MRIF: Multi-resolution Interest Fusion for Recommendation

Shihao Li
Shihao Li
Dekun Yang
Dekun Yang
Bufeng Zhang
Bufeng Zhang

SIGIR '20: The 43rd International ACM SIGIR conference on research and development in Information Retrieval Virtual Event China July, 2020, pp. 1765-1768, 2020.

Cited by: 0|Bibtex|Views39|DOI:https://doi.org/10.1145/3397271.3401240
EI
Other Links: arxiv.org|dl.acm.org|dblp.uni-trier.de|academic.microsoft.com
Weibo:
We propose multi-resolution interest fusion model consisting of interest extraction layer, interest aggregation layer, and attentional fusion structure to address the problem of extracting and combining user preferences at different temporal-ranges

Abstract:

The main task of personalized recommendation is capturing users' interests based on their historical behaviors. Most of recent advances in recommender systems mainly focus on modeling users' preferences accurately using deep learning based approaches. There are two important properties of users' interests, one is that users' interests are...More

Code:

Data:

0
Introduction
  • Recommender systems have been evolving fast.

    As deep learning methods achieve state-of-the-art performances in a lot of fields such as computer vision and natural language processing, several deep learning based recommendation methods are developed by extending the traditional collaborative filtering techniques [2].

    users’ interests are dynamic and change over time, which are hard to express by simple factorization approaches.
  • Recommender systems have been evolving fast.
  • Users’ interests are dynamic and change over time, which are hard to express by simple factorization approaches.
  • The sequential recommender has attracted much attention recently due to its ability to capture users’ intents based on the order and relation of user behaviors.
  • GRU4Rec [3] uses GRU-based RNN to extract information from user interaction sequences.
  • CASER [7] embeds an item sequence into an image and learns sequential patterns via horizontal and vertical convolutional filters
Highlights
  • In recent years, recommender systems have been evolving fast.

    As deep learning methods achieve state-of-the-art performances in a lot of fields such as computer vision and natural language processing, several deep learning based recommendation methods are developed by extending the traditional collaborative filtering techniques [2].

    users’ interests are dynamic and change over time, which are hard to express by simple factorization approaches
  • We evaluate model performances in terms of Area under ROC curve (AUC), Group AUC (GAUC), Normalized Discounted Cumulative
  • BPR and NCF perform better than POP, which is because these two models incorporate user information using collaborative filtering based methods
  • We propose multi-resolution interest fusion model consisting of interest extraction layer, interest aggregation layer, and attentional fusion structure to address the problem of extracting and combining user preferences at different temporal-ranges
  • Interest extraction layer relies on transformer blocks to extract instantaneous user interests at each step
  • Experiments show that our method outperforms state-of-the-art recommendation methods consistently
  • Interest aggregation layer focuses on finding a group of user interests at different resolutions
Methods
  • The authors introduce the MRIF model in detail.
  • As is shown in Fig. 1, the proposed model is composed of three main parts, which are interest extraction layer, interest aggregation layer, and attentional fusion structure.
  • Interest extraction layer extracts instantaneous user interests from embedded behavior sequences.
  • Interest aggregation layer captures users’ interests at different temporalranges.
  • Attentional fusion structure combines users’ interests using attentional mechanisms to make predictions
Results
  • Experimental Results and Analysis

    The authors show in Table 2 the experimental results on two amazon datasets, namely Electronics and Movies.
  • The POP method performs worst in terms of all metrics since it only considers the popularity of items, and no user side information is taken into account.
  • BPR and NCF perform better than POP, which is because these two models incorporate user information using collaborative filtering based methods.
  • DIN achieves better results than BPR and NCF on all metrics, since DIN relies on attention mechanism and attends to user’s historical behaviors using the target item.
  • SASRec outperforms the other three sequential methods with the use of self-attention block.
  • MRIF-attn achieves best results on Movie dataset in terms of all metrics and best results on Electro dataset except under
Conclusion
  • The authors propose multi-resolution interest fusion model consisting of interest extraction layer, interest aggregation layer, and attentional fusion structure to address the problem of extracting and combining user preferences at different temporal-ranges.
  • Interest extraction layer relies on transformer blocks to extract instantaneous user interests at each step.
  • Interest aggregation layer focuses on finding a group of user interests at different resolutions.
  • The interest fusion structure adopts the attention mechanism to integrate multi-resolution interests to make predictions.
  • Experiments on two datasets under seven evaluation metrics demonstrate the superiority of the model
Summary
  • Introduction:

    Recommender systems have been evolving fast.

    As deep learning methods achieve state-of-the-art performances in a lot of fields such as computer vision and natural language processing, several deep learning based recommendation methods are developed by extending the traditional collaborative filtering techniques [2].

    users’ interests are dynamic and change over time, which are hard to express by simple factorization approaches.
  • Recommender systems have been evolving fast.
  • Users’ interests are dynamic and change over time, which are hard to express by simple factorization approaches.
  • The sequential recommender has attracted much attention recently due to its ability to capture users’ intents based on the order and relation of user behaviors.
  • GRU4Rec [3] uses GRU-based RNN to extract information from user interaction sequences.
  • CASER [7] embeds an item sequence into an image and learns sequential patterns via horizontal and vertical convolutional filters
  • Methods:

    The authors introduce the MRIF model in detail.
  • As is shown in Fig. 1, the proposed model is composed of three main parts, which are interest extraction layer, interest aggregation layer, and attentional fusion structure.
  • Interest extraction layer extracts instantaneous user interests from embedded behavior sequences.
  • Interest aggregation layer captures users’ interests at different temporalranges.
  • Attentional fusion structure combines users’ interests using attentional mechanisms to make predictions
  • Results:

    Experimental Results and Analysis

    The authors show in Table 2 the experimental results on two amazon datasets, namely Electronics and Movies.
  • The POP method performs worst in terms of all metrics since it only considers the popularity of items, and no user side information is taken into account.
  • BPR and NCF perform better than POP, which is because these two models incorporate user information using collaborative filtering based methods.
  • DIN achieves better results than BPR and NCF on all metrics, since DIN relies on attention mechanism and attends to user’s historical behaviors using the target item.
  • SASRec outperforms the other three sequential methods with the use of self-attention block.
  • MRIF-attn achieves best results on Movie dataset in terms of all metrics and best results on Electro dataset except under
  • Conclusion:

    The authors propose multi-resolution interest fusion model consisting of interest extraction layer, interest aggregation layer, and attentional fusion structure to address the problem of extracting and combining user preferences at different temporal-ranges.
  • Interest extraction layer relies on transformer blocks to extract instantaneous user interests at each step.
  • Interest aggregation layer focuses on finding a group of user interests at different resolutions.
  • The interest fusion structure adopts the attention mechanism to integrate multi-resolution interests to make predictions.
  • Experiments on two datasets under seven evaluation metrics demonstrate the superiority of the model
Tables
  • Table1: Statistics of datasets
  • Table2: Performance comparison of different recommendation methods
Download tables as Excel
Reference
  • Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In WWW. 507–517.
    Google ScholarLocate open access versionFindings
  • 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
  • Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk.
    Google ScholarFindings
  • 2016. Session-based recommendations with recurrent neural networks. In ICLR.
    Google ScholarFindings
  • [4] Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In ICDM.
    Google ScholarFindings
  • [5] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme.
    Google ScholarFindings
  • 2009. BPR: Bayesian personalized ranking from implicit feedback. In UAI. 452–461.
    Google ScholarLocate open access versionFindings
  • [6] Nachiketa Sahoo, Param Vir Singh, and Tridas Mukhopadhyay. 2010. A hidden Markov model for collaborative filtering. MIS Quarterly (2010).
    Google ScholarLocate open access versionFindings
  • [7] Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In WSDM. 565–573.
    Google ScholarLocate open access versionFindings
  • [8] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS. 5998–6008.
    Google ScholarFindings
  • [9] Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. In AAAI. 5941–5948.
    Google ScholarFindings
  • [10] Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In SIGKDD.
    Google ScholarFindings
Full Text
Your rating :
0

 

Tags
Comments