National Key Basic Research Program (2015CB358700) and Natural Science Foundation (61472206, 61073071) of China, and Yi is sponsored by the National Science Foundation under grant CCF-1101741 and IIS-0953908
Reference
  • G. Adomavicius, B. Mobasher, F. Ricci, and A. Tuzhilin. Context-Aware Recommender Systems. Recommender systems handbook, pages 217–253, 2011.
    Google ScholarFindings
  • L. Baltrunas and X. Amatriain. Towards Time Dependant Recommendation based on Implicit Feedback. CARS, 2009.
    Google ScholarLocate open access versionFindings
  • G. E. P. Box, G. M. Jenkins, and G. C. Reinsel. Time Series Analysis: Forecasting and Control. John Wiley & Sons, 2013.
    Google ScholarFindings
  • K. P. Burnham and D. R. Anderson. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. Springer, 2002.
    Google ScholarFindings
  • P. G. Campos, F. Dıez, and I. Cantador. Time-aware Recommender Systems: A Comprehensive Survey and Analysis of Existing Evaluation Protocols. User modeling & user-adapted interaction, 24:67–119, 2014.
    Google ScholarLocate open access versionFindings
  • T. Chen, W. Han, H. Wang, Y. Zhou, B. Xu, and B. Zang. Content Recommendation System based on Private Dynamic User Profiling. ICMLC, 2007.
    Google ScholarLocate open access versionFindings
  • W. Chen, W. Hsu, and M. Lee. Modeling User’s Receptiveness Over Time for Recommendation. SIGIR, pages 373–382, 2013.
    Google ScholarLocate open access versionFindings
  • H. Choi and H. Varian. Predicting the Present with Google Trends. Economic Record, 88(s1):2–9, 2012.
    Google ScholarLocate open access versionFindings
  • W. Chu and S. Park. Personalized Recommendation on Dynamic Content Using Predictive Bilinear Models. WWW, pages 691–700, 2009.
    Google ScholarLocate open access versionFindings
  • X. Ding, B. Liu, and P. S. Yu. A Holistic Lexicon Based Approach to Opinion Mining. WSDM, 2008.
    Google ScholarLocate open access versionFindings
  • G. Dror, N. Koenigstein, and Y. Koren. Yahoo! music recommendations: Modeling music ratings with temporal dynamics and item taxonomy. RecSys, 2011.
    Google ScholarLocate open access versionFindings
  • Z. Gantner, S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme. MyMediaLite: A Free Recommender System Library. RecSys, 2011.
    Google ScholarLocate open access versionFindings
  • Z. Gantner, S. Rendle, and L. Schmidt-Thieme. Factorization Models for Context-/Time-Aware Movie Recommendations. CAMRa, pages 14–19, 2010.
    Google ScholarLocate open access versionFindings
  • S. Gauch, M. Speretta, A. Chandramouli, and A. Micarelli. User Profiles for Personalized Information Access. The Adaptive Web, pages 54–89, 2007.
    Google ScholarLocate open access versionFindings
  • A. Karatzoglou, X. Amatriain, L. Baltrunas, and N. Oliver. Multiverse Recommendation: N-dimensional Tensor Factorization for Context-aware Collaborative Filtering. RecSys, pages 79–86, 2010.
    Google ScholarLocate open access versionFindings
  • Y. Koren. Collaborative Filtering with Temporal Dynamics. KDD, pages 447–455, 2009.
    Google ScholarLocate open access versionFindings
  • D. D. Lee and H. S. Seung. Algorithms for Non-negative Matrix Factorization. NIPS, 2001.
    Google ScholarLocate open access versionFindings
  • Y. Lu, M. Castellanos, U. Dayal, and C. Zhai. Automatic construction of a context-aware sentiment lexicon: An optimization approach. WWW, 2011.
    Google ScholarLocate open access versionFindings
  • Z. Lu, D. Agarwal, and I. Dhillon. A Spatio Temporal Approach to Collaborative Filtering. RecSys, 2009.
    Google ScholarLocate open access versionFindings
  • K. Oku, S. Nakajima, J. Miyazaki, S. Uemura, and H. Kato. A Recommendation Method Considering Users’ Time Series Contexts. ICUIMC, 2009.
    Google ScholarLocate open access versionFindings
  • M. J. Pazzani and D. Billsus. Content-Based Recommendation Systems. The Adaptive Web LNCS, pages 325–341, 2007.
    Google ScholarFindings
  • A. M. Popescu and O. Etzioni. Extracting Product Features and Opinions from Reviews. EMNLP, 2005.
    Google ScholarLocate open access versionFindings
  • S. Rendle, C. Freudenthaler, Z. Gantner, and L. S. Thieme. BPR: Bayesian Personalized Ranking from Implicit Feedback. UAI, 2009.
    Google ScholarLocate open access versionFindings
  • F. Ricci, L. Rokach, and B. Shapira. Introduction to Recommender Systems Handbook. Springer US, 2011.
    Google ScholarFindings
  • Y. Shi, M. Larson, and A. Hanjalic. Mining Mood-specific Movie Similarity with Matrix Factorization for Context-aware Recommendation. CAMRa, pages 34–40, 2010.
    Google ScholarLocate open access versionFindings
  • R. H. Shumway and D. S. Stoffer. Time Series Analysis and Its Application. Springer, 2010.
    Google ScholarFindings
  • X. Su and T. M. Khoshgoftaar. A Survey of Collaborative Filtering Techniques. Advances in Artificial Intelligence, 4, 2009.
    Google ScholarLocate open access versionFindings
  • G. Takacs, I. Pilaszy, B. Nemeth, and D. Tikk. Investigation of Various Matrix Factorization Methods for Large Recommender Systems. Proc. ICDM, 2008.
    Google ScholarLocate open access versionFindings
  • D. Tang, F. Wei, B. Qin, M. Zhou, and T. Liu. Building Large-Scale Twitter-Specific Sentiment Lexicon: A Representation Learning Approach. COLING, pages 172–182, 2014.
    Google ScholarLocate open access versionFindings
  • C. Vaca, A. Mantrach, A. Jaimes, and M. Saerens. A Time-based Collective Factorization for Topic Discovery and Monitoring in News. WWW, 2014.
    Google ScholarLocate open access versionFindings
  • J. Wang and Y. Zhang. Is It Time For a Career Switch? WWW, pages 1377–1387, 2013.
    Google ScholarLocate open access versionFindings
  • J. Wang and Y. Zhang. Opportunity Models for E-commerce Recommendation: Right Product, Right Time. SIGIR, pages 303–312, 2013.
    Google ScholarLocate open access versionFindings
  • Y. Wu and M. Ester. FLAME: A Probabilistic Model Combining Aspect Based Opinion Mining and Collaborative Filtering. WSDM, pages 199–208, 2015.
    Google ScholarLocate open access versionFindings
  • L. Xiang, Q. Yuan, S. Zhao, L. Chen, X. Zhang, Q. Yang, and J. Sun. Temporal Recommendation on Graphs via Long- and Short-term Preference Fusion. KDD, pages 723–731, 2010.
    Google ScholarLocate open access versionFindings
  • D. Xu, Y. Liu, M. Zhang, S. Ma, A. Cui, and L. Ru. Predicting Epidemic Tendency through Search Behavior Analysis. IJCAI, pages 2361–2366, 2011.
    Google ScholarLocate open access versionFindings
  • A. Yessenalina, Y. Yue, et al. Multi-level structured models for document-level sentiment classification. EMNLP, pages 1046–1056, 2010.
    Google ScholarLocate open access versionFindings
  • Q. Yuan, G. Cong, Z. Ma, A. Sun, and N. M. Thalmann. Time-aware Point-of-interest Recommendation. SIGIR, pages 363–372, 2013.
    Google ScholarLocate open access versionFindings
  • Y. Zhang, G. Lai, M. Zhang, Y. Zhang, Y. Liu, and S. Ma. Explicit Factor Models for Explainable Recommendation based on Phrase-level Sentiment Analysis. SIGIR, pages 83–92, 2014.
    Google ScholarLocate open access versionFindings
  • Y. Zhang, H. Zhang, M. Zhang, Y. Liu, et al. Do Users Rate or Review? Boost Phrase-level Sentiment Labeling with Review-level Sentiment Classification. SIGIR, pages 1027–1030, 2014.
    Google ScholarLocate open access versionFindings
  • Y. Zhang, M. Zhang, Y. Zhang, Y. Liu, and S. Ma. Understanding the Sparsity: Augmented Matrix Factorization with Sampled Constraints on Unobservables. CIKM, 2014.
    Google ScholarLocate open access versionFindings
Your rating :
0

 

Tags
Comments