E-commerce ML

Core concepts in this department

• Contextual Bandits

  Contextual bandits are online learning algorithms that choose an action (a price, a layout, a recommendation) given a context (user features), observe a reward, and update their policy to balance exploration and exploitation. They are the modern foundation of real-time personalization and dynamic pricing.

• Product Embeddings

  Product embeddings are dense vector representations of items in a learned semantic space, such that geometrically close items are similar in the behavioral or content sense. Transformer-based embeddings trained on session sequences capture nuanced substitute/complement relationships that simple collaborative filtering misses.

• Cold Start Problem

  The cold-start problem describes recommendation and ranking systems' inability to serve new users or new items with no interaction history. Few-shot learning, meta-learning (MAML), and prototypical networks address it by learning initializations that adapt quickly from sparse signals.

• Dynamic Pricing

  Dynamic pricing is the practice of adjusting prices in real time based on demand, inventory, user context, competition, or time. Machine-learned pricing uses contextual bandits and demand models, but introduces fairness, perception, and regulatory considerations that static pricing avoids.

• Conformal Prediction

  Conformal prediction is a model-agnostic framework for producing calibrated prediction intervals with finite-sample coverage guarantees. Applied to demand forecasting it replaces opaque point predictions with intervals that provably contain the true demand at a specified confidence level.

• Uplift Modeling

  Uplift modeling estimates the heterogeneous causal effect of a treatment — a promotion, a feature, a message — on each individual. Unlike propensity or response models, it explicitly targets the difference in outcome between the treated and untreated counterfactual, enabling promotion budgets to be spent only on the persuadable segment.

• Learning to Rank

  Learning to Rank is the class of supervised machine learning algorithms that optimize the ordering of a result set — search results, recommendations, product rankings — for revenue, engagement, or relevance. Pairwise (RankNet) and listwise (LambdaMART, ListNet) objectives are the dominant training paradigms.

• Graph Neural Networks

  Graph Neural Networks learn representations over graph-structured data by message-passing between nodes. In e-commerce cross-sell, GNNs ingest the user-item-category graph and produce recommendations that respect product hierarchy, co-purchase relationships, and session structure — outperforming flat collaborative filtering by 15–25% on business metrics.

• Real-Time Personalization

  Real-time personalization adapts product recommendations, content, and pricing within a session based on immediate behavior signals — dwell time, scroll depth, added items, search queries. Contextual-bandit systems with streaming feature stores enable policy updates in milliseconds, producing 10–25% lift over batch-trained models.

• Real-Time Fraud Detection

  Real-time fraud detection scores checkout transactions within latency budgets of 50–200 ms to decide allow, challenge, or block. Production systems combine gradient boosting (feature-rich), graph features (linked-device, shared-card), and autoencoder-based anomaly scoring under extreme class imbalance and adversarial adaptation.

• LLM-Powered Catalog Enrichment

  LLM-powered catalog enrichment uses large language models to generate product descriptions, attributes, categorization, and structured data from sparse inputs (SKU name, supplier feed) at scale. It eliminates the manual-curation bottleneck that has historically limited catalog coverage in marketplace and retail businesses.