我們知道公司不應只專注於為客戶提供最便宜的產品。給客戶提供舒適的消費體驗是更為重要的。舉個例子: 如果一個客戶想買一些冷凍產品，但網站推薦的東西，客戶不感興趣，這是一個奇怪的情況。一個好的電子商務服務應該瞭解客戶的需求，並滿足您的客戶。

為了給客戶帶來更好的體驗與節約人力資源。在我們的實驗中，我們首先想知道的是有多少類型的客戶可以分類。當我們知道客戶的類型，我們就可以使用人力資源去服務客戶將預訂包裝行程或想要購買一些便宜產品。為了實現這一目標，我們考慮了兩種方法。一個是監督學習。我們使用 Frame 提供的標註作為特徵，將顧客提出的第一個要求轉成基於特稱所構成基底的向量再去做聚類，然後使用聚類結果作為標籤以及顧客提出的第一個要作為輸入來分類客戶。對於監督學習（BERT 加上 邏輯斯迴歸），我們可以得到95%的準確性。對於無監督學習。我們僅使用顧客提出的第一個要求作為輸入和不同的嵌入來表示它們來聚類。除了不同的嵌入外，還有一種方法屬於無人監督的學習，稱為SimCLR。雖然條形圖上的最佳精度是 LDA 加 tf-idf，但散射圖上只有 SimCLR 可以將相同類型的客戶聚集在一起。第二，我們想知道的是每句話是什麼樣的主題，那麼我們可以在正確的時間使用正確的人力資源，以滿足客戶的需求。在我們的實驗中，我們的對話主題是由人類決定的。我們使用 BERT 和邏輯斯迴歸來對顧客所提出的要求進行分類。在瞭解客戶類型和每一個句子的主題後，我們想知道的最後一個資訊是每一個對話中的客戶行為。為了觀察客戶行為，我們為每種類型的客戶創建資料隱馬爾可夫模型。當我們了解客戶行為時，我們可以根據對話的一部分預測一些客戶行為。我們可以推薦一些客戶想要購買的產品，也可以區分客戶是否有比較價格的行為，也可以區分客戶是否只想知道價格。

We know that a company should not only focus on offering customers the cheapest product. It is important to give customers a comfortable experience. It’s a weird sit-uation that a customer wants to buy some frozen product but a website recommends something that a customer is not interested in. A good e-commerce service should understand what your customer wants and satisfy your customer.

In order to give customers a better experience and save human resources. In our experiment first thing, we want to know is how many types of customers can be classified. If we know the customer’s type, we can use the human resource on the customers who will book a package or want to buy some bargain products. To achieve this goal, we consider two situations. One is supervised learning. We use Frames annotation as feature to cluster first turn then using clustering result as label and first turn as input to classify a customer. For supervised learning, we can get 95% accuracy. Second is unsupervised learning. We only use first turn as input and different embeddings to represent them to cluster turns. Beside different embeddings, there is a method belongs to unsupervised learning called A Simple Framework for Contrastive Learning of Visual Representations (SimCLR). Although the best accu-racy on bar graph is LDA plus tf-idf, on scatter graph only SimCLR can gather same type of customers together. The second thing we want to know is what kind of topics in each turn. Since we know the topic then we can put the right human resource at the right time in order to fulfill customer demand. In our experiment, our dialogue topic is determined by human decision. We use BERT and logistic regression to classify a turn. After knowing the types of customers and the topic of each turn, the last thing we want to know is customer behavior in each dialogue. To observe cus-tomer behavior, we create profile hidden Markov models for each type of customers. When we know customer behavior, we can predict some customer behavior based on part of a dialogue. We can recommend some products that a customer wants to buy or we can distinguish whether a customer has the behavior of comparing price or not also we can distinguish whether a customer only wants to know the price.

[1] Ali H H, Kadhum L E. K-Means Clustering Algorithm Applications in Data Mining and Pattern Recognition[J]. International Journal of Science and Research, 2017, 6(8): 1577-1584.
[2] Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." arXiv preprint arXiv:1810.04805 (2018).
[3] A.K. Jain, M.N. Murty and P.J. Flynn.‖Data Clustering: A Review‖. ACM Com-puting Surveys. 1999. Vol. 31, No. 3.
[4] Patel V R, Mehta R G. Impact of outlier removal and normalization approach in modified k-means clustering algorithm[J]. International Journal of Computer Sci-ence Issues (IJCSI), 2011, 8(5): 331.
[5] Jean Yan. ―Big Data, Bigger Opportunities- Data.gov’s roles: Promote, lead, contribute, and collaborate in the era of big data‖. 2013. Retrieved from http://www.meritalk.com/pdfs/bdx/bdx-whitepaper-090413.pdf on 14 July 2015.
[6] Gnanaraj T N, Kumar K R, Monica N. Survey on mining clusters using new k-mean algorithm from structured and unstructured data[J]. IJACST, 2014.
[7] Hughes AM (1996). Boosting reponse with RFM. Mark. Tools, 5: 4-10.
[8] Kahan R (1998). Using database marketing techniques to enhance your one-to-one marketing initiatives. J. Consum. Mark., 15(5): 491-493.
[9] Tsai CY, Chiu CC (2004). A purchase-based market segmentation methodology. Expert Syst. Appl., 27: 265-276.
[10] Bult JR, Wansbeek T (1995). Optimal seleciton for direct mail. Market Sci.,
14(4): 378-394.
[11] Bitran GR, Mondschein SV (1996). Mailing decisions in the catalog sales
industry. Manage Sci., 42(9): 1364-1381.
[12] Miglautsch JR (2000). Thoughts on RFM scoring. J. Database Mark., 8(1):
67-72.
[13] Chang EC, Huang SC, Wu HH (2010). Using K-means method and
spectral clustering technique in an outfitter’s value analysis. Qual
Quant., 44(4): 807-815.
[14] Alam GM, Khalifa TB (2009). The impact of introducing a business
marketing approach to education: A study on private HE in Bangladesh.
Afr. J. Bus. Manage. 3 (9): 463-474.
[15] Rabiner L R. A tutorial on hidden Markov models and selected applications in speech recognition. Proc. IEEE. 1989;77:257–286.

[16] Kristoffer Forslund, Erik L. L. Sonnhammer, Predicting protein function from domain content, Bioinformatics, Volume 24, Issue 15, 1 August 2008, Pages 1681–1687,
[17] Di Francesco V, Garnier J, Munson P J. Protein topology recognition from sec-ondary structure sequences: Application of the hidden Markov models to the alpha class proteins. J. Mol. Biol. 1997;267:446–463.
[18] Di Francesco V, Munson P J, Garnier J. FORESST: fold recognition from sec-ondary structure predictions of proteins. Bioinformatics. 1999;15:131–140.
[19] David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent dirichlet al-location. J. Mach. Learn. Res. 3, null (3/1/2003), 993–1022.
[20] Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space." arXiv preprint arXiv:1301.3781 (2013).
[21] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Comput. 9, 8 (November 15, 1997), 1735–1780
[22] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//Advances in neural information processing systems. 2017: 5998-6008.
[23] Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020). A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709.
[24] Schulz H, Zumer J, Asri L E, et al. A frame tracking model for memory-enhanced dialogue systems[J]. arXiv preprint arXiv:1706.01690, 2017.
[25] Radford A, Wu J, Child R, et al. Language models are unsupervised multitask learners[J]. OpenAI blog, 2019, 1(8): 9.
[26] You Y, Gitman I, Ginsburg B. Large batch training of convolutional networks[J]. arXiv preprint arXiv:1708.03888, 2017.
[27] You Y, Li J, Reddi S, et al. Large batch optimization for deep learning: Training bert in 76 minutes[J]. arXiv preprint arXiv:1904.00962, 2019.