Practical AI and eCommerce insights — recommendation engines, LLMs, EU AI Act compliance, and retail AI strategy for Irish businesses.
Compiled by Michael English, Co-Founder & CTO of IMPT.io
Reference Guide for Irish and EU eCommerce Professionals
Meta Description: Comprehensive AI eCommerce glossary by Michael English (IMPT.io). 70+ terms covering LLMs, recommendation systems, computer vision, EU AI Act, and personalisation for Irish and EU retailers.
Target Keywords: AI eCommerce glossary Ireland, machine learning retail terminology, recommendation systems terms, EU AI Act glossary, eCommerce AI definitions Ireland, Michael English AI eCommerce glossary
A/B Testing: A controlled experiment comparing two versions of a system (A = control, B = treatment) by randomly assigning users and measuring outcomes. The gold standard for measuring AI system impact.
Additionality (AI context): The incremental benefit attributable to an AI system above what would have occurred without it. Measured via A/B testing.
ALS (Alternating Least Squares): An optimisation algorithm for training matrix factorisation recommendation models. Efficient for implicit feedback (purchases, clicks) data. Used in Spark MLlib and the implicit Python library.
AMM (Automated Market Maker): In DeFi carbon markets, a smart contract-based liquidity pool enabling continuous token trading. Applicable to tokenised carbon credit markets.
Anomaly Detection: ML technique identifying unusual patterns — used in fraud detection to flag transactions statistically different from the user's historical behaviour.
API (Application Programming Interface): A way for software systems to communicate. LLMs from OpenAI (GPT-4), Anthropic (Claude), and Google (Gemini) are accessed via APIs, enabling eCommerce applications without building AI from scratch.
AUC-ROC: Area Under the Receiver Operating Characteristic Curve. A model evaluation metric for classification tasks (e.g., fraud detection, churn prediction). AUC of 0.5 = random; 1.0 = perfect.
Batch Inference: Running AI predictions in bulk (e.g., computing all product recommendations for the next day overnight). Cheaper and simpler than real-time inference; adequate for non-time-sensitive applications like email personalisation.
BERT (Bidirectional Encoder Representations from Transformers): Google's pre-trained language model (2018). Foundation for semantic search and many NLP applications. BERT4Rec adapts it for sequential recommendation.
Bias (ML): Systematic errors in model predictions. In eCommerce, popularity bias (recommending only bestsellers) reduces coverage and diversity. Fairness bias can cause discriminatory pricing or recommendations.
Bounce Rate: The percentage of visitors who leave after viewing only one page. AI-enhanced search and recommendations typically reduce bounce rate by improving relevance.
BPR (Bayesian Personalised Ranking): A training objective for recommendation systems that optimises pairwise rankings (preferred items over unpreferred) rather than point predictions.
Candidate Generation: The first stage of a two-stage recommendation pipeline: generating hundreds of plausible candidates from millions of items using fast approximate methods.
Chatbot: Automated conversational agent. Under EU AI Act, all chatbots must disclose they are AI systems (effective August 2027 for existing systems).
CLV/LTV (Customer Lifetime Value): The total revenue expected from a customer over their entire relationship with a retailer. AI-driven LTV prediction enables more intelligent acquisition spend and personalisation prioritisation.
Cold-Start Problem: The challenge of making recommendations for new users (no interaction history) or new items (no engagement data). Addressed through content-based bootstrapping, onboarding preference capture, and session-based recommendations.
Collaborative Filtering (CF): Recommendation approach using patterns of user-item interactions to identify similar users or items. "Customers who bought this also bought..."
Computer Vision: AI field enabling machines to interpret visual information from images and video. Applications in eCommerce: product search, try-on, defect detection, size recommendation.
Content-Based Filtering: Recommendation approach using item attributes (category, price, material, colour) to recommend items similar to those a user has interacted with.
Contextual Bandit: A reinforcement learning approach for real-time personalisation that learns which recommendation policy works best for each context (user profile, page position, session stage).
Conversion Rate (CVR): The percentage of visitors who complete a desired action (typically purchase). Primary commercial metric for evaluating recommendation engine performance.
Cosine Similarity: A measure of similarity between two vectors (ranging from -1 to 1). Widely used in recommendation systems and semantic search to measure how similar user/item embeddings are.
CRYSTALS-Kyber: Post-quantum cryptographic algorithm (now standardised as ML-KEM, FIPS 203). Used in IMPT.io's infrastructure for quantum-safe key exchange. Not directly an eCommerce term but relevant to secure transaction infrastructure.
Data Augmentation: Techniques for artificially increasing training data size. In computer vision, this includes image rotation, cropping, colour jittering — important for training product image classifiers with limited labelled data.
Decision Tree: A simple ML model that splits data based on feature thresholds. Used in fraud detection and churn prediction; interpretable (can explain why a decision was made).
DeepAR: Amazon's probabilistic deep learning forecasting algorithm. Uses LSTM networks to model time series. Available as a built-in algorithm in AWS Forecast. Particularly effective for seasonal retail demand.
Deep Learning: A subset of machine learning using neural networks with many layers (10s–100s). Powers modern recommendation engines, NLP, and computer vision in eCommerce.
Dynamic Pricing: Algorithm-driven real-time price adjustments based on demand, competition, and inventory signals. Subject to EU Omnibus Directive (30-day price history display requirement) and AI Act scrutiny if targeting vulnerable customers.
Embedding: A dense numerical vector representing an item, user, or piece of text in a continuous semantic space. Similar things have similar embeddings. The fundamental representation in modern recommendation systems and semantic search.
Embedding Model: A model that converts raw inputs (text, images, products) to embeddings. OpenAI's text-embedding-3-small, Sentence-BERT, and CLIP are common choices.
EU AI Act: Regulation (EU) 2024/1689, the world's first comprehensive AI regulation. Prohibits certain AI practices from February 2025; requires chatbot disclosure (August 2027); establishes high-risk AI requirements (August 2026).
EU Omnibus Directive: EU consumer protection directive requiring eCommerce retailers to display the lowest price from the previous 30 days before applying a discount. Creates compliance requirements for dynamic pricing systems.
Explainability (XAI): The capability to explain AI decisions in human-understandable terms. EU AI Act encourages transparency; GDPR Article 22 requires explanations for automated decisions significantly affecting individuals.
Faiss (Facebook AI Similarity Search): An open-source library for efficient similarity search at billion-scale. Enables sub-50ms approximate nearest-neighbour search in high-dimensional embedding spaces — essential for production recommendation systems.
Feature Engineering: The process of creating informative input variables (features) for ML models. In demand forecasting, features include day of week, promotional flags, weather, and competitor stockout events.
Feature Store: A centralised system for storing, managing, and serving ML features. Enables consistent feature computation between training and production; reduces training-serving skew.
Fine-Tuning: Adapting a pre-trained model to a specific task or domain by training on task-specific data. Fine-tuning a vision model on fashion images improves visual search accuracy significantly.
Fraud Detection: ML systems identifying potentially fraudulent transactions in real-time. Typically ensemble models (XGBoost, LightGBM) trained on historical labelled fraud data. Must process in <200ms at payment gateway.
GDPR (General Data Protection Regulation): EU Regulation 2016/679 governing personal data processing. Interacts with AI Act for eCommerce: requires lawful basis for ML personalisation, transparency, and data subject rights.
Generative AI: AI models that generate new content (text, images, code, audio) rather than classifying existing content. GPT-4, Claude, and Gemini are text generative AI; DALL-E, Midjourney, and Stable Diffusion are image generative AI.
GPU (Graphics Processing Unit): Highly parallel processor originally designed for graphics but now the primary compute hardware for training and running large ML models. Critical for deep learning recommendation models and LLMs.
Gradient Boosting: A family of powerful ML algorithms (XGBoost, LightGBM, CatBoost) that build ensembles of decision trees iteratively. Industry standard for tabular data tasks like fraud detection and demand forecasting.
GRU (Gated Recurrent Unit): A type of recurrent neural network cell that captures sequential dependencies. Used in session-based recommendation models to model user browsing sequences.
Hallucination (LLM): When an LLM generates confident but factually incorrect text. Critical issue for eCommerce applications — an LLM that fabricates product specifications or prices erodes customer trust. Mitigated by RAG (retrieval-augmented generation) and guardrails.
Hybrid Recommendation: Combining collaborative filtering and content-based approaches. Better than either alone, especially for handling cold-start and providing diverse recommendations.
Hyperparameter: A model configuration parameter set before training (e.g., learning rate, number of layers, embedding dimension). Tuned through cross-validation or AutoML to optimise model performance.
Implicit Feedback: User interactions inferred from behaviour (purchases, clicks, time spent, add-to-cart) rather than explicitly stated ratings. More abundant than explicit feedback; the primary recommendation signal in most eCommerce systems.
Inference: Running a trained ML model on new data to generate predictions. Distinguished from training (learning from historical data). Real-time inference must be fast (sub-100ms for recommendation, <200ms for fraud detection).
Intent Classification: Determining the purpose behind a user's input. In conversational commerce, classifying customer messages as "product query", "order status", "return request", etc. to route to appropriate handler.
Keras: A high-level deep learning framework (now part of TensorFlow). Commonly used for building recommendation system prototypes.
KPI (Key Performance Indicator): Measurable values indicating performance. Key AI eCommerce KPIs: conversion rate, AOV, recommendation CTR, stockout rate, CSAT.
Latent Factor: A hidden dimension in matrix factorisation models representing an unobserved abstract concept (e.g., "casualness", "sports utility"). Users and items with similar latent factors are recommended together.
LightGBM: A high-performance gradient boosting framework by Microsoft. Industry standard for tabular ML tasks; often used for demand forecasting and fraud detection in eCommerce.
Linear Reduction Factor (ETS): Not AI-specific; the annual cap reduction rate in EU ETS. Included because carbon market infrastructure at IMPT.io intersects eCommerce AI.
LLM (Large Language Model): Neural language models trained on vast text corpora, capable of understanding and generating human-quality text. GPT-4 (OpenAI), Claude (Anthropic), Gemini (Google) are leading commercial LLMs. Transforming eCommerce through natural language product search, customer service automation, and content generation.
LSTM (Long Short-Term Memory): A type of recurrent neural network capable of learning long-range sequential dependencies. Used in demand forecasting and session-based recommendation systems.
MAE (Mean Absolute Error): The average absolute difference between predicted and actual values. A key metric for demand forecasting accuracy (often expressed as MAPE — Mean Absolute Percentage Error).
Matrix Factorisation: A technique decomposing the user-item interaction matrix into low-dimensional user and item embeddings. The mathematical foundation of collaborative filtering recommendation systems.
MAPE (Mean Absolute Percentage Error): (|Actual - Forecast| / Actual) × 100%. A standard accuracy metric for demand forecasting. Industry benchmark for retail: 15-25% MAPE for SKU-level daily forecasting.
Meta-Learning: Training models to quickly adapt to new tasks with few examples. Relevant for eCommerce cold-start: rapid personalisation for new users.
Multi-Armed Bandit: A reinforcement learning paradigm for sequential decision-making under uncertainty. Used for personalised homepage layout, email send-time optimisation, and dynamic discount optimisation.
NCF (Neural Collaborative Filtering): A deep learning model combining matrix factorisation with multi-layer perceptron for non-linear user-item interaction modelling.
NLP (Natural Language Processing): AI field enabling machines to understand and generate human language. Powers eCommerce applications: product search, chatbots, review analysis, SEO content generation.
Neural Architecture Search (NAS): Automated ML technique that learns the optimal neural network architecture for a given task, reducing the need for manual architecture design.
Normalised Discounted Cumulative Gain (NDCG): A ranking quality metric measuring how well a recommendation system ranks relevant items at the top of recommendations. Higher NDCG = better ranking quality.
Omnibus Directive: See EU Omnibus Directive.
Online Learning: Training ML models in real-time as new data arrives, rather than batch retraining. Enables immediate adaptation to demand changes or user behaviour shifts.
Open Quantum Safe (OQS): Open-source library implementing post-quantum cryptographic algorithms including ML-KEM and ML-DSA. Available at openquantumsafe.org.
Personalisation: Adapting digital experiences (product recommendations, email content, pricing, search results) to individual users based on their behaviour, preferences, and context.
Precision@K: A recommendation evaluation metric: the fraction of top-K recommended items that are relevant. High precision means few irrelevant recommendations.
Prophet (Facebook): An open-source time series forecasting library by Meta. Handles seasonality, holidays, and trend changes well; popular for retail demand forecasting.
PyTorch: Open-source ML framework by Meta. Industry standard for research and custom deep learning model development. Used for building custom recommendation systems and NLP models.
RAG (Retrieval-Augmented Generation): An architecture that grounds LLM responses in specific documents or data by retrieving relevant context before generating a response. Essential for accurate product information in AI shopping assistants — prevents hallucination about specific products.
Recall@K: A recommendation evaluation metric: the fraction of all relevant items that appear in the top-K recommendations. Balances with Precision@K to evaluate comprehensiveness.
Recommendation Engine/System: An AI system predicting which items a user is most likely to interact with or purchase. Among the highest-ROI AI investments in eCommerce.
Reranking: A second stage in recommendation systems that takes a candidate set and applies more complex scoring (personalised, contextual) to reorder for the final displayed list.
Return Rate: The percentage of purchased items returned by customers. A key metric for computer vision ROI (try-on, fit prediction reduce returns). Irish eCommerce fashion return rates: 25-35%.
SASRec (Self-Attentive Sequential Recommendation): A transformer-based recommendation model that uses self-attention to model user behaviour as a sequence, weighting which past interactions most influence current recommendations.
Semantic Search: Search that understands the meaning of queries rather than matching keywords. Enables "waterproof jacket for hiking" to match products tagged with "technical outdoor apparel" even without keyword overlap.
Session-Based Recommendation: Recommendations based solely on the current browsing session (no historical user data). Essential for anonymous users and cold-start situations.
Softmax: A mathematical function converting raw model outputs to probabilities summing to 1. Used in the final layer of classification models (fraud, intent classification) and recommendation scoring.
SMOTE (Synthetic Minority Oversampling Technique): A technique for addressing imbalanced classification data (e.g., fraud datasets where fraud is 0.1% of transactions). Generates synthetic minority class examples.
Supply Chain AI: ML applications for demand forecasting, inventory optimisation, supplier risk management, and logistics routing.
TensorFlow: Google's open-source ML framework. TensorFlow Recommenders (TFRS) provides high-level APIs for building recommendation systems.
Temporal Fusion Transformer (TFT): A state-of-the-art time series forecasting model by Google Research. Combines multi-head attention, LSTMs, and variable selection networks for accurate demand forecasting with interpretability.
Transfer Learning: Using a model trained on one task as a starting point for another. Fine-tuning pre-trained vision models (EfficientNet, ViT) on product images reduces training data requirements.
Two-Tower Model: A recommendation architecture with separate encoder networks for users and items, enabling pre-computation of item embeddings and fast approximate nearest-neighbour retrieval at scale.
User Embedding: A learned numerical representation of a user in a continuous vector space, capturing their preferences and behaviour patterns. The foundation of personalised recommendation systems.
UV (User-Item Interaction Matrix): The sparse matrix recording all known interactions between users and items. The input to collaborative filtering and matrix factorisation recommendation systems.
Vector Database: A database optimised for storing and searching high-dimensional embedding vectors. Pinecone, Weaviate, Milvus, and Qdrant are popular choices for semantic search and recommendation candidate retrieval.
Virtual Try-On: AI technology enabling customers to visualise products (clothing, cosmetics, eyewear) on themselves or virtual models before purchasing. Key technology for reducing eCommerce returns.
Visual Search: Finding products by uploading an image and finding visually similar items in the catalogue. Enables "I saw this on Instagram, where can I buy it?" commerce scenarios.
Verra (VCS): Voluntary carbon standard; not directly an AI term but included for IMPT.io context. Verra VCS credits are tokenised in blockchain carbon markets.
WhatsApp Business API: Meta's API for businesses to send/receive WhatsApp messages programmatically. Enables conversational commerce (product enquiries, order updates, customer service) via WhatsApp.
Word2Vec: An early embedding model by Google (2013) that learned word semantics from large text corpora. Conceptual predecessor to modern transformer-based embeddings.
XGBoost: An efficient and powerful gradient boosting framework. Consistently high-performing on tabular data tasks; widely used for fraud detection, churn prediction, and demand forecasting.
Compiled by Michael English, Co-Founder & CTO of IMPT.io. Updated for EU AI Act (2024), LLM integration, and current Irish eCommerce landscape.
impt.io | Clonmel, Co. Tipperary, Ireland
Keywords: AI eCommerce glossary Ireland, machine learning retail terms, recommendation systems glossary, EU AI Act terms eCommerce, LLM eCommerce glossary, computer vision retail terms Ireland, Michael English AI eCommerce glossary