學術產出-Theses
Article View/Open
Publication Export
-
題名 隱藏狀態模型、貝氏估計與企業營運
Hidden States Models, Bayesian Estimation and Business Operations作者 黃紀維
Huang, Chi-Wei貢獻者 周彥君<br>莊皓鈞
Chou, Yen-Chun<br>Chuang, Hao-Chun
黃紀維
Huang, Chi-Wei關鍵詞 離散時間序列
分布預測
隱藏馬可夫鏈模型
循環神經網路
貝氏估計
馬可夫鏈蒙地卡羅
Discrete Time Series
Distributional Forecast
Hidden Markov Model
Recurrent Neural Network
Bayesian Estimation
Markov chain Monte Carlo日期 2023 上傳時間 2-Aug-2023 14:06:09 (UTC+8) 摘要 企業營運的情境中經常儲存大量的離散時間序列，利用用戶過去需求資料預測未來可能的情況對於企業營運中的規劃與決策舉足輕重。隱藏馬可夫鏈模型 (Hidden Markov Models, HMM) 是機器學習模型中具有能探討隱藏特徵與挖掘隱藏狀態與觀察值之間關係的模型，捕捉不可觀察的內在特徵預測未來長期與短期的模式。有別於HMM簡潔明晰的參數模型，新興的機器學習模型循環神經網路 (Recurrent Neural Networks, RNN) 較為複雜且更具彈性，DeepAR 同樣善於預測序列資料。本研究與先前研究不同，我們運用零售M5競賽資料集對HMM與RNN進行較完備的比較，M5資料集中存在過度分散與0的出現次數過高的問題，我們提出Poisson HMM、Generalized Poisson HMM和NB DeepAR進行需求分布預測。由於HMM模型參數估計不易，因此我們使用馬可夫鏈蒙地卡羅 (Markov Chain Monte Carlo, MCMC) 估計HMM參數。我們發現HMM預測M5資料集的分位數損失勝過DeepAR，Generalized Poisson HMM在預測較高的分位數損失有更好的表現，HMM相較DeepAR能更精準的預測前一期自我回歸係數較高的單品，DeepAR則是對於0的比例較高的單品序列中表現出色，至於GP1 HMM與Poisson HMM僅在Q值較大的分位數損失有明顯差異，這對於如罕見疾病藥物的需求預測具有重要意義。我們所設計的HMM面對數十萬筆的資料能兼顧運算成本與相當的精準度，同時解釋力佳的HMM能夠有效支援管理者於商業營運上的決策。
In the context of enterprise operations, storing large volumes of discrete time series and utilizing past customer demand to predict future scenarios is crucial for planning and decision-making. Hidden Markov Models (HMMs) are machine learning models that explore hidden features and uncover relationships between hidden states and observed values, capturing unobservable underlying patterns for long-term and short-term predictions. Unlike the concise and straightforward parameter model of HMMs, the emerging machine learning model Recurrent Neural Networks (RNNs) is more complex and flexible. DeepAR, in particular, excels in predicting sequential data. In this study, different from previous research, we conducted a comprehensive comparison between HMMs and RNNs using the retail M5 competition dataset. The M5 dataset presents challenges such as overdispersion and zero-inflation. To address these issues, we proposed Poisson HMM, Generalized Poisson HMM, and NB DeepAR for demand distributional forecast. Due to the challenging parameter estimation of HMMs, we employed Markov Chain Monte Carlo (MCMC) for HMM parameter estimation. We found that HMMs outperformed DeepAR in predicting quantile losses for the M5 dataset. The Generalized Poisson HMM demonstrated better performance in predicting higher quantile losses. DeepAR excels in handling single-item sequences with a high proportion of zeros. On the other hand, GP1 HMM and Poisson HMM exhibit significant differences only in terms of quantile losses at larger Q values. This finding holds particular significance for demand prediction in scenarios such as medications for rare diseases. The HMM we designed strikes a balance between computational costs and accuracy when handling hundreds of thousands of data. Additionally, the interpretability of HMMs effectively supports managers in making decisions for business operations.參考文獻 Alwan, L. C., & Weiß, C. H. (2017). INAR implementation of newsvendor model for serially dependent demand counts. International Journal of Production Research, 55(4), 1085-1099.Ascarza E, Netzer O, Hardie BGS (2018) Some customers would rather leave without saying goodbye. Marketing Science 37(1), 54-77.Chapados, N. (2014). Effective Bayesian modeling of groups of related count time series. International conference on machine learning (pp. 1395-1403).Chen, J., & Rao, V. R. (2022). Evaluating strategies for promoting retail mobile channel using a hidden Markov model. Journal of Retailing.Chib, S., & Greenberg, E. (1995). Understanding the metropolis-hastings algorithm. The american statistician, 49(4), 327-335.Consul, P. C., & Jain, G. C. (1973). A generalization of the Poisson distribution. Technometrics, 15(4), 791-799.Goldberg, Y., Adler, M., & Elhadad, M. (2008, June). EM can find pretty good HMM POS-taggers (when given a good start). In Proceedings of ACL-08: HLT (pp. 746-754).Graves, A. (2013). Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850.Hoff, P. D. (2009). A first course in Bayesian statistical methods (Vol. 580). New York: Springer.Kolassa, S. (2016). Evaluating predictive count data distributions in retail sales forecasting. International Journal of Forecasting, 32(3), 788-803.Montoya, R., & Gonzalez, C. (2019). A hidden Markov model to detect on-shelf out-of-stocks using point-of-sale data. Manufacturing & Service Operations Management, 21(4), 932-948.Montoya, R., Netzer, O., & Jedidi, K. (2010). Dynamic allocation of pharmaceutical detailing and sampling for long-term profitability. Marketing Science, 29(5), 909-924.Naumzik, C., Feuerriegel, S., & Weinmann, M. (2022). I will survive: Predicting business failures from customer ratings. Marketing Science, 41(1), 188-207.Netzer O, Lattin JM, Srinivasan V (2008) A hidden Markov model of customer relationship dynamics. Marketing Science 27(2), 185-204.Sahoo, N., Singh, P. V., & Mukhopadhyay, T. (2012). A hidden Markov model for collaborative filtering. MIS quarterly, 1329-1356.Salinas, D., Flunkert, V., Gasthaus, J., & Januschowski, T. (2020). DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3), 1181-1191.Schulte, B., & Sachs, A. L. (2020). The price-setting newsvendor with Poisson demand. European Journal of Operational Research, 283(1), 125-137.Singh, P. V., Tan, Y., & Youn, N. (2011). A hidden Markov model of developer learning dynamics in open source software projects. Information Systems Research, 22(4), 790-807.Valendin, J., Reutterer, T., Platzer, M., & Kalcher, K. (2022). Customer base analysis with recurrent neural networks. International Journal of Research in Marketing, 39(4), 988-1018.Ziel, F. (2022). M5 competition uncertainty: Overdispersion, distributional forecasting, GAMLSS, and beyond. International Journal of Forecasting, 38(4), 1546-1554.Zucchini, W., MacDonald, I.L., & Langrock, R. (2016). Hidden Markov Models for Time Series: An Introduction Using R, Second Edition (2nd ed.). Chapman and Hall/CRC. https://doi.org/10.1201/b20790 描述 碩士
國立政治大學
資訊管理學系
110356039資料來源 http://thesis.lib.nccu.edu.tw/record/#G0110356039 資料類型 thesis dc.contributor.advisor 周彥君<br>莊皓鈞 zh_TW dc.contributor.advisor Chou, Yen-Chun<br>Chuang, Hao-Chun en_US dc.contributor.author (Authors) 黃紀維 zh_TW dc.contributor.author (Authors) Huang, Chi-Wei en_US dc.creator (作者) 黃紀維 zh_TW dc.creator (作者) Huang, Chi-Wei en_US dc.date (日期) 2023 en_US dc.date.accessioned 2-Aug-2023 14:06:09 (UTC+8) - dc.date.available 2-Aug-2023 14:06:09 (UTC+8) - dc.date.issued (上傳時間) 2-Aug-2023 14:06:09 (UTC+8) - dc.identifier (Other Identifiers) G0110356039 en_US dc.identifier.uri (URI) http://nccur.lib.nccu.edu.tw/handle/140.119/146578 - dc.description (描述) 碩士 zh_TW dc.description (描述) 國立政治大學 zh_TW dc.description (描述) 資訊管理學系 zh_TW dc.description (描述) 110356039 zh_TW dc.description.abstract (摘要) 企業營運的情境中經常儲存大量的離散時間序列，利用用戶過去需求資料預測未來可能的情況對於企業營運中的規劃與決策舉足輕重。隱藏馬可夫鏈模型 (Hidden Markov Models, HMM) 是機器學習模型中具有能探討隱藏特徵與挖掘隱藏狀態與觀察值之間關係的模型，捕捉不可觀察的內在特徵預測未來長期與短期的模式。有別於HMM簡潔明晰的參數模型，新興的機器學習模型循環神經網路 (Recurrent Neural Networks, RNN) 較為複雜且更具彈性，DeepAR 同樣善於預測序列資料。本研究與先前研究不同，我們運用零售M5競賽資料集對HMM與RNN進行較完備的比較，M5資料集中存在過度分散與0的出現次數過高的問題，我們提出Poisson HMM、Generalized Poisson HMM和NB DeepAR進行需求分布預測。由於HMM模型參數估計不易，因此我們使用馬可夫鏈蒙地卡羅 (Markov Chain Monte Carlo, MCMC) 估計HMM參數。我們發現HMM預測M5資料集的分位數損失勝過DeepAR，Generalized Poisson HMM在預測較高的分位數損失有更好的表現，HMM相較DeepAR能更精準的預測前一期自我回歸係數較高的單品，DeepAR則是對於0的比例較高的單品序列中表現出色，至於GP1 HMM與Poisson HMM僅在Q值較大的分位數損失有明顯差異，這對於如罕見疾病藥物的需求預測具有重要意義。我們所設計的HMM面對數十萬筆的資料能兼顧運算成本與相當的精準度，同時解釋力佳的HMM能夠有效支援管理者於商業營運上的決策。 zh_TW dc.description.abstract (摘要) In the context of enterprise operations, storing large volumes of discrete time series and utilizing past customer demand to predict future scenarios is crucial for planning and decision-making. Hidden Markov Models (HMMs) are machine learning models that explore hidden features and uncover relationships between hidden states and observed values, capturing unobservable underlying patterns for long-term and short-term predictions. Unlike the concise and straightforward parameter model of HMMs, the emerging machine learning model Recurrent Neural Networks (RNNs) is more complex and flexible. DeepAR, in particular, excels in predicting sequential data. In this study, different from previous research, we conducted a comprehensive comparison between HMMs and RNNs using the retail M5 competition dataset. The M5 dataset presents challenges such as overdispersion and zero-inflation. To address these issues, we proposed Poisson HMM, Generalized Poisson HMM, and NB DeepAR for demand distributional forecast. Due to the challenging parameter estimation of HMMs, we employed Markov Chain Monte Carlo (MCMC) for HMM parameter estimation. We found that HMMs outperformed DeepAR in predicting quantile losses for the M5 dataset. The Generalized Poisson HMM demonstrated better performance in predicting higher quantile losses. DeepAR excels in handling single-item sequences with a high proportion of zeros. On the other hand, GP1 HMM and Poisson HMM exhibit significant differences only in terms of quantile losses at larger Q values. This finding holds particular significance for demand prediction in scenarios such as medications for rare diseases. The HMM we designed strikes a balance between computational costs and accuracy when handling hundreds of thousands of data. Additionally, the interpretability of HMMs effectively supports managers in making decisions for business operations. en_US dc.description.tableofcontents 壹、緒論 1貳、文獻探討 3參、隱藏馬可夫鏈與貝氏估計 5一、隱藏馬可夫鏈模型 5二、貝氏估計MCMC 7肆、自我回歸循環網路 12伍、隱藏狀態模型與實證資料 14一、M5競賽 14二、隱藏狀態模型 16三、實驗結果 18陸、結論 29參考文獻 30 zh_TW dc.format.extent 3153066 bytes - dc.format.mimetype application/pdf - dc.source.uri (資料來源) http://thesis.lib.nccu.edu.tw/record/#G0110356039 en_US dc.subject (關鍵詞) 離散時間序列 zh_TW dc.subject (關鍵詞) 分布預測 zh_TW dc.subject (關鍵詞) 隱藏馬可夫鏈模型 zh_TW dc.subject (關鍵詞) 循環神經網路 zh_TW dc.subject (關鍵詞) 貝氏估計 zh_TW dc.subject (關鍵詞) 馬可夫鏈蒙地卡羅 zh_TW dc.subject (關鍵詞) Discrete Time Series en_US dc.subject (關鍵詞) Distributional Forecast en_US dc.subject (關鍵詞) Hidden Markov Model en_US dc.subject (關鍵詞) Recurrent Neural Network en_US dc.subject (關鍵詞) Bayesian Estimation en_US dc.subject (關鍵詞) Markov chain Monte Carlo en_US dc.title (題名) 隱藏狀態模型、貝氏估計與企業營運 zh_TW dc.title (題名) Hidden States Models, Bayesian Estimation and Business Operations en_US dc.type (資料類型) thesis en_US dc.relation.reference (參考文獻) Alwan, L. C., & Weiß, C. H. (2017). INAR implementation of newsvendor model for serially dependent demand counts. International Journal of Production Research, 55(4), 1085-1099.Ascarza E, Netzer O, Hardie BGS (2018) Some customers would rather leave without saying goodbye. Marketing Science 37(1), 54-77.Chapados, N. (2014). Effective Bayesian modeling of groups of related count time series. International conference on machine learning (pp. 1395-1403).Chen, J., & Rao, V. R. (2022). Evaluating strategies for promoting retail mobile channel using a hidden Markov model. Journal of Retailing.Chib, S., & Greenberg, E. (1995). Understanding the metropolis-hastings algorithm. The american statistician, 49(4), 327-335.Consul, P. C., & Jain, G. C. (1973). A generalization of the Poisson distribution. Technometrics, 15(4), 791-799.Goldberg, Y., Adler, M., & Elhadad, M. (2008, June). EM can find pretty good HMM POS-taggers (when given a good start). In Proceedings of ACL-08: HLT (pp. 746-754).Graves, A. (2013). Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850.Hoff, P. D. (2009). A first course in Bayesian statistical methods (Vol. 580). New York: Springer.Kolassa, S. (2016). Evaluating predictive count data distributions in retail sales forecasting. International Journal of Forecasting, 32(3), 788-803.Montoya, R., & Gonzalez, C. (2019). A hidden Markov model to detect on-shelf out-of-stocks using point-of-sale data. Manufacturing & Service Operations Management, 21(4), 932-948.Montoya, R., Netzer, O., & Jedidi, K. (2010). Dynamic allocation of pharmaceutical detailing and sampling for long-term profitability. Marketing Science, 29(5), 909-924.Naumzik, C., Feuerriegel, S., & Weinmann, M. (2022). I will survive: Predicting business failures from customer ratings. Marketing Science, 41(1), 188-207.Netzer O, Lattin JM, Srinivasan V (2008) A hidden Markov model of customer relationship dynamics. Marketing Science 27(2), 185-204.Sahoo, N., Singh, P. V., & Mukhopadhyay, T. (2012). A hidden Markov model for collaborative filtering. MIS quarterly, 1329-1356.Salinas, D., Flunkert, V., Gasthaus, J., & Januschowski, T. (2020). DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3), 1181-1191.Schulte, B., & Sachs, A. L. (2020). The price-setting newsvendor with Poisson demand. European Journal of Operational Research, 283(1), 125-137.Singh, P. V., Tan, Y., & Youn, N. (2011). A hidden Markov model of developer learning dynamics in open source software projects. Information Systems Research, 22(4), 790-807.Valendin, J., Reutterer, T., Platzer, M., & Kalcher, K. (2022). Customer base analysis with recurrent neural networks. International Journal of Research in Marketing, 39(4), 988-1018.Ziel, F. (2022). M5 competition uncertainty: Overdispersion, distributional forecasting, GAMLSS, and beyond. International Journal of Forecasting, 38(4), 1546-1554.Zucchini, W., MacDonald, I.L., & Langrock, R. (2016). Hidden Markov Models for Time Series: An Introduction Using R, Second Edition (2nd ed.). Chapman and Hall/CRC. https://doi.org/10.1201/b20790 zh_TW
