The use of data to drive positive business results is widely debated and many companies have recognized the potential of artificial intelligence and machine learning in particular to implement powerful data-driven solutions. In this article, I’ll introduce you to some machine learning case studies that will help you understand how machine learning is used to solve business problems.
In practice, companies can benefit from a wide range of machine learning applications, but it is not always clear which solutions are the most feasible and efficient to achieve the desired results. The best techniques available to solve a problem vary depending on the goals, resources, and data of each business.
Machine Learning Case Studies
To get a feel for how machine learning can create business value in different industries using different techniques, I have summarized some machine learning case studies that I have practically conducted.
Each case study describes several business challenges, and all of these machine learning case studies below describe the custom machine learning systems that I have built to solve the respective business problems with the approaches that I have. I chose to solve them.
While working on practical business issues, I have learnt that solving business issues follows some common steps:
- Understand the business challenge.
- Evaluate different approaches to solve the challenge.
- Build proofs of concepts or production-ready systems
- Evaluate the results and perform additional iterations, if necessary.
5 Machine Learning Case Studies
Below are the 2 machine learning case studies that consist of end-to-end solutions that I implemented from scratch and put into production.
I hope the following case studies will help you familiarize yourself with the tremendous opportunities that exist around using data to drive business results and encourage you to embark on a machine learning business in your organization.
Case Study 1: Setting Retail Prices:
The solution to this business problem was designed for a large online consignment market. Before using machine learning, the prices of unique products in an entire catalogue are determined manually in an extremely lengthy process. Pricing is neither cost-effective nor consistent.
After using machine learning, more than half of the products in the catalogue are billed automatically. Prices are determined based on product images, prices of visually similar products, product descriptions, and other attributes.
The case here was to build, train, and deploy a machine learning pipeline with an entirely new infrastructure capable of pricing new items based on product images, product information, and historical sales data.
To solve this business problem, I trained convolutional neural networks to classify products and extract relevant features from images (eg, pattern, neckline, sleeve length). Using image similarity and K nearest neighbors (k-NN) to find the most visually similar products and their prices in the sales database to provide quality data for the algorithm.
Case Study 2: Scaling Image Processing:
This solution was designed for a business problem of a risk management company. Before using machine learning, manual analysis of photos of building rooftops taken by drones to detect damage. The services offered by the company was not scalable due to the tedious nature and limited human resources.
After using machine learning to solve this problem, the system was able to understand the images and automatically identify problems on the roofs, including pooling of water, loose cables, and rust.
The case here was to develop a deep neural network capable of detecting several problems in the imagery of the roof, as well as a pipeline to analyze the incoming images and an API to provide access to the results to external tools.
Solving this problem was not an easy task, given the absence of a preexisting dataset, the first step is to create a custom dataset. To facilitate the marking process, the segmentation problem is approached as a classification problem by classifying small image patches instead of a pixel by pixel.
Even though this is a much faster process, creating a dataset large enough to train a deep neural network from scratch is still very expensive. To improve the training process, a pre-trained ResNet-18 model on ImageNet is used.
When applying the transfer learning technique, the model is trained to capture relevant information from a large data set (ImageNet), then refined on a small data set to learn specific information of its own. to the problem in question.
The model mentioned above is used as a feature extractor. Once good patch representations are generated, a simpler classifier is trained on them to generate the final result.
I hope you liked this article on Machine Learning case studies to learn about the application of Machine Learning in business problems. Feel free to ask your valuable questions in the comments section below.