Elsevier

Information Sciences

Volume 541, December 2020, Pages 332-344
Information Sciences

Evolutive preference analysis with online consumer ratings

https://doi.org/10.1016/j.ins.2020.06.048Get rights and content

Abstract

The rapid growth of online rating information enables firms to learn and monitor consumer preferences over their products. Based on multidimensional rating information systems, we propose a novel class of Evolutive Preference Analysis (EPA) methods to handle the dynamic online ratings with arbitrary rating distribution, which takes all the historical ratings into consideration and delivers a comprehensive ranking evolution. As a new tool to analyze real-time consumer preferences, the EPA class includes the EPA that emphasizes the overall preferences of consumers, and the EPAt that emphasizes the recent preferences that generally provide up-to-date information. Both of them include four indices (the expected priority vector, expected rank, confidence factor, and index of rank probability) to reveal consumer preferences over rated attributes of products along time. The evolution trends of these indices help firms verify whether their advertised products match the preferences of target consumers to improve their products and marketing strategies. Finally, the practical applicability of EPA class is corroborated by several numerical examples and one real-world application on smartphone online ratings, which demonstrates the effectiveness of the proposed EPA class on streaming rating evolution.

Introduction

Consumer preference is a central problem in marketing, and learning consumer preferences from online ratings can be a basis of growing importance in today’s online environment. Monitoring consumer online comments and ratings of products is critical to deeply understand consumer feedback and preferences, which greatly supports firms’ operations and marketing strategies such as capacity planning, product pricing, market segmentation, advertising, and product recommendation.

Previous research in recommendation systems has built some consumer-preference models based on online reviews to predict consumer preferences. Chung and Rao [4] classify the existing consumer preference models into two categories: (1) collaborative filtering models based on consumer preference similarity, and (2) attribute-based preference or choice models based on product attribute similarity. The first category relies on the preferences of other consumers as a whole to predicting the preferences of target consumers. The second category uses observed product attributes and consumer characteristics to predict the preferences. However, these models focus on predicting consumer preferences in terms of ratings on products, rather than analyzing the changes or evolution of consumer preferences. As a result, these models have difficulties in directly providing much insight for firms, such as manufacturers and platforms.

There have been a series of studies relevant to monitoring public opinion evolution. Opinion mining and sentiment analysis have been widely investigated and these approaches aim to extract features, opinions, sentence subjectivity, and emotions from user-generated content. However, these studies mainly focus on text mining, which is different from the ranking problem embedded in consumer preference analysis. Next, we review the literature related to our topic.

Analysis on Online Ratings. The online reviews and ratings from consumers are highly affected by their preferences. For this reason, consumer reviews and ratings are more likely to focus on how or whether a product matches a specific individual’s preference [3]. Consumers use online rating systems to rate their purchased products. In particular, the multidimensional rating systems can help new consumers better match their preferences with product attributes and make purchase decisions [43]. The existing literature focusing on ratings can be roughly divided into three classes: (1) how online ratings reflect product quality or consumer satisfaction [2], (2) how the ratings affect sales [7], and (3) how to detect fake/suspicious ratings and reviews [29]. However, there are few pieces of literature that pay attention to monitoring consumer preference based on the ratings to facilitate firm operations.

Preference Expression. In decision making theories, the ranking methods (or prioritization methods) derive priority weights of compared objects from decision maker’s preference. The first issue before the ranking is to express the preference. The pairwise comparison is a commonly used technique to express the preferences of decision makers. It is the process of comparing objects, such as attributes and alternatives, in pairs to judge which one is preferred. For example, the well-known analytical hierarchy process [23] and analytical network process [25] utilize the comparisons of decision makers to derive rank order of objects. Saaty summarizes their wide applications in the problems of business planning, resource allocation, and organization priority setting [26]. Kou et al. summarize 93 journal articles in 37 peer-reviewed international journals [14], which focus on using the comparisons to deal with decision-making problems. Recent studies on decision making problems based on pairwise comparisons can be found in [1], [15], [16]. Moreover, the applications of the pairwise comparison technique are also utilized widely in other research areas, especially in the area of artificial intelligence [42], such as recommendation systems [38], predicting the purchases of consumers [12] and image quality assessment [20].

Another direction to express the decision makers’ preferences over objects is to model utility functions. Many utility functions have been proposed to explain the preferences, such as the expected utility theory [31] and the prospect theory [13]. However, the scores of utility functions are usually difficult to elicit. For example, the scores elicited from different decision makers become problematic, since the decision makers may not have a uniform scale of their scores [33]. In comparison, it is easy to compare two objects and determine a better one, which is a subjective ratio judgment. The theoretical foundation of ratio judgment is laid by [17], [21]. In a nutshell, it is a natural way to use subjective ratio judgment to elicit subjective preferences in terms of a weighted scale.

Based on the theoretical foundation of subjective ratio judgment, the subjective ratings of attributes can be converted into pairwise comparisons, that is, the subjective ratio judgments. Clark et al. [5] prove that the results from ratings correlate strongly with pairwise comparison methods. Similar conversions are supported by many studies. For example, Gerdsri and Kocaoglu [8] give the comparisons between attributes by comparing the values of attributes within [0,100] assigned by decision makers; Chen and Vulcano [12] construct pairwise companions between products by comparing the sequence of purchase observations of consumers; Ren et al. [22] build comparisons between movies by comparing the movie ratings; Nieto-Garcia et al. [18] compare the online ratings of hotel attributes to construct pairwise comparisons.

Ranking Methods Based on Pairwise Comparison. The values of a pairwise comparison can be deterministic, i.e. crisp, and interval values, or stochastic. For different forms of the comparison, there exist various preference relations or pairwise comparison matrices to express these comparisons for decision analysis. For example, there are (interval) fuzzy preference relations [19], [35], hesitant fuzzy preference relations [34], [39] for deterministic comparisons, and numerical preference relations (NPRs) [41] that allow the comparisons represented as stochastic variables. In particular, the NPR is a general form of the existing preference relations consisting of numerical values. To deal with deterministic pairwise comparisons, many valuable ranking methods with deterministic procedures have been developed to derive the priority weights of compared objects including the geometric mean [6], eigenvector method [24], the least deviation method [37], and the linear programming method [40]. In particular, the class of linear programming methods is simple and efficient in computation. For example, Wand and Chin [32] claim that their linear goal programming method can derive the optimal priorities of compared objects from inconsistent comparisons. Zhu and Xu [40] show that their fuzzy linear programming method can directly derive the optimal priorities from inconsistent and incomplete pairwise comparisons.

For the stochastic pairwise comparisons, some simulation methods based on the multiplicative preference relations have been proposed. Hahn [9] summarizes the ranking methods and claims that the deterministic procedures are special cases of the stochastic procedures. As a generalized stochastic procedure, Zhu and Xu [41] develop a stochastic preference analysis method to deal with the NPRs consisting of stochastic pairwise comparisons. The advantage of SPA is that it provides a series of ranking indices with probability interpretation to assist preference analysis. However, as a Monte Carlo based method, the limitation of SPA is that it requires the distribution of preference information, cannot deal with streaming data, and may collapse if the comparison information is overloaded.

In this paper, we propose a class of Evolution Preference Analysis (EPA) methods based on consumer ratings. Specifically, we focus on the products of firm that each has several attributes rated by online consumers. The online rating information is usually immense, coming in a continuous stream, being constantly updated and often exhibiting some degree of reversal over time. Based on multidimensional rating systems, we contribute to describing the evolution of consumer preference on the attributes based on real-time rating information. The class of EPA methods includes an EPA and an EPA with time decay (short for EPAt). The EPA equally considers all ratings, while EPAt pays more attention to newly arriving ratings because recent ratings generally provide more up-to-date information (see for example [10], [11]). The EPA/EPAt includes four indices that describe consumer preference on rated objects over time. These indices give insights into consumer preferences on the objects by monitoring the evolution trends of these indices. Firms can use the evolution trends to verify whether their advertised products match the preferences of target consumers. We feel this paper makes the following methodological contributions:

  • To the marketing literature, we model consumer preference by converting online ratings into pairwise comparisons.

  • To the decision analysis literature, specifically the ranking methods based on pairwise comparison, we develop a dynamic ranking method based on real-time data stream of ratings.

Section snippets

Preliminary knowledge

In this section, we summarize the concept of numerical preference relations (NPR) and the procedure of stochastic preference analysis (SPA) formalized in Zhu and Xu [41], then give a discussion on the ranking method involved in the SPA. Given a set of objects for comparison X={x1,x2,,xn}, a pairwise relation z(xi,xj) (i,j1,2,,n) is referred to as a pairwise comparison that indicates the dominance of xi over xj [28]. With respect to online ratings, the objects, for example, are attributes of

EPA procedure

Given a set of NPRs {Z(1),Z(2),,Z(T)} for n objects, with the timestamp 1tT, EPA updates all ranking indices of SPA for every moment t. For Z(1), we apply SPA to calculate the priority vector w(1). Then we provide the update rules for EPA as follows:E(w)(t+1)=tt+1E(w)(t)+1t+1w(t+1),where E(w)(t) presents the expected priority vector for timestamp t, and w(t+1) is the current priority vector derived from Z(t+1).

We can see that E(w)(t+1) involves the previous E(w)(t) and the current priority

Numerical examples of EPA class

In this section, we conduct several experiments to illustrate the EPA methods. Specifically, in Example 1, we demonstrate that EPA without time delay degrades into the SPA, which verifies the correctness of EPA. In Example 2, we illustrate the EPA, and analyze the impact of the length of time series on derived expected weights. In Example 2, we illustrate the EPAt, and compare it with the SPA. Following Example 2, we give three observations in Example 3 to further show how the confidence factor

An application

In this section, we apply the EPA methods to help phone manufactures analyze consumer preference on multiple attributes of their products in the real market. Specifically, there are four smartphones that each have five attributes. The four smartphones are VIVO (X23), OPPO (R17), One Plus (6) and Samsung (Galaxy S8), and the five attributes are Endurance, Photo, Performance, Appearance and Price/Performance. We use the data from Zol1, which is one of the largest online

Concluding remarks

Based on rating data, this paper develops the EPA class as a new tool to monitor and analyze consumer preferences over rated objects such as the attributes of products. The EPA class can deal with online rating data coming in as a stream with an arbitrary rating distribution. The EPA class updates the existing SPA to a dynamic version and fixes its hidden peril if the input information is overloaded. The EPA class includes the EPA that equally considers all ratings along timestamps, and the EPAt

CRediT authorship contribution statement

Xue Li: Writing - review & editing, Validation, Formal analysis, Project administration. Hongfu Liu: Conceptualization, Methodology, Writing - original draft, Visualization, Investigation. Bin Zhu: Conceptualization, Methodology, Writing - original draft, Supervision.

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.

Acknowledgements

The authors are grateful to the anonymous reviewers for their constructive and detailed comments on the manuscript. This research is supported by National Natural Science Foundation of China (Grant numbers 71902011, 61503210), and the Fundamental Research Funds for the Central Universities (Grant number 2018QD011).

References (43)

  • T.L. Saaty et al.

    Uncertainty and rank order in the analytic hierarchy process

    Eur. J. Oper. Res.

    (1987)
  • T. Tanino

    Fuzzy preference orderings in group decision making

    Fuzzy Sets Syst.

    (1984)
  • Y.M. Wang et al.

    A linear goal programming priority method for fuzzy analytic hierarchy process and its applications in new product screening

    Int. J. Approximate Reasoning

    (2008)
  • M.M. Xia et al.

    Hesitant fuzzy information aggregation in decision making

    Int. J. Approximate Reasoning

    (2011)
  • Z.S. Xu

    Goal programming models for obtaining the priority vector of incomplete fuzzy preference relation

    Int. J. Approximate Reasoning

    (2004)
  • Z.S. Xu et al.

    A least deviation method to obtain a priority vector of a fuzzy preference relation

    Eur. J. Oper. Res.

    (2005)
  • B. Zhu et al.

    A fuzzy linear programming method for group decision making with additive reciprocal fuzzy preference relations

    Fuzzy Sets Syst.

    (2014)
  • B. Zhu et al.

    Stochastic preference analysis in numerical preference relations

    Eur. J. Oper. Res.

    (2014)
  • M. Zihayat et al.

    A utility-based news recommendation system

    Decis. Support Syst.

    (2019)
  • O. Besbes et al.

    On information distortions in online ratings

    Oper. Res.

    (2018)
  • Y. Chen and J. Xie. Online consumer review: Word-of-mouth as a new element of marketing communication mix. Manage....
  • Cited by (11)

    • Decision making towards large-scale alternatives from multiple online platforms by a multivariate time-series-based method

      2023, Expert Systems with Applications
      Citation Excerpt :

      Online ratings can affect decision-making activities and attitude formation among users across different platforms such as IMDB or Douban (Purnawirawan, Eisend, De Pelsmacker, & Dens, 2015). For online ratings, it is critical to analyze the evolution or changes of user preferences from a dynamic perspective since the evolutional preferences can provide managerial insights for platforms, manufacturers, or cinemas (Li et al., 2020). Users are both opinion providers and consumers in a platform.

    • Information acquisition and assimilation capacities as determinants of technological niche markets

      2022, Journal of Innovation and Knowledge
      Citation Excerpt :

      The current framework improves upon both the game-theoretical models, mostly centering on the diffusion of technology (Beath et al., 1995; Chen et al., 2018; Zhang & Sun, 2020), and the traditional decision-theoretical ones where demand is based on stopping criteria that validate or dismiss the introduction of new technology (Jensen, 1982; McCardle, 1985). At a formal level, the lack of interaction between the strategic diffusion of innovations by firms and the subsequent emergence of demand by DMs represents a drawback that the equilibrium model developed in the current paper helps to solve (Kim Wang and Seidle, 2017; Li et al., 2020; Ren et al., 2021; Tiberius et al., 2021). We describe the main results intuitively and highlight the requirements that need to be satisfied for the corresponding equilibria to exist.

    View all citing articles on Scopus
    View full text