Data-driven product configuration improvement and product line restructuring with text mining and multitask learning

Abstract

In the era of big data, data-driven product configuration improvement and product line restructuring are two important and interrelated problems, and joint decision-making regarding these two problems needs to be tackled. There are two difficulties to achieve the joint decision-making. One is to obtain consumer choice probabilities related to improved product configurations, and the other is to obtain the best product configuration portfolio. In this study, a framework combining text mining and multitask learning is developed to deal with the difficulties. In the framework, improved product configurations are generated using online reviews and transaction data of the target product and its competitors. A one-to-many mapping from customer requirements to improved product configurations is realized by using a multitask support vector machine to obtain consumer choice probabilities for the improved product configurations. The profit maximization models considering the customer choice probabilities are then developed to obtain the best product configuration portfolio. A case study of Huawei P20 series smartphones is used to illustrate the effectiveness of the proposed methods. The results indicate that the multitask support vector machine obtained a higher prediction accuracy than two single-task learning and two other multi-task learning methods, and the proposed framework has the ability to increase the profits produced by the best product configuration portfolio.

Appendices

Appendix 1

Like the standard support vector machine (Chen et al., 2012), the Lagrange multiplier method can be used to obtain the dual of the MT-SVM in Eqs. (7)–(9). The Lagrangian for the MT-SVM is described as:

$$ L\left( {{\mathbf{w}}_{0} ,{\mathbf{v}}_{t} ,a_{it} ,\gamma_{it} } \right) = J\left( {{\mathbf{w}}_{0} ,{\mathbf{v}}_{t} ,\xi_{it} } \right) - \mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{i = 1}^{I} a_{it} \left( {y_{it} \left( {{\mathbf{w}}_{0} + {\mathbf{v}}_{t} } \right) \cdot {\mathbf{x}}_{it} - 1 + \xi_{it} } \right) - \mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{i = 1}^{I} \gamma_{it} \xi_{it} , $$

(14)

where $\alpha_{it}$ and $\gamma_{it}$ are nonnegative Lagrange multipliers. According to the Karush–Kuhn–Tucker (KKT) condition, taking partial derivatives of $L\left( {{\mathbf{w}}_{0} ,{\mathbf{v}}_{t} ,a_{it} ,\gamma_{it} } \right)$ with respect to the primal variables:

$$ \frac{\partial L}{{\partial w_{0} }} = 0 \Rightarrow {\mathbf{w}}_{0} = \frac{1}{{2\lambda_{2} }}\mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{i = 1}^{I} \alpha_{it} y_{it} {\mathbf{x}}_{it} $$

(15)

$$ \frac{\partial L}{{\partial v_{t} }} = 0 \Rightarrow {\mathbf{v}}_{t} = \frac{T}{{2\lambda_{1} }}\mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{i = 1}^{I} \alpha_{it} y_{it} {\mathbf{x}}_{it} $$

(16)

$$ \frac{\partial L}{{\partial \xi_{it} }} = 0 \Rightarrow 1 - \alpha_{it} - \gamma_{it} = 0. $$

(17)

Substituting the results in Eqs. (15)–(17) into (14) leads to the dual formulation of the MT-SVM

$$ {\text{max}}\left\{ {\mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{i = 1}^{I} \alpha_{it} - \frac{1}{2}\mathop \sum \limits_{s = 1}^{S} \mathop \sum \limits_{i = 1}^{I} \mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{j = 1}^{J} \alpha_{is} y_{is} \alpha_{jt} y_{jt} K_{st} \left( {{\mathbf{x}}_{is} ,{\mathbf{x}}_{jt} } \right)} \right\}, $$

(18)

$$ \begin{gathered} {\text{s}}.{\text{t}}.{\mkern 1mu} {\mkern 1mu} {\mkern 1mu} 0 \le \alpha _{{it}} \le C{\mkern 1mu} {\mkern 1mu} {\mkern 1mu} {\text{for}}{\mkern 1mu} {\mkern 1mu} {\mkern 1mu} i \in \left\{ {1,2, \cdots ,I} \right\}{\mkern 1mu} {\mkern 1mu} {\mkern 1mu} {\mkern 1mu} {\text{and}} \hfill \\ t \in \left\{ {1,2, \cdots ,T} \right\}. \hfill \\ \end{gathered} $$

(19)

Let $\alpha_{it}^{*}$ is the optimal solution of the dual problem, the resulting classification. function of the MT-SVM is

$$ f_{t}^{*} \left( x \right) = \mathop \sum \limits_{t = 1}^{T} \mathop \sum \limits_{i = 1}^{I} \alpha_{it}^{*} y_{it} K_{st} \left( {{\mathbf{x}}_{is} ,{\mathbf{x}}} \right). $$

(20)

Appendix 2

See Fig. 4 and Tables (5, 6, 7, 8, 9).