銷售預測對於商業以數據科學的方式進行決策至關重要，由於內外部環境因素的影響，銷售預測是一個具有挑戰性的問題，特別是產品屬於生命週期短的產業。以麵包店為例，由於產品具有易腐及保存期限短的性質，預測每天的需求量對麵包店來說很困難但是重要。雖然傳統上基於統計模型的方法可以做預測，但預測準確度仍不夠滿意。近日機器學習方法應用在銷售預測上雖能得到較好的預測準確度，但還需要仰賴專業領域知識與人為的特徵工程。本研究基於卷積類神經網路發展了一個創新的方法，讓模型能以高效率的方式將銷售數據圖形化後學習並預測。產品每小時的銷售額將以單位像素的形式作為圖形的一部分，而完整圖形可以看出當天銷售的整體情形或分布，且此圖形將會是卷積類神經網路的輸入值。本研究以一家真實位於中國杭州的麵包企業所提供之銷售資料為模型驗證範例，並和多層感知器演算法比較。結果顯示本研究提供之方法錯誤率上顯著的優於多層感知器演算法。

Sales forecast is extremely important for business in making decision with a data-scientific manner. The sales forecast is a challenging problem due to the influence of internal and external environments especially in a business that has short life-cycle. For example, forecasting a daily demand in bakery business is crucial due to the perishable products with short preservation time. Traditional sales forecasting method based on the statistical model have been proved to be effective in forecasting. However, their performances are still far from satisfying. The more recent machine learning approach in sales forecasting have better performance but it requires manual feature engineering and domain knowledge. In this thesis, a novel approach using sales transformation with convolutional neural networks (CNN) to learn the sales as images and predict the sales with a high performance. The sales per hour of a product is represented as a pixel of an image. The generated image containing the sales pattern of the products in a daily based is used as an input of CNN. The effectiveness of this method is evaluated on a large real-world bakery sales dataset in Hang Zhou, China and compared with deep feedforward networks or multi-layer perceptron. The results show that the proposed method significantly outperforms deep feedforward network with a lower error rate.

[1] Kui Zhao and Can Wang. “Sales Forecast in E-commerce using Convolu-tional Neural Network”. In:arXiv:1708.07946 [cs, stat](Aug. 2017). arXiv:1708.07946.url:http://arxiv.org/abs/1708.07946(visited on 11/22/2017).[2] Samaneh Beheshti-Kashi et al. “A survey on retail sales forecasting and predic-tion in fashion markets”. In:Systems Science & Control Engineering3.1 (Dec.2014), pp. 154–161.doi:10.1080/21642583.2014.999389.url:https://doi.org/10.1080%2F21642583.2014.999389.[3] Tak-chung Fu. “A review on time series data mining”. In:Engineering Ap-plications of Artificial Intelligence24.1 (Feb. 2011), pp. 164–181.doi:10.1016/j.engappai.2010.09.007.url:https://doi.org/10.1016%2Fj.engappai.2010.09.007.[4] Rob J. Hyndman Spyros G. Makridakis Steven C. Wheelwright.Forecasting:Methods and Applications, 3rd Edition. Wiley, 1998.[5] Zhan-Li Sun et al. “Sales forecasting using extreme learning machine withapplications in fashion retailing”. In:Decision Support Systems46.1 (Dec.2008), pp. 411–419.doi:10.1016/j.dss.2008.07.009.url:https://doi.org/10.1016%2Fj.dss.2008.07.009.[6] Bogdan Oancea and fffdfffdtefan Cristian Ciucu. “Time series forecasting usingneural networks”. In:arXiv preprint arXiv:1401.1333(2014).[7] Min Xia and W.K. Wong. “A seasonal discrete grey forecasting model forfashion retailing”. In:Knowledge-Based Systems57 (Feb. 2014), pp. 119–126.doi:10.1016/j.knosys.2013.12.014.url:https://doi.org/10.1016%2Fj.knosys.2013.12.014.53
[8] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. “Deep learning”. In:Na-ture521.7553 (May 2015), pp. 436–444.doi:10.1038/nature14539.url:https://doi.org/10.1038%2Fnature14539.[9] Nima Hatami, Yann Gavet, and Johan Debayle. “Classification of Time-SeriesImages Using Deep Convolutional Neural Networks”. In:arXiv preprint arXiv:1710.00886(2017).[10] Tao Huang, Robert Fildes, and Didier Soopramanien. “The value of compet-itive information in forecasting FMCG retail product sales and the variableselection problem”. In:European Journal of Operational Research237.2 (Sept.2014), pp. 738–748.doi:10.1016/j.ejor.2014.02.022.url:https://doi.org/10.1016%2Fj.ejor.2014.02.022.[11] Shaohui Ma, Robert Fildes, and Tao Huang. “Demand forecasting with highdimensional data: The case of SKU retail sales forecasting with intra- andinter-category promotional information”. In:European Journal of OperationalResearch249.1 (Feb. 2016), pp. 245–257.doi:10.1016/j.ejor.2015.08.029.url:https://doi.org/10.1016%2Fj.ejor.2015.08.029.[12] Jean-Louis Bertrand, Xavier Brusset, and Maxime Fortin. “Assessing andhedging the cost of unseasonal weather: Case of the apparel sector”. In:Eu-ropean Journal of Operational Research244.1 (July 2015), pp. 261–276.doi:10.1016/j.ejor.2015.01.012.url:https://doi.org/10.1016%2Fj.ejor.2015.01.012.[13] David J. Olive. “Multiple Linear Regression”. In:Linear Regression. SpringerInternational Publishing, 2017, pp. 17–83.doi:10.1007/978-3-319-55252-1_2.url:https://doi.org/10.1007%2F978-3-319-55252-1_2.[14] Zhaoxia Guo. “A Neural Network-Based Forecasting Model for UnivariateSales Forecasting”. In:Intelligent Decision-making Models for Production and54
Retail Operations. Springer Berlin Heidelberg, 2016, pp. 261–293.doi:10.1007/978-3-662-52681-1_10.url:https://doi.org/10.1007%2F978-3-662-52681-1_10.[15] Zhen-Yao Chen and R. J. Kuo. “Evolutionary Algorithm-Based Radial BasisFunction Neural Network Training for Industrial Personal Computer SalesForecasting”. In:Computational Intelligence33.1 (Sept. 2015), pp. 56–76.doi:10.1111/coin.12073.url:https://doi.org/10.1111%2Fcoin.12073.[16] Xue-Feng Jiang. “The research on sales forecasting based on rapid BP neuralnetwork”. In:2012 International Conference on Computer Science and In-formation Processing (CSIP). IEEE, Aug. 2012.doi:10.1109/csip.2012.6309083.url:https://doi.org/10.1109%2Fcsip.2012.6309083.[17] Esteban Alfaro et al. “Bankruptcy forecasting: An empirical comparison ofAdaBoost and neural networks”. In:Decision Support Systems45.1 (Apr.2008), pp. 110–122.doi:10.1016/j.dss.2007.12.002.url:https://doi.org/10.1016%2Fj.dss.2007.12.002.[18] Kanad Chakraborty et al. “Forecasting the behavior of multivariate time seriesusing neural networks”. In:Neural Networks5.6 (Nov. 1992), pp. 961–970.doi:10.1016/s0893-6080(05)80092-9.url:https://doi.org/10.1016%2Fs0893-6080%2805%2980092-9.[19] “Backpropagation in time-series forecasting”. In:Long Range Planning28.6(Dec. 1995), p. 123.doi:10.1016/0024-6301(95)99962-y.url:https://doi.org/10.1016%2F0024-6301%2895%2999962-y.[20] Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning.http://www.deeplearningbook.org. MIT Press, 2016.55
[21] Teijiro Isokawa, Haruhiko Nishimura, and Nobuyuki Matsui. “QuaternionicMultilayer Perceptron with Local Analyticity”. In:Information3.4 (Nov.2012), pp. 756–770.doi:10.3390/info3040756.url:https://doi.org/10.3390%2Finfo3040756.[22] Mariette Awad and Rahul Khanna. “Deep Neural Networks”. In:EfficientLearning Machines. Apress, 2015, pp. 127–147.doi:10.1007/978-1-4302-5990-9_7.url:https://doi.org/10.1007%2F978-1-4302-5990-9_7.[23] Sangeeta Vhatkar and Jessica Dias. “Oral-Care Goods Sales Forecasting UsingArtificial Neural Network Model”. In:Procedia Computer Science79 (2016),pp. 238–243.[24] Alain Yee Loong Chong et al. “Predicting online product sales via online re-views, sentiments, and promotion strategies: A big data architecture and neu-ral network approach”. In:International Journal of Operations & ProductionManagement36.4 (2016), pp. 358–383.[25] Hong-Ze Li et al. “A hybrid annual power load forecasting model based ongeneralized regression neural network with fruit fly optimization algorithm”.In:Knowledge-Based Systems37 (2013), pp. 378–387.[26] Youqin Pan, Terrance Pohlen, and Saverio Manago. “Hybrid neural networkmodel in forecasting aggregate US retail sales”. In:Advances in business andmanagement forecasting. Emerald Group Publishing Limited, 2013, pp. 153–170.[27] Yu Wei and Mu-Chen Chen. “Forecasting the short-term metro passenger flowwith empirical mode decomposition and neural networks”. In:TransportationResearch Part C: Emerging Technologies21.1 (2012), pp. 148–162.56
[28] Chun-Fu Chen, Ming-Cheng Lai, and Ching-Chiang Yeh. “Forecasting tourismdemand based on empirical mode decomposition and neural network”. In:Knowledge-Based Systems26 (2012), pp. 281–287.[29] Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning.http://www.deeplearningbook.org. MIT Press, 2016.[30] Kunihiko Fukushima. “Neocognitron: A hierarchical neural network capable ofvisual pattern recognition”. In:Neural Networks1.2 (Jan. 1988), pp. 119–130.doi:10.1016/0893-6080(88)90014-7.url:https://doi.org/10.1016%2F0893-6080%2888%2990014-7.[31]CS231n Convolutional Neural Networks for Visual Recognition. Accessed onSat, December 16, 2017.url:http://cs231n.github.io/convolutional-networks/#norm.[32] Yisheng Lv et al. “Traffic flow prediction with big data: a deep learning ap-proach”. In:IEEE Transactions on Intelligent Transportation Systems16.2(2015), pp. 865–873.[33] Xiaolei Ma et al. “Learning traffic as images: a deep convolutional neuralnetwork for large-scale transportation network speed prediction”. In:Sensors17.4 (2017), p. 818.[34] Stanislas Chambon et al. “A deep learning architecture for temporal sleepstage classification using multivariate and multimodal time series”. In:arXivpreprint arXiv:1707.03321(2017).[35] Nima Hatami, Yann Gavet, and Johan Debayle. “Classification of Time-SeriesImages Using Deep Convolutional Neural Networks”. In:CoRRabs/1710.00886(2017). arXiv:1710.00886.url:http://arxiv.org/abs/1710.00886.57
[36] Wenlin Wang et al. “Earliness-Aware Deep Convolutional Networks for EarlyTime Series Classification”. In: (Nov. 2016).[37] A. Elliot and C. H. Hsu. “Time Series Prediction : Predicting Stock Price”.In:ArXiv e-prints(Oct. 2017). arXiv:1710.05751 [stat.ML].[38] Y. Lecun et al. “Gradient-based learning applied to document recognition”.In:Proceedings of the IEEE86.11 (1998), pp. 2278–2324.doi:10.1109/5.726791.url:https://doi.org/10.1109%2F5.726791.[39] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. “ImageNet classifi-cation with deep convolutional neural networks”. In:Communications of theACM60.6 (May 2017), pp. 84–90.doi:10.1145/3065386.url:https://doi.org/10.1145%2F3065386.[40] Ken Chatfield et al. “Return of the Devil in the Details: Delving Deep intoConvolutional Nets”. In:Proceedings of the British Machine Vision Conference2014. British Machine Vision Association, 2014.doi:10.5244/c.28.6.url:https://doi.org/10.5244%2Fc.28.6.[41] Yann LeCun et al. “Efficient BackProp”. In:Lecture Notes in Computer Sci-ence. Springer Berlin Heidelberg, 1998, pp. 9–50.doi:10.1007/3-540-49430-8_2.url:https://doi.org/10.1007%2F3-540-49430-8_2.[42] Diederik Kingma and Jimmy Ba. “Adam: A method for stochastic optimiza-tion”. In:arXiv preprint arXiv:1412.6980(2014).[43] Sebastian Ruder. “An overview of gradient descent optimization algorithms”.In:arXiv preprint arXiv:1609.04747(2016).[44]Multivariate Statistics. Springer New York, 2007.doi:10.1007/978-0-387-73508-5.url:https://doi.org/10.1007%2F978-0-387-73508-5.58
[45] I.A Basheer and M Hajmeer. “Artificial neural networks: fundamentals com-puting, design, and application”. In:Journal of Microbiological Methods43.1(Dec. 2000), pp. 3–31.doi:10.1016/s0167-7012(00)00201-3.url:https://doi.org/10.1016%2Fs0167-7012%2800%2900201-3