Big Data Risk Analytics
Abstract
In this paper, I am empirically validating the importance of analytics from different Machine Learning, Analytic tools and Deep Learning. These modern predicting technologies are now playing a vital role in both technology and business due to the rapid increase in data access and computing power. Advance analytic incorporated with Big data, data availability and processing capabilities are helping to many financial institutions, medical research and manufacturing industries are taking business guidance and redefining their project models. Mostly of financial organizations are using analytics in Credit, Business decisions, Network security and successful loan granting are key for taking the decision with confidence. Many serious diseases are diagnosis early via Machine Learning and deep learning models which save many patient’s life and also saved their hard earn money. Machine learning, Deep Learning are two main pillars of predictive technologies. Data has become Significant for analysis as many organizations start collecting huge amounts of information related to business-oriented, health care and other which contain more useful information about problems and customer’s activity and sentiments such as national security, cyber-crime, online marketing, financial scams and medical data analysis. Top MNC such as IBM, Google, Amazon and Microsoft are analyzing large volumes of data for business analysis and decisions, impacting existing Analyzing Risk and future technology, everything came true because of Big Data and it’s advance analytics. [2]
Index Terms—the Histogram of Oriented Gradients (HOG), Scale Invariant Feature Transform(SIFT), Computer Vision (CV), artificial intelligence (AI), Deep Neural Networking (DNN), Convolution Neural Network (CNN), Recurrent Neural Network (RNN), Generative Adversarial Networks (GANs), Machine Learning (ML).
I. INTRODUCTION
The MAIN vision of advance analytic is to generalization the patterns inside the data and helping to predict for future decisions. More usage of algorithms and data analysis raises many moral questions, many discussions are happening on the growing usage of non-public data statistics which all issues already addressed on Global popular records protection law (GDPR) published on May 2018. The public have fear of the use of personal and unknown algorithms are working on it and it would impact on their rights. Data analysis, visualizations and predictions are core pillars of advance analytic. Machine Learning algorithm’s performance depends on the quality of data if the quality of data is poor means it has missing and redundant information which will degrade the performance of the model. ML model’s performance and prediction depend on the uniqueness and authenticities of data, If Data has more variations and biases, the model will more complex and prediction will more uncertain. Feature engineering and extraction play a big role to develop a better model but for completing this tedious task user have needed domain expert. Nowadays Feature Engineering is slowly moving towards automated work and which require less human interaction requires to extract features from the Dataset and less financial expenses required for this. Most of the algorithms have the common characteristic are to understand the pattern and generalization it like similar object tracking, semantic segmentation, speech identifications, chatbot, defence industries etc. In Big Data Analytics are remain largely untapped power of most of the industries because now industries have already collected more data, the massive amount of data means has more information inside which help on various domains take a decision in business and more on research. Some BigData areas, such as computer vision [17] and speech identification [13], have seen DL applied mainly to enhance the performance of classification modelling. DL’s ability to access complexity high-end abstractions and data presentation from large volumes of data, especially unsupervised data, makes it attractive as a valuable tool for BigData Analytics. Deep Learning is addressing BigData issues with more accurate ways and it is solving business. Deep learning has more analytic options to address a variety of challenges. The conventional ML have many cumbersome tasks like feature engineering, extraction dummy variable conversion, feature scaling etc because it is a very complex process and need domain expert. So Deep Learning is quite easy to handle different type of data without domain expert and no data processing task and its accuracy were quite impressive then ML.
II. RISK ANALYTIC CHARACTERISTIC AND REQUIREMENTS
A. Data Requirement
Secure and reliable data collection always help for better data analysis, second preserving data should have enough features which can help for data insight, what we are preserving in the system we need domain expert who will guide about business understanding. Usually, data collection is considered a fundamental operation for any application. Generally, it is infrastructure limited, and how to conquer these constraints. Presently, many schemes have been proposed for eco-friendly records collection in actual time, which may be more or less divided into 3 categories,i.e., "mathematical model-based", "compressed sensing (CS)-based" and "query-driven based approximate data collection schemes". Fog computing on BigData reduces data processing delay, reduces bandwidth and it is eco-friendly.[15]
B. Analytical Tools
The traditional data processing tools have become incompetent to process the huge magnitude of data being produced over the industry. The conventional information processing gear has grown to be incompetent to process the massive significance of records being produced over the enterprise. Massive statistics analytics presents possibilities to leverage the big data to be had on the disposal of the business enterprise to the advantage in their commercial, ensuing in heightened commercial enterprise technology and better service delivery. Massive data analytics refers to the procedure of analyzing massive quantities of various styles of structured/unstructured data, in an effort to discover hidden styles, unknown correlations and different crucial statistics. Such facts can provide aggressive area over enterprise opponents and also help to analyze risk.[11]
C. Deep Learning
Deep Learning is part of supervised and unsupervised learning, its architecture is mimic of the human brain and it has a structure like neurons, it is internally structured as network, this network is called a neural network. It has multiple hidden layers basic structure of Deep learning have input layer and an output layer which is called Neural Network and in Deep neural network has at least on addition layer in between input and output layer which is called the hidden layer, it is also known as the perceptron. Now deep learning is using in various area of research and industries like image recognition, voice recognition, natural language process etc. [11]
D. Machine Learning
Machine learning is nowadays part of most the research it is adding many concepts of several branches of science and another area like mathematics, physics, biology, chemistry, statistics, aviation, financial etc. The main agenda of machine learning is to utilise historical data from the system and apply different machine learning algorithm and able to predict future data. Machine Learning main grouped into 3 categories. The first category is supervised learning, Second Category is unsupervised learning and the third category is reinforcement learning. Supervised Algorithm is applying on data which have labelled information means have a target class. Supervised Learning is two type Regression and classification. Second Unsupervised Learning is applying on data which have no label means no target class. The third is reinforcement learning which is collecting data from the environment. Generally, reinforcement learning is using for the decision-making problem statement. [16]
III. ANALYTICAL TOOLS
A. Google BigQuery
Google BigQuery is an Analytical Software that permits you to run square queries over your massive records. it is able to run query requests to heaps of servers, reads from tens or loads of thousands of disks immediately, and can go back solutions to complicated questions inside seconds. All parameters along with the computing environment, statistics quantities that can be processed, ease of use, selection making skills, time intake, and pricing norms. the writer concludes that the choice of the class of massive statistics analytical gear relies upon upon the utilization pattern and necessities of the agency.[11][12]
1) Analyze massive Datasets in few Seconds: BigQuery includes executing amazing-rapid, sq.-like queries towards massive datasets, strolling into terabytes of information in the time span of seconds via leveraging the processing strength lying on the disposal of Google.[11][12]
2) Various entry points: BigQuery may be accessed via BigQuery browser software, Google Apps Script, Google Spreadsheets, or via BigQuery rest API.[11][12]
3) Simplified records loading: Information needs to be loaded into BigQuery earlier than it is able to be used for analytics. customers might also easily load their information via streaming, direct add, or through Google Cloud storage.[11][12]
4) Easy Pricing: Google BigQuery strategies initial hundred GB of information totally free first month. They adhere to an easy and obvious pricing list, in which customers want to pay best for the garage they want and the queries they run. [11][12]
5) Easy to evaluate: customers have no cost to pay for small datasets with personal uses, companies, or maybe make it public. [11][12]
6) Analytic features: The strength of BigQuery lies in its potential to interactively execute combination queries on terabytes of information. however, from time to time counts and averages are not enough, hence BigQuery offers diverse superior features consisting of analytical capabilities for rating outcomes, exploring distributions and percentiles, and traversing outcomes without the want for a self be a part of. [11][12]
7) Tremendously Scalable: Excessive ingestion ability of massive quantities of information in mere mins makes Google’s large exceptionally scalable for any information volumes. [11][12] .
B. Datameer
Datameer is a strong massive information-analytical software that provides statistics integration, analytics and visualization capability, all added into one software platform. The distinguishing functions presented by using Datameer Analytics are : [11][13]
1) Records Integration: Datameer permits collecting and joining information from various sources, it is almost more than twenty connectors available for Hadoop. 2) REST API: Datameer allows easy REST API integration with external applications and ensures streamline business need. [11][13]
3) Tabular sheet Data Analysis: Datameer provides spreadsheet-like user-interface that facilitate the business users to analyze structured and unstructured information of any size by using analytic features and performs complex analytics without having to write a complex query. [11][13]
4) Exporting Analytics: This tool helps to export the results to the other sources like RDBMS and NoSql both.[11][13]
5) Quick Graphs: Datameer has "flip side view" feature which helps users to create quick visualization on a different context.[11][13]
6) Enhanced Visualization: Users can create easy and quick dashboards, time-series via Datameer’s available widgets. These visualizations are easily accessible on Mobiles, Desktops and online.[11][13]
C. Alteryx
Alteryx is one of extensively used information-analytical software throughout the enterprise. Alteryx is first-rate recognized for its functionality to hastily procedure all varieties of state-of-the-art data. Alteryx gives an easy-to-use analytics platform for organisations facilitating making essential choices that drives their enterprise approach and organizational business.[11]
1) information blending: The data mixing procedure includes integrating facts from various sources in one workflow that fuels specific enterprise procedure and guarantees progressed selection making. Alteryx seamlessly brings all distinct facts from heterogeneous sources, which includes client records from a conventional data-sources.[11]
2) R Language: facts Analytics in Alteryx is deeply included with statistical language R because of which even the maximum state-of-the-art and complex records units can be dealt with quite simply.[11]
3) Multi-threading: The multi-threading functionality of R integrated into Alteryx Analytics which extra capability to analysts and build quick advance predictive analytics.[11]
D. Pentaho
Pentaho Analytics software provides an entire solution to the mixing of data, analyze and graphical presentation of the large data into insights inside one platform.[11][14]
1) Mixed Analytics: This software provides all type of blending to users. It can be easily integrating with Hadoop NoSQL database, structural or non-structural database and easy to explore hidden secrets inside databases and easy to analyze.[11][14]
2) Interactive Graphical Presentation: It provides a rich source of the dashboard and easily updates as per requirement. Its visualization is web-based responsive. [11][14]
IV. DEEP LEARNING AND MACHINE LEARNING
A. Architectures
1) Recurrent Neural Network: RNNs have been very popular in areas where the sequence of presentation of the information is crucial. As a consequence, they find many applications in real-world domains such as the processing of human language, speech processing, and machine translation. RNNs provided solutions when information uses loop networks and allow information to persist. RNNs allow information in multiple copies of the same neural network, this chain-like shows that RNNs are closely linked to sequences and lists. Limiting
Vanilla Neural Networks: It accepts a specified-size vector as an input (e.g. an image) and generates a set-size vector as an output. Such neural network models use a fixed amount of computational steps to perform this mapping. The main reason recurrent networks are more fascinating is that it allows us to operate over vector sequences Two RNNs architectures are most famous: 1- Bidirectional RNN – In this RNN output depends on previous output as well as future output too. 2- Deep RNN – In RNN it allows multiple layers together in the current step and also generate and the greater rate of learning and accuracy. RNNs is using nowadays in building industry-oriented chatbots for customer interaction on websites. It also helps to understand the sequence of signals, audio wave. RNNs also help to understand audio waves and human language to predict and easily classify with the probability.
Advantages of RNN 1- In comparison to a conventional neural network, an RNNs uses the same parameters overall. It reduces the number of parameters considerably too. 2- RNNs and CNNs may be used to produce accurate descriptions for unlabelled images.
Disadvantages of RNN 1- RNNs find long term dependencies hard to monitor. This is particularly true for long sequences, sentences and paragraphs which have too many words which are available in between noun and verb. 2- It can’t be stack RNNs into very deep models. This is because of the activation function used in RNN models which causes the gradient to decay over layers by layers.
2) Autoencoders: In an unsupervised environment, autoencoders apply the backpropagation principle. Interestingly, auto-encoders are closely resembling PCA (Principal Component Analysis), except they are more flexible. Some of Autoencoders are commonly used in anomaly detection, for example, Fraud Detection in Bank, Discrepancy in financial transactions. The core task of autoencoders is essential to define and decide what constitutes of data, and then identify the outliers or anomalies. Normally, autoencoders depict data through multiple hidden layers so that the output value is as similar to the input value.[8] There are 4 types of encoders available:
Vanilla autoencoder – There is the simplest form of autoencoders, which is neural network with single hidden layer Multilayer autoencoder – Sometime autoencoder need more hidden layer so it can be extended to include more hidden layers.
Convolutional autoencoder – Convolutions are used instead of fully connected layers in the autoencoders.[8]
Regularized autoencoder – This form of autoencoder uses a special loss function that allows the model to have properties that go beyond the simple ability to copy a given input to the output.[8]
Advantages of Autoencoders 1- Autoencoders offer a resulting model-based specifically on the data and not predefined filters[8] 2- Less complex means it can be trained faster.[8]
Disadvantages of Autoencoders[8] 1- Sometimes, it can take more time to train.[8] 2- Sometimes training data is not indicative of the test data, then the details come from the models may be blurred and ambiguous.[8] 3- Some autoencoders, especially of the variational kind, trigger deterministic biases in the model.[8]
3) Generative Adversarial Networks: GANs are a versatile type of neural networks used to learn without supervision. It was developed by Ian J. Goodfellow and launched in 2014.
GANs consist essentially of a system of two competing neural network models that compete with each other and are capable of analyzing, capturing and copying the variations within a dataset. GANs have three constituents:
1- Generative: A generative model which describes how a probabilistic model generates data.[8]
2- Adversarial: To complete model’s training in against of environment. [8]
3- Networks: Using deep neural networks as the training model.[8]
The general architecture of GAN:
Advantages of GAN: 1- GANs allow semisupervised efficient training of the classifiers. 2- GAN have better precision, The data generated GAN is very unique than original data.[8] 3- Unlike variational autoencoders, GANs do not introduce any deterministic biases. Disadvantages of GAN: 1- Operating efficiently by generator and discriminator is critical to GAN’s performance. The entire system fails even if one of them get an error.[8] 2- The generator as well as the discriminator are separate systems and are trained with various loss functions. More time needed to train the whole process.[8]
4) Residual neural network: Most data scientists and AI researchers have widely adopted and used ResNets, or Deep Residual Networks. As we know, CNNs are extremely useful when it comes to solving problems with image detection and visual recognition. As these tasks become more complex, neural network training begins to get much more complicated, when additional deep layers are required to measure and improve model accuracy. Residual learning is designed to address the complex problem and the resulting architecture is popularly referred to as a ResNet. A ResNet is made up of a number of residual modules–where each module is a layer. [8] Every layer consists of a collection of functions that must be performed on the input. ResNet has more residual modules, each of which represents one layer. Each layer is comprised of a set of functions to execute on the input. The depth of a ResNet varies considerably in Microsoft researchers invented ResNet which had 152 layers![8]
Advantages of ResNets: 1- ResNets are accuracy is more and sometimes it is needed less weight than RNNs and LSTMs.[8] 2- It is a complex model. It is possible to add hundreds and thousands of residual layers to create a network, and then train.[8] 3- ResNets can be designed as complex as possible based on the requirement by adding layers. Disadvantages of ResNets: 1- Suppose the layers in a ResNet are very complex and errors can be difficult to detect, and they can not be quickly and correctly propagated back. At the same time, the learning may not be very effective if the layers are too small.[8]
5) Long short-term memory: LSTM is extended version kind of RNNs Networks that include a special memory cell capable of holding information for long periods. A collection of gates is used to determine when specific information enters the memory and when it is erased. LSTM is a better option when we have a sequence of data with long-tern term dependency like TimeSeries.[8]
6) SqueezeNet : It is a special type of DNN architecture which work for Computer Vision deep, it is published in 2016. SqueezeNet developed by researchers of DeepScale. The goal of the researchers in designing of SqueezeNet was to create a smaller neural network with fewer parameters that can easily fit into system memory and it can be easily transmitted over the network. The architecture of squeezeNet is one more powerful architecture which is extremely useful for low bandwidth scenarios such as telecom platforms. This architecture has only used only 4.9 MB of space, while inception uses more than 100 MB.[7][8]
7) cabinet: Capsule Networks, is advance research in DNN and NN modelling. it is mainly using to recognize image accurately and is an advanced extension of CNNs.[7][8]
8) SegNet: it is a special type of CNN architecture used for the wise labelling of semantic pixels. More generally, this problem is called semantic segmentation. It is majorly using to solve the image segmentation issues. SegNet is designed to solve the image segmentation problem. This consists of processing layers (encoders) sequences followed by a corresponding set of decoders for pixel-wise classification.[7][8]
9) Seq2Seq: It is originally implemented by Google for machine translation. The translation earlier worked very naively. Earlier, every word has been transformed to its output language by giving no consideration to its grammar and sentence structure. Seq2seq innovated the language understanding and translation by using deep learning. Nowadays seq2seq deep learning architecture is using for language translation and making chatbots.[7][8]
10) AlexNet : It is the first deep learning architecture to be implemented by the father of the deep learning Geoffrey Hinton and colleagues. It is a simple yet powerful network architecture that has helped pave the way for an important role in deep learning research. AlexNet participated in the Massive scale Image Recognition Challenge in ImageNet. The primary finding of the original research paper was that the model’s complexity was necessary for its high performance, which was computationally expensive but it made possible through by the use of GPUs during preparation.[7][8] In the 1980s, a neural network was developed using CPU. Whereas AlexNet only by using GPU accelerates the training by 10 times. AlexNet is still used as a starting point for applying deep neural networks to all activities, whether it’s computer vision or speech recognition, even though it’s a bit outdated currently.[7][8]
11) VGG Net : The researchers at Visual Graphics Group in Oxford developed the VGG Network. This network architecture is distinguished particularly by its pyramidal form, where the bottom layers closer to the picture are large whereas the upper layers are more in deep. VGG comprises subsequently CNN layers followed by pooling layers as demonstrated in the image. The pooling layers have the task of narrowing the layers. They suggested several such types of networks in their paper, with a shift in the profundity of the design.[7][8]
Advantages: Pre-trained VGG networks are freely available on the internet, so it’s generally used out of the box for different applications. Disadvantage: On the other hand, its principal disadvantage is that if we train model from scratch, it will take more time to train , more than week.[7][8]
12) GoogleNet or Inception Network: It is a deep learning architecture class developed by Google-researchers. GoogleNet was the ImageNet 2014 winner, where it has proven to be a successful platform. The researchers have made a novel approach in this architecture, along with going deeper (it includes 22 layers compared to VGG which had 19 layers), called the Inception module. As seen above, it is a drastic shift from the sequential architectures that we have seen before. Many forms of ”function extractors” occur in a single layer. It also makes the network perform better, as the network has many choices to choose from when solving the challenge at training itself. It can either choose to compress the data or pool it. it’s architecture have multiple of these stacked initiation modules. In GoogleNet, even the training is slightly different, since most top layers have their own output layer. This complexity allows the model to integrate more easily, as there is both a joint training and parallel training for the layers themselves.[7][8]
The advantages of GoogleNet: 1- GoogleNet network train faster than VGG Network. 2- The scale of a GoogleNet pretrained is comparatively smaller than that of VGG. A VGG model can contain more than 500 MB while GoogleNet has a capacity of just 96 MB.[7][8] As such GoogleNet network have no drawback, but it proposes more design improvements which will improve the model’s performance. One such improvement is called an Xception Network, in which the limit of inception module divergence is increased. Now it can be infinite in theory which is called as extreme inception.[7][8]
13) ResNet: It is a straightforward, exceptionally modularized organize engineering for Image classification. Our system is built by rehashing a structure hinder that totals a lot of changes with the equivalent topology. It expands upon the ideas of Inception and Resnet to realize an as good as ever design of Resnet.[7][8]
14) Region-Based CNN, RCNN: R-CNN system provides a solution for finding objects in an image. Here we just go over the entire image with the different block of rectangles when using this method and look at those small images via brute force procedure. The R-CNN is a unique procedure to region proposal method. That will generate almost 2000 various regions that we will need to look at. This sounds like a large number, but it is still very small compared to the sliding window approach with brute force. RCNN have 3 part. First one is R-CNN, second on is fast R-CNN and the third one is faster R-CNN. [7][8]
15) You Only Look Once, YOLO: It compared to other fast RCNNs which detect different regional proposals and thus end up in an image with multiple predictions for various regions, Yolo architecture is very similar to FCNN (fully CNN) and send the image (nxn) through the FCNN and generate output (mxm) as a prediction. A single CNN simultaneously predicts multiple bounding boxes and class probabilities for those boxes. YOLO trains on full images and directly optimizes detection performance. This unified model has several benefits over traditional methods of object detection. First, YOLO is extremely fast.[7][8][9] Since we frame detection as a regression problem, we don’t need a complex pipeline. To predict detections, we simply run our neural network in test time on a new image. With no batch processing on a Titan X GPU, the base network runs at 45 frames per seconds, and a fast version runs at more than 150 frames per seconds. Which implies that with less than milliseconds of latency, we can process streaming video almost real-time.[7][8][9] Second, when making predictions, YOLO reasons globally about the image. Unlike sliding window and region-based proposal based techniques, YOLO sees the entire image during training and testing time so it encodes contextual information about classes and their appearance implicitly. Fast R-CNN, a top detection process, mistakes background patches for artefacts in an image because the larger context is not visible. Compared to Fast R-CNN, YOLO does less than half the number of context errors.[7][8][9] Third, YOLO explores about generalizable object depictions. Trained on real images and checked on the artwork, YOLO outperforms by a wide margin the top filter method such as DPM and R-CNN. Because YOLO is highly applicable when extended to new domains or unpredictable sources it is less likely to break down.[7][8][9]
B. Frameworks
Deep learning frameworks are powering the artificial intelligence revolution. It belongs to the family of machine learning which is connected with Artificial intelligence. Its model architecture is mimicking of Human brain which has neural (neurons) networking to flow day input to output with Hidden layers and have activation function on each node to get nonlinearity in data, Deep learning believes on high computation with huge data. So as more reliable data deep learning provides better prediction. Here In this paper, I am discussing 10 best and industries favourite deep learning framework.[7][8][9]
1) Tensorflow: Google Brain has developed deep learning library in 2015. TensorFlow is a second-generation machine learning framework which is based on python which can run on CPUs, GPUs and TPUs. It is almost platform-independent and it is providing services on personal computer and cell phone devices too. TensorFlow supports C++, R and Python programming language. Most of the researchers are directly using this in his project for innovations some industry are using wrapper libraries of TensorFlow which is called Keras this is also developed my Google Team, May 2017 Google launched TensorFlow Lite for mobile devices which is basically for microcontrollers, Machine learning framework is still growing and evolving so organization are doing much innovation on small devices to make it smarter. Recently 2019 Feb, TensorFlow 2.0 released which is more friendly and lightweight in development.[10]
2) Keras: It is a high-level open-source library which is written in the python programming language. It is designed for rapid development and does experiments in industries because it is easy and friendly coding structure, on backend Tensorflow API, is supporting for development. Tensorflow is quite difficult and complex for the new IT developer, This is supporting Seq2Seq, RNN and CNN model. Microsoft also integrated CNTK with Keras and it is available in CNTK2.0 model. Keras also support development for mobile devices too. Keras has a built-in variety of neural network structures like hidden layer, activation-functions, dropout, pooling which help to production sized deep learning model most easily. [10]
3) CAFFE: It is convolutional deep learning Architecture for Fast-feature Embedding in deep learning model development. It is an open-source framework which is developed in C++ and Python programming language. Caffe has built-in several trained neural network model which user can use easily and in a very short period, he can produce it also. [10]
4) Torch: It is a deep learning development system which is based in Lua and it is an open-source framework. It uses both the C and C++ libraries and CUDA for GPUs. The torch was developed to maximize versatility and make the model development exceptionally simple. Recently, Torch’s Python implementation, called PyTorch, has found success and is rapidly gaining acceptance.[10]
5) PyTorch: It is developed with Python package to build deep neural networks and do complex tensor development. Although Torch is using Lua, PyTorch is exploiting Python’s growing popularity, enabling anyone with some simple Python programming language to get development on complex deep learning model. PyTorch enhances the architecture of Torch which makes the whole process of deep modelling simpler and more straightforward.[10]
6) Deeplearning4j: DL4J is a deep learning system which is developed in Java programming language, it is working on JVM environment, it is a simple platform and mostly it used for distributed model development in focus on the industry. The benefits of this system are java’s entire libraries support for model building, it also gets support from Big Data Frameworks like Apache Hadoop and Spark.[10]
7) MXNet: It is one of the deep learning frameworks that is most supported by languages. It is using by R, Python, C++ and Julia programming language which help to the developer not skip his comfort zone. MXNet’s is developed by C++ and CUDA, it’s memory manages by Theano framework. MXNet is very popular nowadays because it can work on multiple CUDA environment together which is more useful to industries. That is the main reason why Amazon has made it and in AWS declared the accessibility of “ONNX-MXNet” and provided support of python with deep learning libraries for building model development.[10]
8) Microsoft Cognitive Toolkit: It is known as CNTK and it is an open-source framework for building custom deep learning model. It is a sequence of the computational network like directed graph and user can easily utilize DNNs, CNNs and RNNs and LSTMs, GANs via using CNTK framework. CNTK apply SGD (Stochastic gradient descent) learning with various Graphic processing units together which enhance its capabilities. CNTK framework has python, C++ and Csharp libraries which are supporting to machine learning model development.[10]
9) deeplearn.js: It helps to train neural models on the program and build an AI model, it isn’t required substantial framework. it is created by Google Brain, it an open-source system, which is JS-based deep learning library which runs on both WebGL 1.0 as well as WebGL 2.0 too. It is a browser-based interactive machine learning tool. In this framework have a various builtin trained model which can easily customize and less training required to build new models and production it. [10]
10) BigDL: BigDL is a deep learning framework for Apache Spark, it is supporting a distributed framework and it can easily with more machine together. BigDL, support directly to build deep learning models on Hadoop or Spark. It has a various deep learning library and uses Intel’s Math Kernel Library (MKL) to guarantee superior. Utilizing BigDL, user can use pre-prepared Torch or Caffe models into Spark. Here user can add BigDL deep learning functionalities to a Big data which is stored on the various clusters.[10]
V. DEEP LEARNING AND MACHINE LEARNING IN RISK ANALYTICS
A. Risk Analytics
I have used only a few recent analytics for this survey paper which will give insight about the growing impact in entire industries.
1) Credit Risk Analytics: A modern era when most of the financial institutions are collecting data for future decisions and investments. Most of the banks are now using BigData for data storing and processing and then with the help of Machine Learning and Deep learning DataScientist are trying to know credit risk, fund tracking and effective loan granting for this purpose they are building model on stored historical customer’s data as well as they are growing business with fewer uncertainties. In new digital transformation, transparency is the main area of worries for granting loan and investment so what algorithms are using in building model which should be easy to understand and it should not behave like Blackbox and also it should be controllable for any kind of abuses. In this paper author tried to use 7 models which are Logistic Regression, Random forest, Gradient boosting and 4 Deep learning models and compared model’s performance-based on the under the curve (AUC) and RMSE. Sample data have main concern was imbalance data and due to this, it was creating biased and weakens the estimate procedure and accuracy of model’s evaluation also so applied SMOT algorithm purpose for minority over-sampled by creating ’synthetic’. In research, author had given priorities to 7 different models based on ROC, AUC and RMSE criteria. He found Random forest and Gradient boosting algorithms have better AUC and lowest RMSE in among all other 7 models. So banks can have a serious impact on decision making and to adapt and groom employees to grow business and make correct decisions with the help of predictive models and author suggest before using their algorithms, Data Scientist should use carefully these black boxes and he should have enough experience in modelling.[17]
2) Breast Cancer Analytic: Breast Cancer has become a crucial and fatal social issue and every year, it is spreading mainly in women due to increasing obesity, hormonal issues, decreasing fertility, age and genetic factors are some main reasons. Medical treatment and diagnosis is a very important step and also an expert medical officer in this domain but all are very costly and time taking process and cancer is very dangerous diseases and it’s diagnosis as early as possible can save patient better way and if patient will get the diagnosis in an early stage then it will easy to save lives as well as hard earn money both. Machine Learning and deep learning are nowadays playing a vital role in prediction in most of the diseases. [19] [20] In Machine learning, the author used ensemble learning with 5 machine learning algorithms which are Random Forest, Support Vector Machine, Naive Bayes, Kth nearest neighbour. He used UCI machine learning database for building a predictive model and also used 5 cross-fold validation for better model training. Machine Learning model was capable to predict almost more than 98 per cent which was appreciating. [11] In Deep learning, a breast cancer patient whose age is more than 65 years and physicians are taking the decision not to apply chemotherapy because it is very painful treatment so it is very difficult to survive during medical treatment so the author decided to build the model and predict diseases on early stage only. The model was able to predict disease almost more than 74 per cent which is appreciating but still in this model can enhance its accuracy.[20]
3) Heart Failure Analytic: ”Chronic Heart Failure” (CHF) is a nowadays serious issue in the whole world almost more than 20+ millions of patients are suffering from this and every year it is increasing by two per cent which is very horrifying and it is very difficult to identify via traditional medical instruments. Author research on CHF via the heart’s sounds and he used a combination of classic machine learning and deep learning method to build the model. In this method, machine learning trained by special features of data and deep learning gets train from ”spectro-temporal” of the signals. Features extracted from sound waves via OpenSMILE tool. The recording-based classifier is a combination of ML and deep learning model. In model development 10 fold cross-validation used for better training. In Machine learning Random forest used in model development. Model accuracy was more than 90 per cent but this accuracy was a little biased because the sample size was just 44. This model is encouraged to reduce Heart attack risk for future research and definitely, this research will help to build a better model.[19]
4) Diabetic Retinopathy Analytic: Diabetic Retinopathy(DR) is known as diabetic eye diseases, it is spreading in patient’s eyes due to diabetes Mellitus. DR is growing in the patient due to High blood sugar which is blocking thin eye blood vessels which help to the retina for its fitness but due to weak blood vessels, there has always a chance for bleeding from patient’s eyes then it has more chance that patients can lose his eyesight permanently but due to DR research and early prediction are saving patient’s eyes which is almost 98 per cent.[21]
In DR, many researchers already build model via different ways, one of the previous researchers was automatically scanning DR images from general computer vision techniques which is manually doing feature extracting process and then applied deep learning classification which had more manual design and more specific, long time-consuming. Now the author used CNN with MobileNet-Dense and ensemble it with existing. It reduces model complexity and increases the performance of the model. EyePACS and Messidor two partners had provided retina sample images for the model-building to the researcher. Model accuracy is around more 95 per cent which is appreciable and helping to reduces the risk of diabetic eyes disease.[21]
5) Tuberculosis Detection Analytic: Tuberculosis disease, TB is a growing world due to mainly air pollution it is very dangerous if the patient will not get treatment on time due to low immune system he can die also. in 2017 more than ten million people were infected by TB, since TB is a very common social problem so machine learning or deep learning will definitely help to minimize this disease because of early prediction. Here author used VGG16 model which is retained and he used transfer learning technique so minimal data pre-processing required and got high accuracy, He got almost more 94 per cent accuracy in his model.[22]
B. Lesson Learned
We highlighted the role of quality of data, proper computation and optimization play a vital role in innovation and production it. Predictive models are really playing a vital role in most of the industry’s innovation and companies are using data and analyzing pattern and taking better decision without financial risk.
VI. CHALLENGES AND FUTURE DIRECTION
1) Data Sample Size: Researcher’s predictive model depends on the size of sample data and computational systems, it is big question that how much size of train data used for the development of predictive model and how much redundancy in the dataset. So better model needs the quality of uniqueness in data and bigger sample size which can easily give pattern inside and such models only help in future predictions.[23]
2) Balance DataSet: Researchers are not validating properly imbalance test for sample data and in future when the model goes to production it does not provide better accuracy with future real data. [23]
3) Data Leakage: In Industry generally researchers are using third party data to build a model which have additional information which make the model inaccurate when this model reached on production and it’s accuracy reduces. To avoid this issue organization should preserve data internally and collect information as per requirement and it’s sample size should be enough big for making a better model and proper use of cross-validation. [23]
4) Pipeline Architecture: Researchers should use better pipeline structures which can help when the model goes on production because it is needed only data and it can easily transform and do prediction seamlessly.[23]
5) Model Understanding: Variety Models are exploring my industries as per need, now it is import for researchers to understand which features are more useful and helping for better model’s prediction only, those features should only use in model building and also explore inside those features and use inferences in features which help to understand better machine learning and deep learning model.[23]
6) Automate ML Model and Optimization: Machine learning model building have cumbersome tasks data cleaning, feature scaling, data transformation, feature selection and modelling and evaluation which is most common in most of the model building and most the researches struggle to do such work in day to day works so automate such work and provide better pipeline architecture which can do such work efficiently and researches can do other more significant work, in this direction already some innovation started like auto-ml , auto sklearn.[23]
7) Better Custom model training with small sample dataset: we know deep learning and machine learning required bigger size of sample data to build the better model and most of the company struggle and investing huge money in data collection now few kinds of research happened like data augmentation and transfer learning which require fewer data and better model accuracy.[23]
VII. CONCLUSION
Advanced analytic techniques are the main attention of innovators in recent years. These technology trends have added positive impact on daily life, medical treatments, researches. In this, we have reviewed the challenges of analytic tools, Machine learning and Deep learning methods. I notice better data collection (less redundant information and more unique ) help for better analysis and researchers can get better pattern inside, so in future they can easily predict for future data, It always beneficial for policy-issuers, laws regulators to take better decision with the use of analytics and overlook discriminations in the case of worst decision. I recommend the analytic matrices should be wider and follow more criteria like AUC, ROC, accuracy metrics. I have also noticed hyperparameters tuning and grid-search help to tune better custom machine learning model similarly in deep learning different activation functions, different epoch configuration, drop-out (remove biases) and different neural architecture of neural network gives a different result. Nowadays pre-trained models are playing vital role in predictions because it is already trained with huge networks and better architecture and developer need only to change input structure and less training requirements and industry can easily get the desired result with better accuracy.[8]
REFERENCES
[1] “Risks — Free Full-Text — Credit Risk Analysis Using Machine and Deep Learning Models.” [Online]. Available:https://www.mdpi.com/2227-9091/6/2/38. [Accessed: 21-Jan-2020].
[2] M. M. Najafabadi, F. Villanustre, T. M. Khoshgoftaar, N. Seliya, R.Wald, and E. Muharemagic, “Deep learning applications and challenges in big data analytics,” Journal of Big Data, vol. 2, no. 1, p. 1, Dec. 2015,doi: 10.1186/s40537-014-0007-7.
[3] P. Domingos, “A few useful things to know about machine learning,”Commun. ACM, vol. 55, no. 10, pp. 78–87, Oct. 2012, doi:10.1145/2347736.2347755.
[4] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), 2005, vol. 1, pp. 886–893 vol. 1, doi: 10.1109/CVPR.2005.177.
[5] D. G. Lowe, “Object recognition from local scale-invariant features,” in Proceedings of the Seventh IEEE International Conference on Computer Vision, 1999, vol. 2, pp. 1150–1157 vol.2, doi:10.1109/ICCV.1999.790410.
[6] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90.
[7] “10 Advanced Deep Learning Architectures Data Scientists Should Know!,” Analytics Vidhya, 09-Aug-2017. [Online]. Available: https://www.analyticsvidhya.com/blog/2017/08/10-advanced-deeplearning-architectures-data-scientists/. [Accessed: 29-Jan-2020].
[8] A. Varangaonkar, “Top 5 Deep Learning Architectures,” Packt Hub,24-Jul-2018. [Online]. Available: https://hub.packtpub.com/top-5-deeplearning-architectures/. [Accessed: 29-Jan-2020].
[9] M. Chablani, “YOLO — You only look once, real time object detection explained,” Medium, 31-Aug-2017. [Online]. Available: https://towardsdatascience.com/yolo-you-only-look-once-real-timeobject-detection-explained-492dc9230006. [Accessed: 29-Jan-2020].
[10] A. Varangaonkar, “Top 10 deep learning frameworks everyone should know,” Packt Hub, 25-May-2017. [Online]. Available: https://hub.packtpub.com/top-10-deep-learning-frameworks/. [Accessed:29-Jan-2020].
[11] H. Singh, G. S. Matharu, A. K. Dardi, and J. S. Matharu,“Empirical Investigation of Big Data Analytical Tools: Comparative Analysis,” in 2019 3rd International Conference on Trends in Electronics and Informatics (ICOEI), 2019, pp. 1264–1269, doi:10.1109/ICOEI.2019.8862739.
[12] G. T, “Google BigQuery Analytics - PDF Books.”
[13] “Combine Data Silos in Minutes,” Datameer. [Online]. Available:https://www.datameer.com/. [Accessed: 31-Jan-2020].
[14] N. Leite, I. Pedrosa, and J. Bernardino, “Open Source Business Intelligence on a SME: A Case Study using Pentaho,” in 2019 14th Iberian Conference on Information Systems and Technologies (CISTI), 2019,pp. 1–7, doi: 10.23919/CISTI.2019.8760740.
[15] S. Chen, X. Zhu, H. Zhang, C. Zhao, G. Yang, and K. Wang, “Efficient Privacy-Preserving Data Collection and Computation Offloading for Fog-Assisted IoT,” IEEE Transactions on Sustainable Computing, pp. 1–1,2020, doi: 10.1109/TSUSC.2020.2968589.
[16] S. Athmaja, M. Hanumanthappa, and V. Kavitha, “A survey of machine learning algorithms for big data analytics,” in 2017 International Conference on Innovations in Information, Embedded and Communication Systems (ICIIECS), 2017, pp. 1–4, doi: 10.1109/ICIIECS.2017.8276028.
[17] P. M. Addo, D. Guegan, and B. Hassani, “Credit Risk Analysis Using Machine and Deep Learning Models,” Risks, vol. 6, no. 2, p. 38, Jun. 2018, doi: 10.3390/risks6020038.
[18] M. Gjoreski, A. Gradiˇsek, B. Budna, M. Gams, and G. Poglajen, “Machine Learning and End-to-end Deep Learning for the Detection of Chronic Heart Failure from Heart Sounds,” IEEE Access, pp. 1–1,2020, doi: 10.1109/ACCESS.2020.2968900.
[19] H. Bed¨uk, M. S. Bed¨uk, A. Dirican, and M. ˙I. Tu?glu, “Estimating Chemotherapy Side Effect in Histopathologically Diagnosed Patients Over 65 With Metastatic Breast Cancer Using Deep Learning,” in 2019 Innovations in Intelligent Systems and Applications Conference (ASYU), 2019, pp. 1–6, doi: 10.1109/ASYU48272.2019.8946443.
[20] S. Das and D. Biswas, “Prediction of Breast Cancer Using Ensemble Learning,” in 2019 5th International Conference on Advances in Electrical Engineering (ICAEE), 2019, pp. 804–808, doi:10.1109/ICAEE48663.2019.8975544.
[21] J. Gao, C. Leung, and C. Miao, “Diabetic Retinopathy Classification Using an Efficient Convolutional Neural Network,” in 2019 IEEE International Conference on Agents (ICA), 2019, pp. 80–85, doi:10.1109/AGENTS.2019.8929191.
[22] M. Ahsan, R. Gomes, and A. Denton, “Application of a Convolutional Neural Network using transfer learning for tuberculosis detection,” in 2019 IEEE International Conference on Electro Information Technology (EIT), 2019, pp. 427–433, doi: 10.1109/EIT.2019.8833768.
[23] M. Cearns, T. Hahn, and B. T. Baune, “Recommendations and future directions for supervised machine learning in psychiatry,” Translational Psychiatry, vol. 9, no. 1, pp. 1–12, Oct. 2019, doi: 10.1038/s41398-019-0607-2.