| ACKNOWLEDGEMENT | 第5-12页 |
| ABSTRACT | 第12-13页 |
| 论文文摘/Translated Abstract | 第14-16页 |
| CHAPTER 1 INTRODUCTION | 第16-22页 |
| 1.1 Background | 第16页 |
| 1.2 Motivation | 第16-17页 |
| 1.3 Objectives of the study | 第17页 |
| 1.4 Problem description | 第17-19页 |
| 1.4.1 Behavior exhibiting features on data | 第18-19页 |
| 1.5 Implementation and deployment of deep learning | 第19-20页 |
| 1.6 Study contributions | 第20页 |
| 1.7 Organization of the Thesis | 第20-22页 |
| CHAPTER 2 BACKGROUND LITERATURE REVIEW | 第22-26页 |
| 2.1 Related works | 第22-23页 |
| 2.2 Behavior Modeling and Anomaly Detection Literature | 第23-24页 |
| 2.3 Deep Data Analytics and Behavior Modeling | 第24-26页 |
| CHAPTER 3 METHODOLOGY OF DEEP LEARNING USING RNNs | 第26-38页 |
| 3.1 RNNs and Deep Learning | 第26-27页 |
| 3.1.1 Characteristics of Deep RNNs | 第26-27页 |
| 3.2 Methodology and mathematical concepts in RNN and deep learning | 第27-34页 |
| 3.2.1 Perceptron as a basic neuron | 第27-29页 |
| 3.2.2 How learning in neural net algorithms happen | 第29-30页 |
| 3.2.3 Learning using Deep RNNs | 第30-31页 |
| 3.2.4 Gating mechanism in LSTM to create context awareness | 第31-34页 |
| 3.3 Features engineering using doc2vec modeling approach | 第34-38页 |
| 3.3.1 Doc2vec modeling | 第35-37页 |
| 3.3.2 Necessity of a doc2vec model in language modeling | 第37-38页 |
| CHAPTER 4 IMPLEMENTATION OVERVIEW | 第38-51页 |
| 4.1 Introduction to Behavior modeling (BM) | 第38页 |
| 4.2 Functional requirements for the study experiments | 第38页 |
| 4.3 Experiment setup analysis | 第38-51页 |
| 4.3.1 How to prevent the Vanishing Gradient Problem | 第38页 |
| 4.3.2 Theano Library | 第38-40页 |
| 4.3.3 Data collection, Features Analysis &Preparation | 第40-42页 |
| 4.3.3.2 Features engineering for inputs preparation | 第41-42页 |
| 4.3.3.3 The Importance of word vectors in a vector space model | 第42页 |
| 4.3.4 Keras- deep learning library and RNN Model | 第42-43页 |
| 4.3.4.1 Distributed representation Words in Doc2vec Matrix | 第43页 |
| 4.3.4.2 Extracting inputs from the doc2vec model | 第43页 |
| 4.3.5 The Algorithm Design and Architecture | 第43-46页 |
| 4.3.5.1 Training | 第44-45页 |
| 4.3.5.2 Input layer | 第45页 |
| 4.3.5.3 Output layer | 第45-46页 |
| 4.3.5.4 Objective function activation using Softmax | 第46页 |
| 4.3.5.5 Compiling and fitting the model | 第46页 |
| 4.3.6 Choice of Efficient parameters | 第46-50页 |
| 4.3.6.1 Training modes | 第47页 |
| 4.3.6.2 Initialization in layers | 第47页 |
| 4.3.6.3 Activation function | 第47-48页 |
| 4.3.6.3 Adding Dropout between layers | 第48页 |
| 4.3.6.4 Batch Normalization | 第48页 |
| 4.3.6.5 Choice for model optimization | 第48-49页 |
| 4.3.6.6 Choice for the Batch size | 第49页 |
| 4.3.6.7 Using Callbacks | 第49-50页 |
| 4.3.7 Prediction inference | 第50-51页 |
| CHAPTER 5 S A CASE STUDIES, BENCHMARK AND RESULTS | 第51-68页 |
| 5.1 Introduction | 第51-53页 |
| 5.1.1 People's Data Acquisition with a web application | 第51-52页 |
| 5.1.2 Steps followed to realize a case study experiment | 第52-53页 |
| 5.2 Movie Reviews Case study 1-The IMDbModel | 第53页 |
| 5.3 Dataset Source | 第53-58页 |
| 5.3.2 Pre-processing and Data cleaning | 第53-54页 |
| 5.3.3 Feature vectors engineering through Doc2vec Modeling | 第54页 |
| 5.3.4 Learning features similarity via the distributed memory(DM) words modeling | 第54-58页 |
| 5.3.5 Inputs vectors and formatting | 第58页 |
| 5.4 Classification with Keras | 第58-63页 |
| 5.4.1 Extract the Training and testing vectors from the doc2vec vector space model | 第59-60页 |
| 5.4.2 Initialize a new model | 第60页 |
| 5.4.3 Adding hidden layers to the model | 第60页 |
| 5.4.4 Model compiling and parameter tuning | 第60页 |
| 5.4.5 Performing Training with Keras classifier algorithm | 第60-61页 |
| 5.4.6 Performance results and model evaluation | 第61页 |
| 5.4.7 Generating visualizations | 第61-62页 |
| 5.4.8 Perform output predictions | 第62-63页 |
| 5.4.9 Using the Model for deployment | 第63页 |
| 5.5 USA Travelers'Airlines sentiments Case Study 2-The TwitterDataModel | 第63-65页 |
| 5.6 Benchmarks setup and Baselines results | 第65-68页 |
| 5.6.1 kNN classifier | 第65页 |
| 5.6.2 Random forest classifier | 第65-66页 |
| 5.6.3 Passive-Aggressive Classifier | 第66页 |
| 5.6.4 Benchmarks Setup and results | 第66-68页 |
| CHAPTER 6 EVALUATION, ANALYSIS AND BASELINES COMPARISON | 第68-80页 |
| 6.1 Metrics presentation and Evaluation | 第68-69页 |
| 6.2 Algorithms evaluation | 第69-76页 |
| 6.2.1 Training time | 第69页 |
| 6.2.2 Accuracy and Loss Measures | 第69-70页 |
| 6.2.3 Effect of increasing the number of epochs | 第70页 |
| 6.2.4 The Existence of unbalanced training sets | 第70页 |
| 6.2.5 Confusion Matrix (CM) | 第70-71页 |
| 6.2.6 ROC Curve and Graph | 第71-73页 |
| 6.2.7 Accuracy and Loss response curves | 第73-75页 |
| 6.2.8 Visualizing predictions and class probabilities | 第75-76页 |
| 6.3 Summary discussion from visual graphs | 第76-77页 |
| 6.4 Evaluation of Baselines＆Performance Comparisons | 第77-78页 |
| 6.5 Challenges/Limitations encountered | 第78-79页 |
| 6.6 Deployment | 第79-80页 |
| CONCLUSION AND FURTHER SUGGESTIONS | 第80-82页 |
| Conclusion | 第80-81页 |
| Suggestions on future works improvements | 第81-82页 |
| REFERENCES | 第82-84页 |
| Appendix A During the period of Study for a Master's Degree | 第84页 |
