Our team can help you with development and launching of new product or services for Financial and Banking industry, based on cutting edge Artificial Intelligence and Machine Learning solutions.
Business Understanding — Step 0
In this step, we help our clients to determine the objectives and requirements of the project. Together, we must identify whats is client goals and tasks to accomplish and how client want to measure its success. Additionally, define the required resources and project requirements. In short, together with client, we must aim to answer “What does the business need?”.
After defining the business objectives, we determine technical goals, i.e., what success looks like from a technical perspective. A technical requirement plan must are created in this step, describing the required technologies and tools.
Next phase, to conduct a cost-benefit analysis.
A framework to conduct the cost-benefit analysis in AI projects called the expectation-complexity framework. You can learn more about this framework here: If You Consider Using AI In Your Business, Read This.
Data Understanding — Step 1
In this step, together with client, we must identify the data fields needed to train machine learning or ML models. Then, we write a plan for the customer to collect an initial set of data. The initial data is needed to explore and analyze. Many unknowns such as quality and sufficiency would be clarified during this process. In short, we need aim to answer, “What data do we have/need? Is it clean? Is it enough?”
For clients we identify required data, create a data collection plan, and provide tips to analyze the data. Anything that can help the client to accomplish the project goals.
Then, we help the client to create a plan for large-scale data collection. We document all the quality issues and surface properties of data in this step. For example, the relationships among the data must be determined. Visualization always helps to dig deeper into the data. You can read more here: On the Path to Conducting a Large Scale Data Collection.
Data Preparation — Step 2
In this step, together with client, we decide on the required data to conduct large-scale data collection. The quality of data must also be recorded during the data collection. A large-scale data collection is an expensive process; so, we make a plan to start conducting it. In short, together with client, we must aim to answer, “How does the company collect and organize the data?”
The final dataset must become clean before going to the next step. A common practice to clean or cure data is to correct, impute, or remove erroneous values. This is often the lengthiest step in the process. Without proper data curation, the project will encounter with “garbage-in, garbage-out” scenario.
Without a proper data curation, the project will encounter with “garbage-in, garbage-out” scenario.
The AI team may need to drive new attributes or features from the raw data to construct new data, a.k.a., feature engineering. According to model architecture, we may also need to combine the existing data and reformat them. For example, in many applications, string values are converted to numeric values. That helps utilize mathematical operations on textual data. A famous example of data reformatting is the Word2Vec model that is regularly used in text processing. You can read more about Word2Vec models here: Word2Vec Models are Simple Yet Revolutionary.
Model Training — Step 3
In this step, we train and assess various ML models with different algorithms (e.g., random forest, XGBoost, or deep learning). That helps determine the best modeling techniques for the problem. The selected models will need further tunning and evaluation anyway. In short, together with client, we must aim to answer, “What modeling techniques should be used?”
No model can solve all the problems. An ML model that fits problems with tabular data may not work for those with image data and vice versa. Plus, an ML model that fits problems with small datasets may not work for problems with large datasets. And, many more!
Many people think model training or building is the most important part of an AI project. This is not true, at least, anymore.
Many people think model training or building is the most important part of an AI project. This is not true, at least, anymore. Using a large set of tools and libraries, such as SciKitLearn library, the model building is summarized into a few lines of code. The AI team must compare multiple models against each other and interpret their results based on domain knowledge and performance metrics. You can learn more about the story of an ensemble classifier that wins its competition here: The Story of an Ensemble Classifier Acquired By Facebook.
Model Evaluation — Step 4
In this step, we must extensively evaluate models and identify a model that meets the business requirements. Step 3 includes a series of evaluation tasks; however, its focus was mostly on standard performance metrics. Most industrial problems require problem-specific metrics, and standard performance metrics are no longer sufficient for them. In short, together with client, we must aim to answer, “Which model best meets the business objectives?”
Many industry problems are constrained by business and technical requirements other than machine learning criteria. For example, a business case that needs responses in less than a second must ensure using low-computation techniques.
We constantly review the work accomplished using an experiment management system that helps us summarize results and identify mistakes if any.
To build an AI product, we must train a large number of models with different parameter configurations. These models are trained using a training dataset that evolves. The performance metrics can also be changed according to various business requirements. Nevertheless, we must manage this complex process and identify the best model that meets the business objectives. We use an experiment management system to manage the evaluation process. You can learn more about the experiment management system here: Why Experiment Management is the Key to Success in Data Science.
Model Deployment — Step 5
In this step, we create a thorough plan for deployment, monitoring, and maintenance. Most companies and experts look down at this step; however, that may generate issues during the operational phase. The best model must be deployed on the cloud to let customers access it. In short, together with client, we must aim to answer, “How do stakeholders access the results?”
A model is not useful unless the customer can access its results, it becomes updated to address unpredicted issues, and it would comply with the customer’s technical and business requirements. For example, cloud computing services can become costly especially for ML models that are computationally intensive.
A model is not useful unless the customer can access its results, it becomes updated to address unpredicted issues, and it would comply with the customer’s technical and business requirements.
In the end, together with client, we review the whole process by conducting a project retrospective. Together with client, we must constantly review the process by answering these questions: [1] what went well, [2] what could have been better, and [3] how to improve in the future. If you want to be successful in building AI products, you must learn about common mistakes. You can learn about four common mistakes to build an ML product here: Build An ML Product — 4 Mistakes To Avoid.
Out methodology based on articles from world known specialist Pedram Ataee, PhD
If you are interested in our experience please use a contact form: