Backorder Prediction Using Machine Learning For Danish Craft Beer Breweries
Author
Li, Yuqi
Term
4. term
Publication year
2017
Submitted on
2017-09-02
Abstract
Denne afhandling undersøger, hvordan maskinlæring kan bruges til at forudsige backorders (restordrer) i danske mikrobryggerier, der oplever stigende efterspørgsel efter specialøl efter afgiftslettelser. Når bryggerier med begrænset kapacitet planlægger ud fra erfaring, kan de få svært ved at imødekomme udsving i markedet, hvilket øger risikoen for restordrer og bullwhip-effekter i forsyningskæden. Projektet opstiller en praktisk arbejdsgang: fra forståelse af konteksten og databehov, over forbehandling af historiske data (binarisering, håndtering af manglende værdier, ubalancerede klasser, outliers og redundans samt dimensionalitetsreduktion med PCA), til modellering og evaluering. Med et relevant open source-datasæt screenes flere supervisionerede klassifikatorer, herunder Naive Bayes, supportvektormaskiner (med kerner), beslutningstræer, k-nærmeste naboer og ensemblemetoder, og der designes forsøg med to træningssæt (rensede vs. balancerede data). Modellerne vurderes med bl.a. k-fold cross-validation, konfusionsmatricer, sensitivitet/specificitet og AUC. Screening peger på KNN, ensemble-KNN og SVM med Gaussisk kerne som de bedste kandidater; i én opsætning nåede et ensemble-subspace KNN omtrent 83 % nøjagtighed. Resultater rapporteres på tværs af forbehandlingsvalg og sammenholdes med en traditionel tommelfingerregel-baseret tilgang. Arbejdet giver anbefalinger til prognoser og et grundlag, som danske bryggerier kan bruge til at forbedre planlægning og reducere risikoen for restordrer, med forbehold for, at yderligere bryggerispecifikke data kan styrke anvendelsen i praksis.
This thesis examines how machine learning can be used to predict backorders in Danish craft breweries facing rising demand for specialty beers following tax reductions. For small producers with limited capacity, experience-based planning may fail to absorb demand volatility, increasing the risk of backorders and bullwhip effects in the supply chain. The project proposes a practical workflow: from contextualization and data requirements, through preprocessing of historical data (binarization, handling missing values, class imbalance, outliers, and redundancy, plus dimensionality reduction with PCA), to modeling and evaluation. Using a relevant open-source dataset, multiple supervised classifiers are screened, including Naive Bayes, support vector machines (with kernels), decision trees, k-nearest neighbours, and ensemble methods, with experiments designed around two training sets (cleaned versus balanced data). Models are assessed with k-fold cross-validation, confusion matrices, sensitivity/specificity, and AUC. Screening identifies KNN, ensemble KNN, and Gaussian-kernel SVM as top performers; in one configuration an ensemble subspace KNN achieved about 83% accuracy. Results are reported across preprocessing choices and compared with a traditional rule-of-thumb approach. The work offers forecasting recommendations and a baseline that Danish breweries can use to improve planning and reduce the risk of backorders, noting that additional brewery-specific data would further support practical deployment.
[This summary has been generated with the help of AI directly from the project (PDF)]
Keywords
Documents