🤖 AI / ML

Neural Network

neural network · neural net · artificial neural network · ann · nn · deep neural network · neuralnetwork

Definition

A neural network is a parametric machine learning model built from successive layers of linear transformations and non-linear activations, trained by backpropagation of the prediction error.

The structure of a neural network

A neural network is organised into layers. Data passes through them left to right — from the input
layer to the output layer:

Input layer — receives the features: user ID, item ID, session attributes, behavioural
embeddings.
Hidden layers — apply linear transformations (multiplication by a weight matrix) and
non-linear activations. This is where the network learns the features it actually needs.
Output layer — produces the prediction: a purchase probability, a relevance score, an
embedding vector.

Input: [user_id, item_id, session_depth, hour_of_day, ...]
  → Dense(256, ReLU)
  → Dense(128, ReLU)
  → Dense(64, ReLU)
  → Dense(1, Sigmoid)  → P(purchase) = 0.73

How a neural network trains

Training is an iterative process. At every step:

Forward pass: the data flows through the layers and the network produces a prediction.
Error computation: the loss function compares the prediction with the true answer.
Backpropagation: the gradient of the error is propagated backwards through the layers.
Weight update: an optimiser (SGD, Adam) moves the parameters in the direction that reduces
the error.

After millions of such steps the network settles on weights that minimise the error on the training
set.

Neural networks in e-commerce

Task	Architecture	Input
Recommendations	Two-Tower, MLP	User and item embeddings
Search ranking	BERT / Transformer	Text queries plus product attributes
Churn prediction	MLP, LSTM	Purchase sequence, visit frequency
Visual search	CNN + CLIP	Product images
AI assistant	Transformer (LLM)	Conversation context plus catalogue

Important: Neural networks need considerably more data to train than traditional methods such
as gradient boosting or collaborative filtering. Below roughly 100K–500K transactions a month they
often lose to simpler models because they overfit.

Neural networks versus traditional recommendation algorithms

Collaborative filtering (matrix factorization) works well on large user–item interaction datasets,
but cannot use side features such as session context, time of day or product attributes. Neural
networks combine every feature type inside one model, which is where they gain the advantage once a
rich feature space is available.

Frequently asked questions

How is a neural network different from a decision tree or gradient boosting?

Trees and boosting split the feature space into rectangular regions and work well on tabular data. Neural networks learn to build features themselves out of raw inputs — images, text, click sequences — which is decisive for tasks where manual feature engineering is impossible.

What is an activation function and why is it needed?

Without a non-linear activation function between layers — ReLU, sigmoid, tanh — a neural network would be mathematically equivalent to a single linear transformation and could not approximate complex relationships. The activation introduces non-linearity, which is what makes the model able to describe arbitrary functions.

How many layers does a recommender network need?

For ranking tasks in e-commerce, 2 to 4 hidden layers of 64 to 256 neurons are often enough. Deep learning starts at 10 or more layers and is used in computer vision and NLP. The principle: the harder the task and the larger the dataset, the more layers.