|
We investigate a data-driven single-period inventory management problem with uncertain demand, where large amounts of textual online reviews and historical data are accessible. Unlike two-step frameworks (i.e., predict-then-optimization), we propose an end-to-end (E2E) framework that directly suggests the order quantity by leveraging a deep learning model that inputs textual online reviews and other demand-related feature data , without any intermediate steps such as text sentiment analysis. The E2E model does not require any prior assumptions about the demand distribution and can automatically determine the order quantity that minimizes the newsvendor cost by employing the information from real-world data. Our experiments, using publicly available real-world data, demonstrate that our method can significantly reduce the sum of overage and underage costs, outperforming other data-driven models proposed in recent years. Specifically, the inclusion of textual online review data improves ordering decisions by a 28.7% cost reduction. Introduction The data-driven newsvendor problem has become a hot topic in management science in recent years (Ban and Rudin, 2019; Huber et al., 2019; Lin et al., 2021; Liu et al., 2022a; Oroojlooyjadid et al., 2019; Pirayesh Neghab et al., 2022). The newsvendor problem denotes a single-period inventory management problem in which the demand is uncertain (Arrow et al., 1951). If the ordered quantity exceeds the actual demand, overage cost occurs; conversely, underage cost is incurred. Generally, the unit overage cost is not equal to the unit underage cost. If the probability distribution of the uncertain demand is known, the optimal order quantity can be obtained through classical theory. However, managers cannot clearly determine the probability distribution of the random demand, either in the form of distribution or in specific parameters. Furthermore, consumer preferences and market conditions are subject to change over time, resulting in different probability distributions of demand observations across various inspection time ranges. The development of the Internet, the application of big data technology, and the maturity of artificial intelligence (AI) have paved the way for new solutions in inventory management. Increasing the availability of large consumption datasets generated by users on online retail platforms has the potential to improve the decision-making effectiveness of ordering goods in practical applications (Maheshwari et al., 2020). User-generated data contain a wealth of features concerning demand and facilitate better decisions based on real information (Babai et al., 2021; Kamble and Gunasekaran, 2019; Kuo and Kusiak, 2018). Studies on the data-driven newsvendor problem aim to utilize demand-related features and appropriate methodologies to tackle the challenge of demand uncertainty and determine the optimal order quantity. Data-driven methods are nonparametric approaches that rely on actual data and do not require any assumptions about the demand distribution. Previous studies have demonstrated that data-driven methods can enhance the accuracy of order quantity estimation, thereby reducing costs and increasing profitability for enterprises in highly competitive business environments (Ban and Rudin, 2019; Oroojlooyjadid et al., 2019; Pirayesh Neghab et al., 2022; Qi et al., 2020, 2023). Existing studies on the data-driven newsvendor problem only utilize structured feature data, which can be stored and managed in a fixed format with clear data structures and data types, such as climate temperature, price indices, unemployment rates, and dates (Ban and Rudin, 2019; Oroojlooyjadid et al., 2019; Pirayesh Neghab et al., 2022). However, the information contained in structured data is limited and may not fully reflect changes in market demand. In contrast, unstructured data refer to data that lack a fixed format and are not easily represented or processed with clear structures or standardized metadata. Unstructured data typically include text, images, audio, video, etc. In the era of big data, unstructured data are more common and contain massive amounts of information. Textual online review data play a crucial role in business forecasts and decision-making, as they reflect customers' perceptions of product quality and satisfaction levels. Online reviews provide valuable reference information for other users' consumption behavior (Qiu et al., 2018; Vana and Lambrecht, 2021). Users voluntarily publish online reviews expressing rich emotions (e.g., happiness, anger, sadness, criticism, and praise), and reflecting reviewers' real experiences of using products in various real situations. As a result, online reviews represent a valuable resource of user-oriented information (Chen and Xie, 2008). Empirical studies confirmed that implied information in textual online reviews can significantly influence the product demand, leading to more accurate demand forecasts and reducing the demand uncertainty (Chong et al., 2016; Cui et al., 2017; Duan et al., 2008; Kim and Allenby, 2022; Schneider and Gupta, 2016; Ye et al., 2009; C. Zhang et al., 2022a). Additionally, online reviews include information about product descriptions and consumer preferences for product attributes, which may further affect future customer demand. The utilization of online review data to support operational decisions presents noteworthy research value. This paper aims to explore whether integrating online reviews into inventory management can assist managers in making more informed ordering decisions. However, textual reviews cannot be used directly in traditional decision-making models, and managing unstructured text data is a major challenge. Consequently, devising a novel method for inventory decision-making that can effectively integrate textual data is necessary. In this study, we propose an end-to-end (E2E) framework that directly outputs the order quantity directly from various input feature data without any intermediate steps (Qi et al., 2023). The E2E approach differs from the separated estimation and optimization (SEO) method (Elmachtoub and Grigas, 2022). Fig. 1 illustrates the decision-making processes for the two methods. The SEO method estimates and predicts the demand distribution from historical data as an intermediate result, and then uses this intermediate result to calculate the final optimal order quantity. In contrast, the E2E method integrates the estimation and optimization processes generally using an AI model that directly estimates the suggested order quantity from the original data. Our proposed E2E method unifies demand estimation and inventory optimization in one deep learning model. Existing studies have demonstrated that the E2E framework generally outperforms SEO in decision-making tasks (Liu et al., 2022a; Oroojlooyjadid et al., 2019; Pirayesh Neghab et al., 2022; Qi et al., 2023). The reason may be that SEO separates the estimation stage from optimization stages, and the objectives of the two stages are inconsistent. Consequently, useful information may be lost substantially in the post-prediction optimization process, and the error in the estimation stage is amplified in the optimization stage, greatly affecting the optimization results. The proposed E2E deep learning framework consists of a word embedding submodule, a bidirectional gated recurrent unit network (BiGRU) submodule, a multi-layer perceptron network (MLP) submodule, and full connection layers. For textual reviews, the framework does not conduct sentiment analysis but instead utilizes the term frequency-inverse document frequency (TF-IDF) method firstly to identify and encode important words that are useful for decision-making from the review documents. Subsequently, these important words are converted into a dense matrix by training the word embedding layer. Finally, the dense matrix is transmitted to the BiGRU networks to directly mine the implicit relationship between the feature information and the future demand. In addition to textual reviews, structured feature data that are important for demand forecasting, such as search trends (D'Amuri and Marcucci, 2017; Li et al., 2018; Niesert et al., 2020; Yu et al., 2019) and macroeconomic indicators (Zhang et al., 2020a,-c), are also utilized. First, these structured features are processed by the MLP submodule, and then the output neurons of the MLP submodule are concatenated with the neurons output from the textual review processing submodule. Our deep learning framework is a multi-input multi-task framework: inputs include both unstructured review texts and structured features, and outputs are the suggested order quantity and the demand forecasting value. The demand forecasting serves as an auxiliary task that guides faster model training, monitors model performance, and improves the model's generalizability to inhibit overfitting (Qi et al., 2023). Our proposed framework offers several advantages in three main aspects. Firstly, the framework avoids the significant information loss caused by sentiment classification errors because this framework does not perform sentiment analysis. Moreover, our framework eliminates the need for manual tagging of large amounts of online review data and improves the decision-making efficiency as building and training a sentiment classification model is unnecessary. Secondly, processing a review text as a time series of words requires an approach that can handle long intervals and delay information. As a recurrent neural network, the BiGRU network is well-suited for this task, and its bidirectional nature enables the context information of the text to be taken into account simultaneously, improving the framework's performance. Thirdly, the multi-input multi-task E2E deep learning model has excellent generalizability and information mining abilities, with the obtained order quantity capable of reducing newsvendor cost for enterprises. Our experiments, conducted using public real-world data, show that our method outperforms benchmark methods proposed in recent years, with a 31.8% lower average cost compared to the benchmark model with the best performance. Additionally, the inclusion of textual online review data improves ordering decisions by a 28.7% cost reduction. Appendix F further verifies the robustness of our method through comparative experiments with additional products. Our study differs from existing studies in the following three aspects. Firstly, we propose an E2E model that can process online textual review data. This novel model extends existing data-driven newsvendor models, which have limitations in handling unstructured data. Our model eliminates the need for separate sentiment analysis of online reviews, avoiding errors and information loss during intermediate processing. Secondly, we innovate the design of a deep neural network architecture and construct a multi-input multi-task network structure. This AI structure fully capitalizes on the strengths of multi-layer neural networks for fitting nonlinear relationships, word embedding layer for mining important features from text content, and BiGRU for capturing word sequence information. This multi-task structure improves the generalizability of the model and achieves sound performance. Lastly, we demonstrate the superior performance of our method relative to multiple benchmark methods through extensive experiments with real-world data. Additionally, Appendix E provides a comprehensive sensitivity analysis. The remainder of this paper is organized as follows. Section 2 summarizes the related literature in recent years. Section 3 describes the detailed framework of our proposed method. Section 4 reports on the experiments with real-world data, including the data, experiment design, results and discussions. Finally, Section 5 provides the conclusions of the study and outlines future prospects. Additionally, Appendix A explains the principles of bidirectional GRU and Adam optimizer, Appendix B provides detailed descriptive statistics about the data used, Appendix C provides the selection method for the structured features, and Appendix D presents the trainable parameters and training time for the machine learning models used in our experiments. Section snippets Data-driven newsvendor problem Data-driven approaches to inventory decisions have become a frontier academic topic in recent years. The newsvendor problem, a classic single-period inventory management problem, serves as a starting point of data-driven research (Qi et al., 2020). Traditional methods for solving the newsvendor problem involve estimating or forecasting the demand distribution before optimization (separated estimation and optimization, SEO), where the demand is assumed to adhere to a distribution family with Problem description For a given product, let c b be the unit underage cost, and c h be the unit overage cost. The cost function for period t can be expressed as: C t = c h ( q t − d t ) + + c b ( d t − q t ) + . The available data about this product include the structured feature data (e.g., the search traffic data related to the product, and the price of raw materials for producing this product. Appendix C provides the specific feature selection method) and the unstructured online review data. The dataset can be denoted as D = { ( x 1 , R 1 ) , … , ( x T , R Experiment with real-world data To evaluate the performance of our E2ETF model and benchmark models, we utilized real-world automobile product data because of data availability. These data can be used to proxy for other types of online products as they possess similar feature-demand relationships. We utilized PyTorch (Paszke et al., 2017) to implement all neural network models and trained them on a computer with an NVIDIA GeForce RTX 3060 GPU. The Adam optimizer was employed (Kingma and Ba, 2015), and its specific principle Conclusions In this paper, we propose a multi-input, multi-task E2E deep learning method for the data-driven newsvendor problem with access to unstructured online review data and structured feature data. Our method can automatically extract key features for inventory decisions from textual reviews without requiring sentiment analysis. This approach enables enterprises to make optimal inventory decisions in uncertain market environments, effectively reducing costs and increasing profits. Through numerous Credit author statement Yu-Xin Tian : Conceptualization, Methodology, Software, Visualization, Writing – original draft, Validation, Writing – review & editing. Chuan Zhang : Supervision, Conceptualization, Funding acquisition, Writing – review & editing. Acknowledgements This work was supported by the National Social Science Fund of China ( 19BGL229 ). References (79) Z. Ahmad et al. Borrow from rich cousin: transfer learning for emotion detection using cross lingual embedding Expert Syst. Appl. (2020) Y. Cao et al. Quantile forecasting and data-driven inventory management under nonstationary demand Oper. Res. Lett. (2019) Y.-C. Chang et al. Predicting aspect-based sentiment using deep learning and information visualization: the impact of COVID-19 on the airline industry Inf. Manag. (2022) F. D'Amuri et al. The predictive power of Google searches in forecasting US unemployment Int. J. Forecast. (2017) W. Duan et al. Do online reviews matter? - an empirical investigation of panel data Decis. Support Syst. (2008) J. Huber et al. A data-driven newsvendor problem: from data to decision Eur. J. Oper. Res. (2019) D. Kim et al. Multi-co-training for document classification using various document representations: TF–IDF, LDA, and Doc2Vec Inf. Sci. (2019) B. Kratzwald et al. Deep learning for affective computing: text-based emotion recognition in decision support Decis. Support Syst. (2018) D. Li et al. Interactive double states emotion cell model for textual dialogue emotion prediction Knowl. Base Syst. (2020) S. Li et al. Effective tourist volume forecasting supported by PCA and improved BPNN using Baidu index Tourism Manag. (2018) C. Liu et al. Newsvendor problems: an integrated method for estimation and optimisation Eur. J. Oper. Res. (2022) L.H. Liyanage et al. A practical inventory control policy using operational statistics Oper. Res. Lett. (2005) R.F. Niesert et al. Can Google search data help predict macroeconomic series? Int. J. Forecast. (2020) S. Peng et al. A survey on deep learning for textual emotion analysis in social networks Digital Communications and Networks (2022) D. Pirayesh Neghab et al. An integrated data-driven method using deep learning for a newsvendor problem with unobservable features Eur. J. Oper. Res. (2022) J. Qiu et al. Leveraging sentiment analysis at the aspects level to predict ratings of reviews Inf. Sci. (2018) M.J. Schneider et al. Forecasting sales of new and existing products using consumer reviews: a random projections approach Int. J. Forecast. (2016) X. Xu et al. Utilizing the platform economy effect through EWOM: does the platform matter? Int. J. Prod. Econ. (2020) X. Yang et al. Forecasting Chinese tourist volume with search engine data Tourism Manag. (2015) Q. Ye et al. The impact of online user reviews on hotel room sales Int. J. Hospit. Manag. (2009) L. Yu et al. Online big data-driven oil consumption forecasting with Google trends Int. J. Forecast. (2019) C. Zhang et al. Forecasting sales using online review and search engine data: a method based on PCA–DSFOA–BPNN Int. J. Forecast. (2022) C. Zhang et al. Forecasting the box offices of movies coming soon using social media analysis: a method based on improved Bass models Expert Syst. Appl. (2022) M. Zhang et al. Using neutral sentiment reviews to improve customer requirement identification and product design strategies Int. J. Prod. Econ. (2022) W. Zhang et al. A comparative study of TF*IDF, LSI and multi-words for text classification Expert Syst. Appl. (2011) M.S. Akhtar et al. How intense are you? Predicting intensities of emotions and sentiments using stacked ensemble [application notes] IEEE Comput. Intell. Mag. (2020) F. Almeida et al. Word Embeddings: A Survey (2019) S. Anbaee Farimani et al. Investigating the informativeness of technical indicators and news sentiment in financial market price prediction Knowl. Base Syst. (2022) N. Archak et al. Deriving the pricing power of product features by mining consumer reviews Manag. Sci. (2011) K.J. Arrow et al. Optimal inventory policy Econometrica (1951) M.Z. Babai et al. Demand forecasting in supply chains: a review of aggregation and hierarchical approaches Int. J. Prod. Res. (2021) G.Y. Ban et al. The big Data newsvendor: practical insights from machine learning Oper. Res. (2019) D. Bertsimas et al. A Data-Driven Approach to Newsvendor Problems (2005) J. Bu et al. Offline pricing and demand learning with censored data Manag. Sci. (2023) B. Chen et al. Technical note—data-based dynamic pricing and inventory control with censored demand and limited price changes Oper. Res. (2020) K. Chen et al. Turning from TF-IDF to TF-IGM for term weighting in text classification Expert Syst. Appl. (2016) Y. Chen et al. Online consumer review: word-of-mouth as a new element of marketing communication mix Manag. Sci. (2008) A.Y.L. Chong et al. Predicting online product sales via online reviews, sentiments, and promotion strategies: a big data architecture and neural network approach Int. J. Oper. Prod. Manag. (2016) J. Chung et al. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling (2014) View more references Cited by (0) Recommended articles (6) Research article How economic depreciation shapes the relationship of uncertainty with investments’ size & timing International Journal of Production Economics, Volume 260, 2023, Article 108836 Show abstract This paper identifies and analyzes the effects of the rate of economic depreciation of capital stock on a monopolist’s investment option and capacity decision in a dynamic and uncertain market environment, where continuous economic depreciation cannot be fully offset. We find that the firm’s capital stock is increasing in the rate of depreciation for low rates and decreasing for higher rates. Further, when considering the timing of investment, we show that the effect of uncertainty on investment is level-dependent on the rate of depreciation: only for sufficiently high rates of depreciation there is a positive relationship between capital investment and uncertainty, and the impact of uncertainty on the present value of the firm is mitigated. The fact that the impact of economic depreciation on the firm’s investment problem is level dependent demonstrates that its consideration by investors and managers is not trivial. Research article Data pooling for multiple single-component systems under population heterogeneity International Journal of Production Economics, Volume 250, 2022, Article 108665 Show abstract We consider multiple newly designed single-component systems with a known finite lifespan. The component in each system can be replaced preventively to avoid a costly failure. This component is also in use for the first time, and therefore, there is no historical data on the lifetime of the component. In this case, the probability distribution of the component lifetime can be estimated based on expert opinions. However, there can be different opinions on the lifetime distribution. There are two populations where components can come from: a weak and a strong population. We assume that the components always come from the same population. However, the true type of the population is unknown. We build a discrete-time partially observable Markov decision process model to find the optimal replacement policy which minimizes the expected total cost throughout the lifespan. To resolve the uncertainty regarding population heterogeneity, we update the belief by using the data collected from all systems, allowing us to investigate the effect of so-called data pooling. First, we generate insights about the structure of the optimal policy. We then compare the cost per system under the optimal policy with the cost per system under two benchmark heuristics that follow the single-system optimal policy with and without data pooling, respectively. In our numerical experiments, we show that the cost reduction relative to the worst benchmark heuristic can be up to 5.6% for data pooling with two systems, and this increases up to 14% for 20 systems. Additionally, the effect of various input parameters on the costs is analyzed. Research article A novel hybrid assembly method to reduce operational costs of selective assembly International Journal of Production Economics, Volume 264, 2023, Article 108966 Show abstract Selective assembly is a technique whereby components are smartly matched depending on their measured feature values. Although it can reduce scrap by allowing more relaxed tolerances on components when compared to traditional random assembly, it is often not adopted due to the high related operational costs. Therefore, this study proposes a novel hybrid strategy whereby a number of traditional assembly rounds are performed before switching to selective assembly. At each round, the components which did not lead to a feasible assembly are put into their respective unmatched piles and are used for the next traditional assembly round. When N such rounds have been performed, the remaining unmatched components will finally be assembled via selective assembly. To determine the optimal number of traditional assembly rounds before moving on to selective assembly, a cost function depending on the matching probabilities at each round and the relative costs of selective and traditional assembly is developed together with an optimality condition. Since the analytic calculation of the matching probabilities is in most of the practical cases infeasible, these are estimated using simulation. Particular attention is paid to reducing the variance of the estimators for the matching probabilities, and to their impact on the total variability of the cost function. The application of this innovative method to two case studies demonstrates that the proposed hybrid technique can significantly reduce operational costs of selective assembly, while still taking advantage of its typical benefits. Research article Collaborative decision-making based on effort level: power battery recycling alliance Procedia Computer Science, Volume 221, 2023, pp. 1155-1161 Show abstract As new energy vehicles continue to advance, the sphere pertaining to power battery recycling is experiencing steady growth. This has sparked widespread discussion and attention regarding the formation of an alliance dedicated to power battery recycling. In order to establish a closed-loop supply chain (CLSC), and faster the amalgamation within the realm of battery recycling, this research examines the influence of the extent of recycling efforts on power battery demand in new energy vehicles. By utilizing a Stackelberg game, the consideration is given to the inclusion of a power battery manufacturer, a new energy vehicle enterprise, and a third-party recycler. The examination revolves around the consequences of various recycling alliance modes about decision-making and profits within the CLSC. The findings indicate that it is more advantageous for the power battery manufacturer to establish a recycling alliance with a new energy enterprise and recycler rather than establishing a separate alliance with a third-party, as this approach enhances the benefits to the battery manufacturer. In other words, the presence of multiple alliance members in the recycling realm positively correlates with the overall profitability of the CLSC. The power battery manufacturer should establish appropriate recycling sensitivity coefficients to incentivize enterprises to intensify recycling efforts and augment the recycled volume of used power batteries. Furthermore, the benefits of CLSC can be amplified by government subsidies, and the distribution ratio of these subsidies can affect the choice of a recycling alliance model for CLSC members. Research article Installment for advance payment and pricing decisions for an inventory system under power demand pattern and all-units discount International Journal of Production Economics, 2023, Article 108951 Show abstract Advance payment method has emerged as an effective mechanism to control not only any company's cash flow but also order cancellations. To encourage customers to use the prepayment mechanism, companies offer a reduction opportunity on the products per unit purchase cost based on the purchased quantity. This work analyzes a retailer's joint installment for advance payment, pricing, and inventory policies to maximize profit under a joined order-sensitive advance payment and all-units discount system. To make the model more representative, the demand is incorporated as a multiplicative form of the effects of selling price as logit and time as power patterns, and shortages are also taken into account during the model's formulation. Theoretical expressions are derived to identify whether the inventory process is unprofitable under a joint link-to-order advance payment and all-units discount arrangement. Utilizing all the derived analytical expressions, an efficacious algorithm is built to attain the optimal installment frequency for prepayment, selling price, order quantity, and shortage amount for the inventory manager. To further study the effects of a combined link-to-order prepayment and all-units discount scheme, the outcomes of several numerical examples are documented. Finally, an exploration is carried out by varying the values of all demand parameters and inventory cost parameters under or in the absence of the combined link-to-order prepayment and all-units discount scheme to provide useful guidelines for the industry manager. A noteworthy and advantageous finding for the inventory management practitioner is that by adopting a linked-to-order prepayment and discount system, certain inventory strategies previously considered unprofitable in the absence of quantity-based advance payment and discount programs have the potential to become profitable. To minimize the cost related to prepayment amounts, the inventory manager should increase the optimal installment number when the installment cost is low but decrease it when the installment cost is high. Moreover, to mitigate the impact of high price sensitivity on demand, the practitioner ought to reduce the optimal selling price and decrease the order quantity per cycle, thereby maintaining customer demand. Research article The optimisation of the location of front distribution centre: A spatio-temporal joint perspective International Journal of Production Economics, Volume 263, 2023, Article 108950 Show abstract Front Distribution Centre (FDC) is a new terminal warehouse which is closer to customers, with its location selection being crucial for e-commerce and customer time satisfaction. We introduce in this paper a joint distribution function of demand based on time and space, which constructs two spatio-time models: spatio-time clustering model and spatio-time optimisation model. A staged clustering algorithm is designed to obtain the candidate FDCs, and an intelligent algorithm based on NSGA-Ⅱ (Non-dominated Sorting Genetic Algorithm II) is applied to determine the final FDCs, in which the location selection problem is formulated as a bi-objective programming model to minimise total costs and maximise customer time satisfaction. Our results indicate that the model considering spatio-temporal joint attribute of demand performs better than the traditional spatial model. Furthermore, when compared with the k-means clustering algorithm, Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) and its improved algorithm Multi-Objective Evolutionary Algorithm based on the Adaptive Neighborhood Adjustment strategy (MOEA/D-ANA), Multi-Objective Particle Swarm Optimisation” (MOPSO) and its enhancing algorithm Competitive Multi-Objective Particle Swarm Optimiser (CMOPSO), the solving method based on staged clustering and NSGA-II absolutely performs more stable and can get a greater number of pareto-optimal solutions with higher qualities. Especially when compared with K-means clustering algorithms, it can reduce total costs by up to 38.84% and improve customer time satisfaction by up to 36.22%. View full text © 2023 Elsevier B.V. All rights reserved. About ScienceDirect Remote access Shopping cart Advertise Contact and support Terms and conditions Privacy policy We use cookies to help provide and enhance our service and tailor content and ads. By continuing you agree to the use of cookies . Copyright © 2023 Elsevier B.V. or its licensors or contributors. ScienceDirect® is a registered trademark of Elsevier B.V. ScienceDirect® is a registered trademark of Elsevier B.V.
|
