TDTMF: A recommendation model based on user temporal interest drift and latent review topic evolution with regularization factor

https://doi.org/10.1016/j.ipm.2022.103037Get rights and content

Abstract

This paper constructs a novel enhanced latent semantic model based on users’ comments, and employs regularization factors to capture the temporal evolution characteristics of users’ potential topics for each commodity, so as to improve the accuracy of recommendation. The adaptive temporal weighting of multiple preference features is also improved to calculate the preferences of different users at different time periods using human forgetting features, item interest overlap, and similarity at the semantic level of the review text to improve the accuracy of sparse evaluation data. The paper conducts comparison experiments with six temporal matrix-based decomposition baseline methods in nine datasets, and the results show that the accuracy is 31.64% better than TimeSVD++, 21.08% better than BTMF, 15.51% better than TMRevCo, 13.99% better than BPTF, 9.24% better than TCMF, and 3.19% better than MUTPD ,which indicates that the model is more effective in capturing users’ temporal interest drift and better reflects the evolutionary relationship between users’ latent topics and item ratings.

Introduction

With the development of big data era, recommendation system has played a vital role in life, and recommendation system is widely used in various fields (Bah, Aala, & Sm, 2020). Capturing the dynamic preference patterns of users is one of the challenges of current recommendation systems (Belém, Silva, Andrade, Person, & Gonalves, 2020). Dynamic preference patterns may change as time proceeds, and ignoring changes in user preferences and item characteristics can affect the accuracy of recommendations. Therefore, capturing the dynamics of user preferences using multiple temporal dynamics information is critical to better predict future user behavior.

In order to accurately identify the evolution of temporal dynamic preference features, various methods have been proposed to capture user dynamic preferences (Jianyun et al., 2019, Panagiotakis et al., 2017). Koren (2009) extend the singular value decomposition (SVD) and combine it with a time-varying deviation model to model user activity level and item popularity. However, this method cannot effectively solve the sparsity problem. Zhang, Wang, Yu, Sun, and Lim (2014) proposed the Temporal Matrix Factorization (TMF) model based on the assumption that user preferences change gradually over time. TMF models the temporal relevance of each user by means of a time-invariant transition matrix between two consecutive periods. However, users’ interests do not always evolve gradually and may also change essentially between two consecutive periods. Rafailidis, Kefalas, and Manolopoulos (2017) assume that users’ preferences in each time period t are determined by user preferences in all time periods before t. They model these multiple associations in the user latent space by constructing joint objective functions. Although there are some solutions to the data sparsity problem by exploiting the multimodal information of user-item interactions, they do not capture the dynamics of users’ temporal preferences well, especially without considering both the evolution of user behavior and the effect of time on the relationship between users’ sequential behaviors. All of the models above assume that the rating data is static and that the role of ratings changes over time in a real recommendation system. Sun and Dong (2017) applied short-term, long-term, and cyclical effects to a matrix of temporal impact factors and used the singular value method to decompose and predict unknown rating items. However, the impact of the weights of each time period is not fully considered. Li, Jin, Wu and Chen (2019) proposed a combined recommendation algorithm based on similarity and forgetting curve. Although the forgetting pattern of human interest is modeled to some extent, the process of evolution of user review information in time series is not taken into account. So far, there have been few studies of models that apply temporal weights to the combination of latent factor transformation and textual themes (Luo, Sun, Wang, Li, & Shang, 2017).

Therefore, a recommendation model (TDTMF) based on temporal drift and enhanced latent topics is proposed in this paper to try to better combine the user’s interest drift about each category of items over time with the review evolution. The proposed model considers the historical rating behavior of users and the latent topics of each item review text for modeling, introduces a temporal regularization factor to balance the temporal influence, and captures the evolution of user preferences through a transformation matrix.

We summarize the main contributions of this work as follows:

  • (1)

    We improve the adaptive temporal weights of multiple preference features. Human forgetting features, item interest similarity, and review text semantic level similarity are used to calculate preferences of different users at different time periods.

  • (2)

    We construct a novel enhanced latent topic model based on user review information, which captures the temporal evolution pattern of users’ latent topics for each product through regularization factors and solves the sparsity problem by using the evolution relationship between topics and ratings of each product in different periods.

  • (3)

    To the best of our knowledge, we construct a multiple non-negative matrix factorization with a temporal regularization factor to consider the synergistic effect of external factors on user preferences, and improve the accuracy of the model’s prediction of the dynamic evolution of temporal preferences for the first time.

Section snippets

Recommendations based on temporal dynamic capture

The TimeSVD++ algorithm proposed by Koren et al. earlier considered the user’s time-varying preferences, and introduced the time factor by improving two existing collaborative filtering methods, and achieved better recommendation results (Koren, 2009). This method focuses on the influence of global time on users only. It ignores the relationship between users’ preference changes at a certain time, which is easy to miss the important preference time-varying features. The Dynamic matrix

Problem formulation

In this paper, a temporal latent factorization recommendation model based on user ratings and review texts is proposed. In this section, the preliminary work to implement the model in this paper is presented and the definition and description of the basic concepts of the model are given. In addition, the motivation for the work in this paper is explained. Some common notations used in this paper are defined in Table 1.

Temporal latent interest drift transformation

Based on the definition in Section 3, this section represents the objective function as a minimization problem that minimizes the difference between the change in user item ratings and the objective matrix error R(t) at each timestep to capture the change in user preferences at each timestep. First define the minimization problem for the initial matrices R(t) and R(tT) corresponding at t and tT moments, respectively, as defined in Eqs. (5), (6). argminUR(t),VR(t)R(t)UR(t)VR(t)F2 argminUR(t

DATASETS

In this paper, nine datasets are used to test the performance of the model algorithm, all of which are selected from the Amazon dataset collected by Stanford University (Ni, Li, & McAuley, 2019). Each dataset includes user ratings, product information and time information corresponding to user behavior, in addition to the other corpus contained in the dataset corresponding to the various item types available from the public-facing Amazon.com website, with rating values ranging from 1–5

Conclusion

The purpose of the study is to construct a user interest model by capturing dynamic time-series preference features (ratings, reviews, etc.) of users at different time stages and using latent factor decomposition techniques, which not only enables dynamic tracking of time-varying features of user preferences, but also considers the influence of review evolution on users, and provides accurate personalized recommendations for users from a large amount of information. It is found that current

CRediT authorship contribution statement

Hao Ding: Conceptualization, Methodology, SoftWare, Writing – original draft, Writing–review & editing. Qing Liu: Conceptualization, Methodology, Writing – review & editing. Guangwei Hu: Conceptualization, Methodology, Supervision, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work is supported by the Major Program of the National Social Science Foundation of China “Research on the accurate construction of urban and rural community service system driven by big data” (Grant No. 20&ZD154).

References (38)

  • YangC. et al.

    Leveraging semantic features for recommendation: Sentence-level emotion analysis

    Information Processing & Management

    (2021)
  • AvazpourI. et al.

    Dimensions and metrics for evaluating recommendation systems

  • BahA. et al.

    Towards a real-time processing framework based on improved distributed recurrent neural network variants with fasttext for social big data analytics

    Information Processing & Management

    (2020)
  • Cheng, Z., Ding, Y., Zhu, L., & Kankanhalli, M. (2018). Aspect-aware latent factor model: Rating prediction with...
  • Cheng, Y., Qiu, G., Bu, J., Liu, K., Han, Y., & Wang, C., et al. (2008). Model bloggers’ interests based on forgetting...
  • ChuaF.C.T. et al.

    Modeling temporal adoptions using dynamic matrix factorization

  • EbbinghausH.

    Memory: A contribution to experimental psychology

    Annals of Neurosciences

    (2013)
  • Ipeirotis, P. G., Provost, F., & Wang, J. (2010). Quality management on amazon mechanical turk. In Proceedings of the...
  • JianyunH. et al.

    A multi-trans matrix factorization model with improved time weight in temporal recommender systems

    IEEE Access

    (2019)
  • Cited by (0)

    View full text