Learning persona-driven personalized sentimental representation for review-based recommendation

Highlights

• A persona-driven sentimental model is proposed for review-based recommendation.

• The usage habits and analogous tones are explored from textual reviews.

• Seq-based and frag-based features are considered to learn sentimental representation.

• Experimental results demonstrate the effectiveness of the proposed PSAR.

Abstract

A large amount of information exists in many e-commerce and review websites as a valuable source for recommender systems. Recent solutions focus on exploring the correlation between sentiment and textual reviews in the review-based recommendation. However, these studies usually pay less attention to the differences of different users in sentimental expression styles or language usage habits when a user writes reviews. In this work, we argue that the individual reviewing behavior is closely related to personality, and sentimental expression is a manifestation of personality. Therefore, we propose a novel Persona-driven Sentimental Attentive Recommendation model (named PSAR) via personalized sentimental interactive representation learning for the review-based recommendation. The proposed model is devised to learn fragment-level and sequence-level personalized sentimental representation simultaneously from reviews. Besides, an attentive persona-driven interaction module is designed to capture word-level usage habits and sentence-level analogous tones. Comprehensive experimental results on four real-world datasets demonstrate that our model outperforms the state-of-the-art methods.

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.