Learning persona-driven personalized sentimental representation for review-based recommendation
Introduction
With the increasing growth of available online information, personalized recommendation (Liu, Zheng, Li, Zhang, et al., 2022, Shen et al., 2021, Tewari and Barman, 2018) is playing an increasingly critical role in alleviating information overload. It is widely used in various fields, such as e-commerce (Bao et al., 2014, Chong et al., 2020, Li et al., 2021), social media (Liu, Zheng, Li, Zhang, et al., 2022, Ning et al., 2019) and industrial (Liu, Zheng, Li, Shen et al., 2022) fields. Not only can it help sellers gain revenue growth through accurate marketing, but users can spend less time discovering the items or services they are interested in. Collaborative Filtering (CF) (Ebesu et al., 2018, Xue et al., 2019) is extensively used in the most successful personalized recommendation, which models user preferences and item features based on historical interaction records such as user ratings and click behaviors. Most CF technologies are based on matrix factorization (MF) to compute the user–item satisfaction score. MF based on bayesian personalized ranking (BPRMF) (Rendle et al., 2009) and generalized MF (GMF) (He et al., 2017) are usually the basic models to improve top-N recommendation. However, these MF methods (Yi et al., 2019) suffer from sparsity problems, and it is also hard to model users’ fine-grained preferences and furnish explainable recommendations.
In most online e-commerce platforms (e.g., Amazon), users are allowed to express their attitudes or opinions utilizing reviews apart from ratings. The textual reviews usually contain rich information and fine-grained features. Therefore, review-based methods (Lei et al., 2016, Li et al., 2021, Wu et al., 2018) are proposed and Li et al. (2021) consider historical rating behavior and review latent factor representation learning. Although these studies have achieved significant improvement, we argue that they still have inherent limitations. Most existing methods neglect the differences of different users in sentimental expression styles or language usage habits. In essence, sentimental expression is a manifestation of personality, and different users with different personalities have different sentimental expression habits. For example, when a user writes reviews, on the same sentimental five-star rank, some users usually use gentle words, while others use stimulating words.
Specifically, personality refers to the characteristic pattern in a person’s thinking, feeling, and decision making (Siddique et al., 2019). In other words, it is a quality that a person has acquired during a comparatively long period. While the sentiment is a manifestation of personality and it can be perceived in an instant. Generally speaking, personality is to sentiment what climate is to weather. Personality traits exhibit the difference of preference and habit in people, and it is an integral part of social multiple interactions (Chen, 2011, Wu, Li et al., 2019). Therefore, research on personality has become increasingly significant in many related applications (see Fig. 1) such as social user profiles, personalized marketing, game design, job recruitment, and so on. Above all, in sociological research, words, phrases, and sentences written by people could reflect personality to some extent (Lin et al., 2019, Yarkoni, 2010). As illustrated in Fig. 1, inspired by the advancement of these applications, we hold that the free-text reviews could reflect the users’ personalities. In Fig. 2, we show review examples of different sentimental expression on Amazon. Fig. 2(A) is the review documents of different users on the same item, which shows that different users differ in language usage and tone when expressing positive (five-star) or negative (one-star) sentiments on the same item. Fig. 2(B) is the review documents of the user in different items, indicating the differences of sentimental expression styles of users in writing reviews of items. For example, on the left of Fig. 2(A), differently, two users use “love love love and I was so surprised” and “I believe are made of the lost feathers of angel’s wings” to express their likes respectively. Intuitively, the difference in user personality leads to the difference in sentimental expression styles or language usage habits. In this case, we aim to utilize personality (same as “persona” below) to drive interactive sentimental representation learning from reviews and help to address the above-mentioned concerns.
To this end, we propose a Persona-driven Sentimental Attentive Recommendation model (PSAR), which can cope with the mentioned problem effectively. As shown in Fig. 3, as for the interactive sentimental representation learning, we first adopt interactive fragmentary and sequential features to learn sentimental expression from reviews respectively, and then concatenate them to obtain the final personalized sentimental representation. In this part, we design an Attentive Persona-based Interaction (API) module to learn the interaction of similar and synonymous words or phrases, as well as the interaction of sentences with the analogous sentimental tone. Besides, we employ two neural attention networks in the API module to select informative interactions and pick up the users’ most concerned item features separately. In addition, the holistic interaction modeling enhances further the connotation of the original sentences. Finally, the results of interactive sentimental representation learning and holistic interaction modeling are concatenated as the input to the output layer, predicating the final satisfaction score. The main contributions of our work can be summarized as follows.
- •
To the best of our knowledge, we are the first to explore persona-driven sentimental representation modeling from textual reviews, enabling to understand the user’s behaviors. Specifically, we explore the complementary benefits of the sequence-based and fragment-based reviews to discover differences in personalized sentimental expression.
- •
We propose a novel Persona-driven Sentimental Attentive Recommendation model (PSAR) with personalized sentimental representation. PSAR effectively captures sentimental expression habits by using the interaction of similar and synonymous words or phrases, as well as sentences with the analogous sentimental tone.
- •
Experimental results on four real-world datasets demonstrate that the proposed model is significantly better than existing state-of-the-art methods on tasks of prediction precision, review mining, and recommendation interpretability.
The remainder of this paper is organized as follows. In Section 2, the existing works related to our method are presented. Section 3 describes the proposed PSAR thoroughly. In Section 4, we introduce experimental settings and evaluate the performance of PSAR by comparing some baselines on various real-world datasets. Finally, this paper is concluded and some future research points are shown in Section 5.
Section snippets
Related work
As mentioned before, most existing efforts focused on free-text reviews written by users to reach recommendation precision by a large margin and provide the interpretable recommended results (Wang et al., 2018). In the light of our target points, we roughly review the related work from two aspects: review-based recommendation, personality-driven recommendation.
Persona-based sentimental attentive recommendation
In this section, we introduce the proposed PSAR model in detail. Our PSAR achieves interactive sentimental representation learning. As a core module, API module is described to learn sequence-based (sentence-level) and fragment-based (word-level) attentive interactions severally and enhances the feature connections. In addition, the holistic interaction modeling is designed to enhance further the connotation of the original sentences. Afterward, the rating prediction and network learning are
Experiments
In this section, we first describe our experimental setup including datasets, comparison baselines, evaluation metrics, and experiment details. We then analyze our experimental results and discuss the performance of our model compared to baselines in terms of overall performance, ablation study, interaction and interpretability, and sensitivity study. We have a discussion at the end of this section.
Conclusion and future work
We have presented a novel Persona-driven Sentimental Attentive Recommendation model (PSAR) via personalized sentimental representation learning. As a core component, the personalized sentimental representation captures sentimental expression habits effectively by employing the interaction of similar and synonymous words or phrases, as well as sentences with the analogous sentimental tone. Experimental results on the public datasets show that our model greatly improves the recommendation
CRediT authorship contribution statement
Peipei Wang: Methodology, Software, Investigation, Resources, Writing – review & editing. Lin Li: Supervision, Conceptualization, Formal analysis, Resources. Ru Wang: Methodology, Investigation, Formal analysis. Xinhao Zheng: Resources, Software. Jiaxi He: Supervision, Writing – review & editing. Guandong Xu: Supervision, Data curation.
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.
Acknowledgment
This work is supported by Key Research and Development Project of Hubei Province (No. 2021BAA030) and Youth Project of National Natural Science Foundation of China (No. 62106070).
References (44)
- et al.
CARM: confidence-aware recommender model via review representation learning and historical rating behavior in the online platforms
Neurocomputing
(2021) - et al.
EDMF: Efficient deep matrix factorization with review feature learning for industrial recommender system
IEEE Transactions on Industrial Informatics
(2022) - et al.
Multi-perspective social recommendation method with graph representation learning
Neurocomputing
(2022) - et al.
Sequencing of items in personalized recommendations using multiple recommendation techniques
Expert Systems with Applications
(2018) - et al.
NeuO: Exploiting the sentimental bias between ratings and reviews with neural networks
Neural Networks
(2019) Personality in 100,000 words: A large-scale analysis of personality and word use among bloggers
Journal of Research in Personality
(2010)- et al.
TopicMF: Simultaneously exploiting ratings and reviews for recommendation
Interface and interaction design for group and social recommender systems
- et al.
Neural attentional rating regression with review-level explanations
- et al.
ANR: aspect-based neural recommender
Hierarchical visual-aware minimax ranking based on co-purchase data for personalized recommendation
Personality-aware product recommendation system based on user interests mining and metapath discovery
IEEE Transactions on Computational Social Systems
Collaborative memory network for recommendation systems
Recommendation via user’s personality and social contextual
Attentive aspect modeling for review-aware recommendation
ACM Transactions on Information Systems
TriRank: Review-aware explainable recommendation by modeling aspects
Neural collaborative filtering
Explainable recommendation with fusion of aspect information
World Wide Web
Enhancing collaborative filtering systems with personality information
The influence of listener personality on music choices
Computer Science
Rating prediction based on social sentiment from textual reviews
IEEE Transactions Multimedia
A persona-based neural conversation model
Cited by (5)
Personalized Recommendation Algorithm with Review Sentiments and Importance
2024, Data Analysis and Knowledge DiscoveryGMAT: A Graph Modeling Method for Group Preference Prediction
2024, Journal of Systems Science and Systems EngineeringGraphs get personal: learning representation with contextual pretraining for collaborative filtering
2023, Applied Intelligence