Himanshu Ramchandani
Posts
7-Step Framework for Designing ML Systems - YouTube example

7-Step Framework for Designing ML Systems - YouTube example

Generative AI resources, learnings and course for Leaders/Product managers

Himanshu Ramchandani
October 24, 2023 • Estimated Reading Time: 6 minutes

Welcome Back, Hero!

It’s been a long time since we last met. I appreciate the patience you have.

If you are building ML systems, today’s content is for you. Happy Learning!

Are you a leader/product manager working on Generative AI knowledge to integrate with your business → Roadmap

Today’s Content →

How to Leverage Data, Products & AI for Your Business 🏢
7-Step Framework for Designing ML Systems for Almost All Real-World Business Problems
1 Action Tip from Data Experts for Leaders 🎬 → Prevent ML system failures.
For Developers 🧑‍💻 → Generative AI Learning resources
Career & Job in the AI field 🚀 → Websites for AI Jobs

How to Leverage Data, Products & AI for Your Business 🏢

7-Step Framework for Designing ML Systems for Almost All Real-World Business Problems

While dealing with clients I found out that, Most of the ML models fail in production.

It is crucial to create the right design for the end-to-end data project.

To tackle this, you need to follow a framework for all the business problems.

These steps will be the same but the process inside them may change depending on the business problem, domain knowledge, dataset preparation, etc.

Requirements Gathering
Business Problem to Machine Learning Assignment
Preparing Data
Developing the Model
Evaluation
Deployment
Monitoring

1 — Requirements Gathering

What is the Business Objective?
It can be, increasing the user base on the website, increasing profit, etc.
Do we have the data to use as features for the ML model?
Like recommend to friend feature or post share count on the platform.
Is the data large enough, labeled, and from where you are pulling it?
Will the cloud be used?
How big the user base is?

2 — Business Problem to Machine Learning Assignment

In the case of YouTube
Business Problem - How to increase user engagement?
ML Assignment - Increase user watch time (You have to keep the user on the platform longer)

Which ML algorithm you are going to need for the same?
Supervised, Unsupervised, Reinforcement.
Classification, Regression
Clustering, Dimensionality Reduction
Markov Models

When to use Which ML algorithm? (only for regression right now)

3 — Preparing Data

You need a Data Engineering team that can help you gather all the data from different sources.

This process will include data sources, data engineering, data storage, and ETL(Extract, Transform, Load).

Which tools to use will depend on whether the data is
- structured(go for Machine Learning Algorithms) - that have a schema (relational databases, data warehouses). names, contact details, employeID, etc.

- unstructured(go for Deep Learning Algorithms) - that have no schema (NoSQL database, Data Lakes). Text files, audio, video, image files, etc.

4 — Developing the Model

Based on the business objective create the ML assignment.
Feature Engineering Team - Extract the useful data, clean missing data, data transformation, normalization, etc.

Privacy - Is the Data Sensitve? Can we use manpower for the data or do we need to use the algorithms? Where to store the user data?

Selecting the right model →
create a base model > test it on different algorithms > pick the best one

5 — Evaluation

This is crucial as you don’t want your model to perform badly in the real world.

Performance Metrics that we use in different scenarios →

Regression - Mean Square Error, Mean Absolute Error.
Classification - Precision, Recall, F1-score, confusion matrix.
NLP - METEOR(Metric for Evaluation of Translation with Explicit ORdering)

6 — Deployment

Deploy on the cloud and check for Cost, Network Latency, hardware, privacy, and if the internet is needed 24/7

Model Compression is used to reduce the model size.

Test the model in production by different methods like, A/B testing, or shadow deployment.

7 — Monitoring

It is important to track the model and measure metrics. If the system fails we can easily know what to do.

→ why do ML models fail in production
the most common reason for failure is data distribution shift (when the model is trained on the dataset is different from the data that is used in the real world).

Keep an eye on the input and output, creating versions, is there are any drifts in the data.

PS → The only community-based data integration, infrastructure, analytics, and AI subscription that makes your competitors look outdated.