A strategic initiative to reduce the cost of operations for OEM, along with provisioning for just in time inventory availability and ensuring to meet the targeted production plans.
Initiative to leverage machine learning to predict demand, enhance production continuity, minimize inventory holding costs, and enable data-driven decision-making in vehicle manufacturing.
From Data to Decisions: Our Journey Through Machine Learning to Neural Networks
When it comes to any AI or ML case study, the foundation of success lies in one thing: data. Before jumping into models and predictions, it’s crucial to invest time in deep data analysis. This includes ensuring the data is in the right format—through data loading, cleaning, and transformation—which directly impacts the quality of predictions and minimizes errors during training.
Phase 1: Solving with Machine Learning
Since our task was regression, we explored several classical ML models with hyperparameter tuning, including Linear Regression and different suitable Machine Learning Models. Random Forest consistently delivered the best results, effectively capturing the dataset’s complexity.
Evaluation Metrics Used:
- Mean Squared Error (MSE)
- R² Score
- Mean Absolute Percentage Error (MAPE)
- Symmetric MAPE (sMAPE)
These metrics offered a comprehensive view of model performance and guided further optimization.
Moving Beyond: Why We Chose Neural Networks
While Random Forest offered strong results, our goal extended beyond just good performance on historical data. We wanted a model that could capture real-time complexity and generalize well to unseen or edge-case scenarios—areas where traditional ML models often struggle or “hallucinate” with overconfident predictions.
To address this, we transitioned to Neural Networks, which are inherently better at learning deeper, non-linear patterns in data. This shift allowed us to model more intricate relationships and prepare for production environments where data can be messy, dynamic, and unpredictable.
What are Neural Networks?
Neural Network is a subfield of Deep Learning. A neural network is a type of machine learning model inspired by how the human brain works. It’s made up of layers of interconnected “neurons” that process and transform data step by step—learning to make predictions or decisions based on examples
How This Neural Network Works – A Simple Walkthrough
The diagram above shows a deep neural network architecture designed to process a combination of numerical and categorical features.
The model processes both numerical and categorical data by embedding categorical IDs and combining them with raw features into a unified input. It uses a deep neural network with multiple hidden layers to learn complex patterns, leveraging techniques like batch normalization, dropout, and activation for stability, regularization, and non-linearity. The final output layer generates a single prediction, making it well suited for regression tasks.
Overview:
- Input Layer:
- Numerical features are passed directly on.
- Categorical ID columns (e.g., part IDs) are label encoded and then transformed via an embedding layer before being concatenated with numerical inputs.
- Hidden Layers:
- Four fully connected layers:
- Layer 1: 128 Neurons → BatchNorm → Dropout → Activation Function
- Layer 2: 64 Neurons → BatchNorm → Dropout → Activation Function
- Layer 3: 32 Neurons → Activation Function
- Layer 4: 1 Neurons → Activation Function
- Activation Function: ReLU
- Regularization: Dropout and Batch Normalization are used to improve generalization and stabilize training.
- Four fully connected layers:
- Output Layer:
- Produces a single value per input (e.g., forecast or regression target) with shape [batch_size, 1].
This architecture balances feature representation (via embeddings), depth (for complexity), and regularization, making it well-suited for forecasting, recommendation systems, or demand prediction.
Summary
This is a deep feedforward neural network that intelligently combines numerical and categorical inputs using embedding layers, and improves performance using batch normalization, dropout, and ReLU activations making it well-suited for tasks like forecasting, classification, or recommendation