The identification of dynamic product innovation opportunities using the multi-phase QFD: the customer requirement and technology development perspectives
In this section, we detail our method as outlined in Fig. 1. First, we collect and preprocess online reviews. Second, to identify PIOs from the customer requirements perspective, we assess how each feature influences customer satisfaction in time slice \({t}_{k}\). Third, from the technology development perspective, we evaluate technology levels using three criteria. Fourth, we apply a multi-phase QFD to compute innovation values by integrating dynamic customer needs and technology development. Finally, we identify static and dynamic PIOs based on innovation values and suggest corresponding innovation strategies.
The figure introduces the details of the five steps of our proposed method.
Data collection and text preprocessing
The first step of our method is data collection and text preprocessing. For the researched product, we first select several salable product models and collect their online reviews. These product models are denoted as \({P}_{1},{P}_{2},\ldots ,{P}_{{NP}}\). \({NP}\) is the number of selected models. Then we split the dates that online reviews generated into \(T\) time slices, which are represented as \({t}_{1},\) \({t}_{2},…,\) \({t}_{T}\). Reviews generated in the slice \({t}_{k}\) is denoted as \({R}_{k}\). Reviews of the product \({P}_{j}\) in the slice \({t}_{k}\) is denoted as \({R}_{j,k}\). Before analyzing reviews, we implement text preprocessing, which includes the removal of stopwords, common words, and useless symbols. Stopwords and common words have high frequencies but are less helpful in understanding texts. For example, when we research the online reviews of a certain product, the product name will occur frequently, but the name is useless in understanding the customer opinions expressed in reviews. Useless symbols such as URLs and meaningless foreign characters will interrupt the accuracy of text analysis and thus are deleted.
Furthermore, we retrieve relevant patents for the researched product from patent databases and collect patent information, including patent names, patent abstracts, technical keywords of patents, and patent filing dates. All patents will be split into \(T\) time slices according to their filing dates. Patents in the slice \({t}_{k}\) denoted as \(P{D}_{k}\).
Product performance in meeting customer requirements in the time slice \({{\rm{t}}}_{{\rm{k}}}\)
To estimate the performance of specific products in satisfying customer requirements, we first derive product features that customers are concerned about from online reviews. Then we evaluate customer opinions on these features using sentiment analysis. By measuring customer preference to satisfaction or dissatisfaction, we finally assess the extent to which customer satisfaction is stimulated by different features.
Product feature extraction using BERTopic
We use an effective topic modeling method named BERTopic to derive product features from preprocessed reviews. BERTopic obtains topics in four steps: document embeddings, dimensionality reduction, cluster analysis, and topic representation (Grootendorst, 2022). Based on Sentence-BERT, BERTopic firstly transfers complicated texts into vectors. Given that these text vectors are usually high-dimensional, BERtopic then uses descending dimension methods to reduce the dimensions of text vectors and enhance processing effectiveness. Based on the vectors after dimensionality reduction, BERTopic leverages cluster methods such as HDBSCAN to classify texts into various topics. At last, BERTopic utilizes Class-based Term Frequency-Inverse Document Frequency (c-TF-IDF) to select keywords from the texts relevant to different topics. Topic names are decided by researching the meanings of keywords related to the same topic. For the advantages of accuracy and effectiveness in obtaining topics from the large corpus, BERTopic has been utilized by recent studies to collect product features from social media data (Ma et al., 2024; Wang et al., 2025; Yi et al., 2025).
In our study, we apply BERTopic to \({R}_{k}\) and acquire \({D}_{k}\) topics with their keywords. To ensure the accuracy of product feature derivation, we invite an expert to further read BERTopic results manually and identify product features with their keywords. Product features identified at different time slices will be added to the feature set named \({\boldsymbol{F}}\). Features in \({\boldsymbol{F}}\) are indicated as \({F}_{1},{F}_{2},\ldots ,{F}_{{NF}}\). \({NF}\) is the number of product features. The set of keywords related to the feature \({F}_{i}\) is represented as \({\boldsymbol{FK}}{{\boldsymbol{W}}}_{i}\). \({FK}{W}_{i,n}\) is the nth word in \({\boldsymbol{FK}}{{\boldsymbol{W}}}_{i}\).
Product feature performance evaluation using sentiment analysis
After obtaining product features, we use sentiment analysis to assess the performance of different features in the slice \({t}_{k}\). We first group reviews in \({R}_{k}\) into different clusters according to their relevance to various features. For the review \({r}_{k}\) in \({R}_{k}\), if \({r}_{k}\) contains the keywords related to the feature \({F}_{i}\), \({r}_{k}\) will be classified into the cluster related to \({F}_{i}\). The review of \({P}_{j}\) related to \({F}_{i}\) at slice \({t}_{k}\) is denoted as \({R}_{{ji}}^{k}\).
Given that customers’ sentiments reflect their satisfaction or dissatisfaction with the product performance, we utilize effective toolkits of sentiment analysis, such as NLTK, SpaCy, and SnowNLP, to measure the feature performance. For the feature \({F}_{i}\) \({P}_{j}\), its performance score equals the average sentiment score of \({R}_{{ji}}^{k}\). The performance scores of all features at various slices will all be in \([\mathrm{0,1}]\). As for \({P}_{j}\) its performance on \({F}_{i}\) in the slice \({t}_{k}\) is denoted as \({per}{f}_{k}\left({P}_{j},{F}_{i}\right)\).
Measuring the extent to which customer satisfaction is stimulated
If a feature effectively evokes customer satisfaction, it has the potential to meet customer requirements and is prone to be a PIO. We adopt the idea developed in the study of Qi et al. (2016) to measure the extent to which customer satisfaction is stimulated (denoted as CRExtent). Qi et al. viewed customer rates of reviews as total satisfaction, which is induced by the performance of different features. They used regression analysis to measure the impact of different features on overall satisfaction, thereby studying the role of various features in provoking customer satisfaction.
For the review \(r\) in \({R}_{{ji}}^{k}\), its sentiment score is indicated as \({senti}(r)\). If \({senti}(r)\) is larger than the preset threshold \(\lambda\), we assume that \(r\) expresses a positive sentiment, implying that customers are satisfied. Conversely, \(r\) expresses a negative sentiment if \({senti}(r)\) is not larger than \(\lambda\). Furthermore, the larger the score of \({senti}(r)\), the more positive the sentiment is. Thus, we propose a method (displayed in Eqs. (1)–(6)) to estimate the impact of \({P}_{j}\)’s features on customer satisfaction.
$$s{s}_{{ji}}^{k}=\frac{1}{\left|{R}_{{ji}}^{k}\right|}\mathop{\sum }\limits_{r\in {R}_{{ji}}^{k}}\left({senti}\left(r\right)-\lambda \right).$$
(1)
$${X}_{{jik}}^{{pos}}=\left\{\begin{array}{c}s{s}_{{ji}}^{k},s{s}_{{ji}}^{k} > 0\\ 0,s{s}_{{ji}}^{k}\le 0\end{array}\right..$$
(2)
$${X}_{{jik}}^{{neg}}=\left\{\begin{array}{c}0,s{s}_{{ji}}^{k} > 0\\ s{s}_{{ji}}^{k},s{s}_{{ji}}^{k}\le 0\end{array}\right..$$
(3)
$$R{S}_{j}^{k}=\mathop{\sum }\limits_{i=1}^{{NF}}\left({\alpha }_{{ji}}^{k}{X}_{{ji}}^{{pos}}+{\beta }_{{ji}}^{k}{X}_{{ji}}^{{neg}}\right)+{c}_{j}.$$
(4)
$${Norm}\left(x\right)=\frac{x}{{x}_{\max }}.$$
(5)
$${CRExten}{t}_{k}({P}_{j},{F}_{i})={Norm}\left({\alpha }_{{ji}}^{k}-{\beta }_{{ji}}^{k}\right).$$
(6)
\(s{s}_{{ji}}^{k}\) means the score of positive/negative sentiments for the \({P}_{j}\)’s reviews of \({F}_{i}\). \({X}_{{jik}}^{{pos}}\) (or \({X}_{{jik}}^{{neg}}\)) denotes whether customers hold a positive (or negative) sentiment on \({F}_{i}\) of \({P}_{j}\). \(R{S}_{j}^{k}\) is the average review rate of \({R}_{{ji}}^{k}\). \({\alpha }_{{ji}}^{k}\) and \({\beta }_{{ji}}^{k}\) represent the customer preference to express positive/negative sentiments, respectively. \({c}_{j}\) is a constant. Equation(5) is a normalization function which transfers values into \([\mathrm{0,1}]\). \(x\) denotes a raw value (e.g., performance score \({\alpha }_{{ji}}^{k}-{\beta }_{{ji}}^{k}\), or any other calculated metric), and \({x}_{\max }\) represents the maximum value of \(x\) within the same slice. \({CRExten}{t}_{k}({P}_{j},{F}_{i})\) is the CRExtent score for \({F}_{i}\) of \({P}_{j}\) in the slice \({t}_{k}\) and \({CRExten}{t}_{{ji}}^{k}\in [\mathrm{0,1}]\). A larger \({CRExten}{t}_{k}({P}_{j},{F}_{i})\) implies that \({F}_{i}\) of \({P}_{j}\) can effectively evoke satisfaction and satisfy customer requirements.
Measuring the levels of technology development in the time slice \({t}_{k}\)
To estimate the levels of technology development, we need to derive key technologies related to different features from patents. We use feature keywords to find patents related to various features. For a patent in \(P{D}_{k}\), if its name and abstract contain the keywords of \({F}_{i}\), we assume that this patent is related to \({F}_{i}\). Patents relevant to \({F}_{i}\) in the slice \({t}_{k}\) are indicated as \(P{D}_{k,i}\). Based on \(P{D}_{k,i}\), we collect technical key technologies related to \({F}_{i}\) in the following steps:
-
(1)
Collecting technical keywords from the patents in \(P{D}_{k,i}\). If the patent does not have technical keywords, we use the preprocessed patent name as technical keywords.
-
(2)
Technical keywords which are common words and do not appear in the patent name and abstract will be removed.
-
(3)
All technical keywords related to \({F}_{i}\) are denoted as \({TK}{W}_{i}\). We leverage Term Frequency-Inverse Document Frequency (TF-IDF) to measure the weights of words in \({TK}{W}_{i}\) and rank these words based on their weights. For each feature, its top \({N}_{{tkw}}\) technical words will be filtered and selected by experts as the key technologies of \({F}_{i}\).
With the steps above, we obtain the key technologies of different features, which will be added to the set of key technologies in the slice \({t}_{k}\) (denoted as \(K{T}_{k}\)). \({NT}\) is the number of all key technologies. Based on the obtained technologies, we propose three criteria to evaluate the levels of technology development of different technologies.
For the key technology \(K{T}_{h}\), if more patents refer to \(K{T}_{h}\), its level of technology development is higher than other technologies. Thus, the first aspect of technology development evaluation is the number of patents related to a certain key technology (denoted as NumTech). The measurement is shown in Eq. (7),
$${NumTech}\left(K{T}_{h}\right)={Norm}\left({N}_{h}^{{patent}}\right),$$
(7)
where \({N}_{h}^{{patent}}\) is the number of patents mentioned \(K{T}_{h}\) \({NumTech}\left(K{T}_{h}\right)\in [\mathrm{0,1}]\)? The larger the \({NumTech}\left(K{T}_{h}\right)\), the more abundant the research outcomes related to \(K{T}_{h}\).
The number of patents reflects how popular the key technology is. However, this criterion cannot estimate how cutting-edge the key technology is. For the key technology \(K{T}_{h}\), more recent patent filing dates \(K{T}_{h}\) indicate \(K{T}_{h}\)’s cutting-edge nature and signify its importance as a current trend in technology development. Thus, the second aspect of technology development evaluation is the cutting-edge degree of key technologies (denoted as CutEdge). The measurement is introduced in Eqs. (8, 9),
$${CutEdge}\left(p{d}_{h}\right)=\frac{{Date}\left(p{d}_{h}\right)-{{EDate}}_{k}}{{{LDate}}_{k}-{{EDate}}_{k}},$$
(8)
$${CutEdge}\left(T{E}_{h}\right)=\frac{1}{\left|P{D}_{k,h}\right|}\mathop{\sum}\limits _{p{d}_{h}\in P{D}_{k,h}}{CutEdge}\left(p{d}_{h}\right),$$
(9)
where \(p{d}_{h}\) is the hth patent in the set of patent data related to \(K{T}_{h}\), which is indicated as \(P{D}_{k,h}\). \({Date}\left(p{d}_{h}\right)\) is the filing date of \(p{d}_{h}\). \({{EDate}}_{k}\) and \({{LDate}}_{k}\) are the earliest and latest date of the time slice \({t}_{k}\). \({CutEdge}\left(p{d}_{h}\right)\) and \({CutEdge}\left(T{E}_{h}\right)\) are the cutting-edge degree of \(p{d}_{h}\) and \(T{E}_{h}\), respectively. They are all in \([\mathrm{0,1}]\).
Moreover, if there is a continuous stream of patents related to \(K{T}_{h}\) in the time slice \({t}_{k}\), \(K{T}_{h}\) is in an ongoing state of development, suggesting a longer technological lifecycle. Therefore, its level of technological advancement is higher than that of technologies with shorter development periods. Based on the observation above, we propose the third aspect of technology development evaluation which is the continuity of technology development (denoted as TechCont). The measurement is presented in Eqs. (10–13).
$${\chi }^{2}=\mathop{\sum }\limits_{{n}_{1}=1}^{{N}_{{inter}}}\frac{{\left({N}_{{n}_{1}}^{{obs}}-{E}_{{n}_{1}}\right)}^{2}}{{E}_{{n}_{1}}}.$$
(10)
$$p={p}_{{value}}\left({\chi }^{2},{N}_{{inter}}-1\right).$$
(11)
$${Continuity}\left(T{E}_{h}\right)=\left\{\begin{array}{c}p,p\ge {p}^{* }\\ 0,p < {p}^{* }\end{array}\right..$$
(12)
$${TechCont}\left(T{E}_{h}\right)={Norm}\left(\frac{{LDat}{e}_{k,h}-{EDat}{e}_{k,h}}{{{LDate}}_{k}-{{EDate}}_{k}}+{Continuity}\left(T{E}_{h}\right)\right).$$
(13)
In Eqs. (10–12), We use the chi-square test to determine whether the filing dates of patents related to \(K{T}_{h}\) are uniformly distributed within the time slice \({t}_{k}\). The period \({t}_{k}\) is uniformly divided into \({N}_{{inter}}\) intervals. \({N}_{{n}_{1}}^{{obs}}\) represents the observed date data in the n1th interval. \({E}_{{n}_{1}}\) denotes the expected value for the n1th interval and \(p\) is the p-value obtained based on chi-square values and degrees of freedom. \({Continuity}\left(T{E}_{h}\right)\) indicates the degree of uniform distribution of patent filing dates related to \(T{E}_{h}\). \({p}^{* }\) is the significance level. In Eq. (13), \(\frac{{LDat}{e}_{k,h}-{EDat}{e}_{k,h}}{{{LDate}}_{k}-{{EDate}}_{k}}\) assesses the range of filing dates for patents related to \(T{E}_{h}\). \({TechCont}\left(T{E}_{h}\right)\) means the continuity of technology development for \(T{E}_{h}\) and \({TechCont}\left(T{E}_{h}\right)\in [\mathrm{0,1}]\).
The multi-phase QFD for innovation value evaluation
With the results of CRExtent, NumTech, CutEdge, and TechCont obtained in different slices, we propose a multi-phase QFD to evaluate the innovation values of various features.
As Fig. 2a shows, our proposed QFD contains \(T\) phases, and each phase indicates a slice. For the tkth phase, we use the results of CRExtent, NumTech, CutEdge, and TechCont obtained in the slice \({t}_{k}\) to calculate innovation values of features based on the results in the previous phase. Furthermore, the results of the tkth phase will also be considered in the calculation of the (tk + 1)th phase. By researching the results of a specific phase, we identify the static PIOs. The results of different slices reflect the evolution of PIOs, which helps us reveal dynamic PIOs and derive dynamic and informative insights on product innovation.
a The tth phase of the QFD discovers PIOs in the slice tk based on the PIOs results of the slice tk−1. PIOs identified in each slice are denoted as static PIOs. Dynamic PIOs are determined based on the trends of static PIOs identified in all T slices. b We derive gaps in customer requirements by comparing the feature performance of various products. Gaps in technology development are obtained by considering the evaluation results of key technologies and the relationships between features and key technologies. With two gaps, we decide PIOs in the slice tk from perspectives of customer requirements and technology development.
We use the tkth phase of QFD shown in Fig. 2b to illustrate the calculation of innovation values. The tkth phase of QFD consists of three parts which are Feature-Technology Relevance Matrix (denoted as \({FTR}{M}_{k}\)), Product-Feature Performance in the Customer Requirement Matrix (denoted as \({PF}{{CR}M}_{k}\)), and Technology-Evaluation Matrix (denoted as \({TE}{M}_{k}\)).
\({FTR}{M}_{k}\) describes the relevance between features and key technologies in the slice \({t}_{k}\). Its arrays and columns represent key technologies and features, respectively. Hence, \({FTR}{M}_{k}\) is a \({NT}\times {NF}\) matrix. For the feature \({F}_{i}\) and the key technology \(K{T}_{h}\), their relevance is indicated as \({FTR}{M}_{k}(K{T}_{h},{F}_{i})\), which is calculated using Eq.(14),
$${FTR}{M}_{k}\left(K{T}_{h},{F}_{i}\right)={Norm}\left({N}_{i,h}^{{patent}}\right),$$
(14)
where \({N}_{i,h}^{{patent}}\) means the number of patents containing \(K{T}_{h}\) and the keywords of \({F}_{i}\) \({FTR}{M}_{k}(K{T}_{h},{F}_{i})\in [\mathrm{0,1}]\). The larger the value of \({FTR}{M}_{k}(K{T}_{h},{F}_{i})\), the higher the correlation between \(K{T}_{h}\) and \({F}_{i}\).
\({PF}{CR}{M}_{k}\) describes the performance of various features in meeting customer requirements in the slice \({t}_{k}\). Its arrays and columns represent products and features, respectively. \({PF}{CR}{M}_{k}\) is a \({NP}\times {NF}\) matrix and its value is denoted as \({PF}{CR}{M}_{k}({P}_{j},{F}_{i})\), which is measured via Eq. (15).
$${PF}{{CRM}}_{k}\left({P}_{j},{F}_{i}\right)=\sqrt{{per}{f}_{k}\left({P}_{j},{F}_{i}\right)\times {CRExten}{t}_{k}\left({P}_{j},{F}_{i}\right)}.$$
(15)
\({PF}{{CRM}}_{k}\left({P}_{j},{F}_{i}\right)\) assesses feature performance on satisfying customer requirements from the facets of product feature performance and the ability of features to provoke customer satisfaction. The larger value of \({PF}{CR}{M}_{k}({P}_{j},{F}_{i})\), the better the performance of \({P}_{j}\) in \({F}_{i}\) is in meeting customer requirements. \({PF}{CR}{M}_{k}\) will be further normalized into \([\mathrm{0,1}]\).
\({TE}{M}_{k}\) represents the evaluation results of technology development for various key technologies. \({TE}{M}_{k}\)’s arrays and columns are key technologies and three evaluation criteria (NumTech, CutEdge, and TechCont). For brevity, we use \({E}_{1}\) \({E}_{2}\) and \({E}_{3}\) to denote three criteria. \({TE}{M}_{k}(K{T}_{h},{E}_{m})\) means the evaluation results of \(K{T}_{h}\) on the evaluation criterion \({E}_{m}\). \(m\)= 1, 2, 3.
When identifying PIOs with the multi-phase QFD, we use the results in the previous phase in the calculation of the focal phase. For the tkth phase, if \({F}_{i}\) can better satisfy customer requirements or has higher levels of technology development than the previous phase, \({F}_{i}\) makes progress and is more likely to be a PIO. Thus, we adjust \({PFP}{EM}_{k}\) and \({TE}{M}_{k}\) according to the trends of \({F}_{i}\)’s results. The adjusted \({PFP}{EM}_{k}\) and \({TE}{M}_{k}\) are indicated as \(N{PFP}{EM}_{k}\) and \(N{TE}{M}_{k}\), respectively. They are measured via Eqs. (16), (17). Both \(N{PFP}{EM}_{k}\) and \(N{TE}{M}_{k}\) will be normalized into \([\mathrm{0,1}]\).
$${NPFP}{{EM}}_{k}={PFP}{{EM}}_{k}-{PFP}{{EM}}_{k-1}.$$
(16)
$${NTE}{M}_{k}={TE}{M}_{k}-{TE}{M}_{k-1}.$$
(17)
Based on \({FTR}{M}_{k}\), \(N{PFP}{EM}_{k}\), and \(N{TE}{M}_{k}\), we estimate the innovation value of different features from the perspectives of customer requirements and technology development. From the perspective of customer requirements, we compare \({F}_{{i}_{1}}\) and \({F}_{{i}_{2}}\) using Eq. (18).
$${CRGa}{p}_{k}\left({F}_{{i}_{1}},{F}_{{i}_{2}}\right)=\mathop{\sum }\limits_{j=1}^{{NP}}\left({NPFP}{{EM}}_{k}\left({P}_{j},{F}_{{i}_{1}}\right)-{NPFP}{{EM}}_{k}\left({P}_{j},{F}_{{i}_{2}}\right)\right),$$
(18)
\({CRGa}{p}_{k}\left({F}_{{i}_{1}},{F}_{{i}_{2}}\right)\) is the gap of \({F}_{{i}_{1}}\) and \({F}_{{i}_{2}}\) in meeting customer requirements. If \({CRGa}{p}_{k}\left({F}_{{i}_{1}},{F}_{{i}_{2}}\right) > 0\), \({F}_{{i}_{1}}\) outperforms \({F}_{{i}_{2}}\) in satisfying customer requirements and thus \({F}_{{i}_{1}}\) has more innovation value than \({F}_{{i}_{2}}\).
From the perspective of technology development, we compare \({F}_{{i}_{1}}\) and \({F}_{{i}_{2}}\) using Eq. (19).
$$\begin{array}{c}{TDGa}{p}_{k}\left({F}_{{i}_{1}},{F}_{{i}_{2}}\right)=\\ \mathop{\sum }\limits_{{h}_{1}}^{{NT}}\mathop{\sum }\limits_{{h}_{2}}^{{NT}}\mathop{\sum }\limits_{m=1}^{3}\left({FTRM}\left({F}_{{i}_{1}},K{T}_{{h}_{1}}\right)\cdot {NTEM}\left(K{T}_{{h}_{1}},{E}_{m}\right)-{FTRM}\left({F}_{{i}_{2}},K{T}_{{h}_{2}}\right)\cdot {NTEM}\left(K{T}_{{h}_{2}},{E}_{m}\right)\right),\end{array}$$
(19)
\({TDGa}{p}_{k}\left({F}_{{i}_{1}},{F}_{{i}_{2}}\right)\) is the gap of \({F}_{{i}_{1}}\) and \({F}_{{i}_{2}}\) in the levels of technology development. If \({TDGa}{p}_{k}\left({F}_{{i}_{1}},{F}_{{i}_{2}}\right) > 0\), \({F}_{{i}_{1}}\) has a higher level of technology development than \({F}_{{i}_{2}}\) and thus \({F}_{{i}_{1}}\) has more innovation value than \({F}_{{i}_{2}}\).
By integrating the gaps estimated from two perspectives, we finally estimate the innovation value of \({F}_{i}\) (denoted as \({{IV}}_{k}\left({F}_{i}\right)\)) via Eq. (20),
$${{IV}}_{k}\left({F}_{i}\right)=\mathop{\sum }\limits_{l=1}^{{NF}}{Norm}{\left({CRGa}{p}_{k}\left({F}_{i},{F}_{l}\right)\right)}^{{w}_{1}}\times {Norm}{\left({TDGa}{p}_{k}\left({F}_{i},{F}_{l}\right)\right)}^{{w}_{2}},$$
(20)
where \({w}_{1}\) and \({w}_{2}\) represent the weights of gaps in customer requirements and technology development, respectively \({w}_{1}+{w}_{2}=1\).
Identifying PIOs from the static and dynamic perspectives
After acquiring the innovation value of features, we identify PIOs from the static and dynamic perspectives using the rules in Tables 1, 2. When uncovering dynamic PIOs, We use a predefined parameter \(\delta\) to dynamically determine whether a feature belongs to PIOs. If the number of times that \({F}_{i}\) is recognized as a PIO across all time slices exceeding \(\delta\), \({F}_{i}\) is considered a dynamic PIO. Additionally, we develop targeted product innovation strategies for different PIOs in these tables.
link
