Machine Learning for Intelligent Decision Support

Apply machine learning to make data-driven decisions at a speed demanded by your business. Multidimensional problems that cannot be easily analyzed by the human brain can be resolved using a wide range of machine learning techniques. By identifying latent structures in data, revealing new insights, and making accurate predictions from data, machine learning algorithms can contextualize the information contained in huge datasets. Leveraging machine learning, you can optimize information-centric business processes, customize solutions per customer requirement, drive productivity, forecast demand, among a host of other possibilities.

Stages in Machine Learning and Steps We Take

Data Discovery

Study the Business DomainWe start by building a high-level understanding of your company.

Define the Business ProblemIs it about increasing click-through rates? Identifying an ideal location for your store? Whatever the objective, it is clearly framed.

Identify Data SourcesQualitative or quantitative, structured or unstructured, archived or streaming, data relevant to the analysis and their sources are identified.

Data Preparation

Data Cleansing & TransformationThe most labor-intensive step, this involves preprocessing data based on the analysis requirement. Checking for missing values, removing inconsistencies, normalizing datasets, we make sure the data is ready for exploration.

Tools: Data Wrangler, OpenRefine

Data Exploration

Exploratory Data Analysis (EDA)At this stage, we get an intuition of the data, identify important variables and their relationships, detect outliers, check assumptions, and choose relevant techniques for modeling. Graphical statistical techniques from the simple histogram and scatter plot to probability plot and seasonal subseries plot are used.

Tools: SPSS, Weka, R

Data Modeling

Based on the insights from EDA and the requirement, we develop predictive or descriptive models—from simple regression to deep learning. After multiple iterations, the most suitable model is created. These models are then validated with the test data and approved by our domain experts.

Tools: R, Python, Scikit, TensorFlow

Communication

Using statistical descriptions and visualization techniques, we communicate the key findings from the models implemented. With the help of interactive tools, business users can change the variables in the model and explore potential outcomes to strategize or refine their decisions.

Tools: Gephi, D3.js, GGobi

Sample Machine Learning Problems We Solve

Supervised (Known Target)

Classification

Classify emails into spam and ham
Categorize articles based on topics
Segment customers based on criteria
Detect anomalies

Regression

Predict housing prices/health outcomes based on input variables
Predict customer response to campaigns

Decision Tree, k-Nearest Neighbors, Naive Bayes, Support Vector Machines

Unsupervised (Unknown Target)

Clustering

Identify homogenous groups of customers
Determine number of customers in each group

K-Means, Fuzzy Clustering, Heirarchical Clustering

The algorithms are chosen depending on the type of data, nature of questions to be answered, size of the dataset available for learning, and computational capabilities of the system.

Deep Learning: The Smarter AI

Recognition of speech, sounds, and images comes naturally to the human brain. In deep learning, machines simulate this functionality of the brain with the aid of massive computational power and advanced algorithms. Multi-layered artificial neural networks are exposed to millions of images and sound samples from which machines automatically learn to pick out patterns. Deep learning algorithms make it possible for machines to understand spoken words in real time, recognize and describe images, play games, and even diagnose diseases more accurately.

Applications

Energy Demand Forecasting

Machine learning forecasting systems can predict future energy demand using past energy consumption data and weather parameters. Hybrid prediction models combining time-tested SARIMA models and new machine learning techniques are also evolving. Power companies can now control power generation and optimize schedules and thereby reduce costs and energy wastage.

Fraud Detection

Models built on known cases of legitimate and fraudulent transactions can assign suspicion scores for new transactions and thus help identify credit card fraud. A host of algorithms including decision trees, neural networks, regression, k-means clustering, Support Vector Machines are applied for this. Decision trees and Bayesian network are used to predict and flag fraud in insurance claims.

Predictive Maintenance

Continuous monitoring of pumps in geographically dispersed oil wells is crucial for smooth field operations. Detection algorithms can identify the deteriorating condition of pumps by analyzing the real-time vibration data against historical data. Thus operators can initiate predictive maintenance, preventing irreversible damage to assets.

Annotation of Medical Records

Although electronic health records are a rich source of patient data, they do not lend themselves to analysis as they are highly unstructured. Using Machine Learning in NLP, entities such as symptoms, diseases, and treatments can be parsed and tagged making them easily retrievable at the time of clinical decision-making.

Health Informatics

Knowledge created by medical research is more than what practitioners can cope up with. An intelligent system that incorporates NLP with semantic knowledge processing and machine learning can help practitioners look up research literature on specific problems much faster.

Medical Image Analysis

Supervised machine learning techniques are deployed in medical image analysis for computer-assisted diagnosis of certain brain disorders. Models trained on large datasets of labeled images (such as CT and MRI scans) can automatically detect indicators of a disease and help doctors make a prognosis.

Intelligent Billboards

Using real-time image recognition applications, retailers can segment customers based on their gender, age, and ethnicity. They can use this intelligence to display targeted advertisements on digital billboards to enhance brand awareness.

Product Recommendation

Content-based and collaborative filtering algorithms can be used to generate user-specific recommendations. These recommendations may include a set of similar items based on the common features of products chosen by users as well as items preferred by similar users.

Sentiment Analysis

Gauging sentiments of people from voters to customers has become vital for campaigns in fields as diverse as politics and retail. Deploying natural language processing, sentiments can be mined to help build more responsive campaigns and modify brand positioning.

START A CONVERSATION