Data-Driven Decision Support System for Vending Machine Planograms Using Machine Learning and Heuristic Algorithms

Sammanfattning

This thesis studies how transaction data from vending machines can be used to support productselection and planogram decisions. Vending machines have limited capacity, while product demand varies across machines, locations, seasons and assortments, so operators face a recurringquestion of which products to keep on the shelf and which to swap out. The aim is to build a datadriven decision support pipeline that estimates product demand in each machine context and usesthese estimates to rank candidate products for both insertion and removal.The empirical work is based on around 7.7 million transaction rows from vending machines operated in Sweden, aggregated into a weekly machine-product panel. Features describe time, location, geography, price, product lifecycle, assortment structure, brand information and cold-startconditions. Light Gradient Boosting Machine (LightGBM) models with a Tweedie objective predictunit demand at one-week and four-week horizons. Separate LambdaRank models rank productsalready on the shelf by their likelihood of underperforming, while insertion candidates are rankedby predicted Tweedie demand. All models are evaluated under walk-forward cross-validation onheld-out folds.The Tweedie regression model achieved a mean absolute error (MAE) of 1.80 units and a coefficient of determination R2 of 0.72 at the one-week horizon, and an MAE of 4.75 units with R2 of0.90 at the four-week horizon. In plain terms, a typical forecast is within about two units of theactual weekly sales of a single product on a single machine at one week, and the model explainsaround 72 percent of the variation in weekly unit sales at one week and around 90 percent at fourweeks. The LambdaRank model identified which of any two products in the same machine andweek would sell more about 77 percent of the time at one week and 81 percent at four weeks,and was sharpest at the top of the inverted list, with median Normalised Discounted CumulativeGain at rank 15 (NDCG@15) values of 0.92 and 0.97, which supports its use for selecting a smallnumber of swap-out candidates per machine.The contribution of this thesis is a complete machine-product-level demand and assortment-rankingpipeline that combines Tweedieregression with LambdaRankranking,evaluatedonarealSwedishvending dataset. The results above demonstrate accurate demand estimates and reliable swaprecommendations at industrial scale on held-out walk-forward folds.