TDTMF: A recommendation model based on user temporal interest drift and latent review topic evolution with regularization factor
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 are determined by user preferences in all time periods before . 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 at each timestep to capture the change in user preferences at each timestep. First define the minimization problem for the initial matrices and corresponding at and moments, respectively, as defined in Eqs. (5), (6).
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)
- et al.
“Fixing the curse of the bad product descriptions” ? Search-boosted tag recommendation for E-commerce products
Information Processing & Management
(2020) - et al.
A deep recommendation model of cross-grained sentiments of user reviews and ratings
Information Processing & Management
(2022) - et al.
Tensor factorization method based on review text semantic similarity for rating prediction
Expert Systems with Applications
(2018) - et al.
Recommendation system exploiting aspect-based opinion mining with deep learning method
Information Sciences
(2020) - et al.
An enhanced tree-LSTM architecture for sentence semantic modeling using typed dependencies
Information Processing & Management
(2020) - et al.
Combined recommendation algorithm based on improved similarity and forgetting curve
Information
(2019) - et al.
Sparse online collaborative filtering with dynamic regularization
Information Sciences
(2019) - et al.
Time and ontology for resource recommendation system
Physica A: Statistical Mechanics and its Applications
(2019) - et al.
Preference dynamics with multimodal user-item interactions in social media recommendation
Expert Systems with Applications
(2017) - et al.
A novel temporal recommender system based on multiple transitions in user preference drift and topic review evolution
Expert Systems with Applications
(2021)