Practical AI and eCommerce insights — recommendation engines, LLMs, EU AI Act compliance, and retail AI strategy for Irish businesses.
Curated by Michael English, Co-Founder & CTO of IMPT.io
Reading list for Irish and EU eCommerce professionals building AI-powered retail systems
Meta Description: Essential AI eCommerce research papers curated by Michael English (IMPT.io CTO). From recommendation systems to LLMs, EU AI Act, and computer vision — the academic foundations for retail AI.
Target Keywords: AI eCommerce research papers Ireland, recommendation systems papers, LLM retail research, machine learning eCommerce literature, EU AI Act research, Michael English AI eCommerce research
The AI applications transforming eCommerce today emerged from decades of academic research. Understanding the foundational papers gives practitioners context for why current architectures work, what their limitations are, and where they're headed. This curated index covers the essential literature for each major AI eCommerce domain.
Koren, Y., Bell, R., & Volinsky, C. (2009). "Matrix Factorization Techniques for Recommender Systems."
IEEE Computer, 42(8), 30–37. DOI: 10.1109/MC.2009.263
Summary: The paper that established matrix factorisation as the dominant recommendation paradigm. Koren's work winning the Netflix Prize ($1M challenge) demonstrated that SVD and ALS can capture latent user-item preferences with exceptional accuracy. The fundamental ideas in ML-based personalisation trace back here.
Why it matters: Every eCommerce recommendation engine deployed today builds on the concepts in this paper, whether directly or through extensions in deep learning frameworks.
He, X., Liao, L., Zhang, H., Nie, L., Hu, X., & Chua, T.S. (2017). "Neural Collaborative Filtering."
Proceedings of the 26th International Conference on World Wide Web (WWW). DOI: 10.1145/3038912.3052569
Summary: Introduces NCF (Neural Collaborative Filtering), which replaces the dot product in matrix factorisation with a multi-layer neural network. Demonstrates that non-linear interaction modelling significantly improves recommendation accuracy over linear methods.
Why it matters: Foundational paper for deep learning recommendations. Explains why modern recommendation models add neural layers over embeddings rather than using pure matrix factorisation.
Kang, W.C., & McAuley, J. (2018). "Self-Attentive Sequential Recommendation."
Proceedings of the IEEE International Conference on Data Mining (ICDM). DOI: 10.1109/ICDM.2018.00035
Summary: Applies transformer self-attention to model user browsing/purchase sequences. Demonstrates that recent interactions should be weighted more heavily and that items attended to in context are more predictive than simple recency. State-of-the-art on multiple benchmarks.
Why it matters: The academic basis for the transformer-based recommendation models now deployed by major eCommerce platforms. Essential reading for teams building next-generation recommendation systems.
Covington, P., Adams, J., & Sargin, E. (2016). "Deep Neural Networks for YouTube Recommendations."
Proceedings of the 10th ACM Conference on Recommender Systems (RecSys), pp. 191–198. DOI: 10.1145/2959100.2959190
Summary: Describes YouTube's two-stage recommendation architecture: candidate generation (fast, approximate, retrieve hundreds from billions) followed by ranking (slower, precise, score hundreds to final ranked list). This two-stage "candidate generation + ranking" paradigm is now industry standard.
Why it matters: The most influential practical paper on recommendation system architecture. The two-stage paradigm it describes is used by Google, Amazon, Alibaba, and virtually every major recommendation system. Essential reading for architects.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., & Polosukhin, I. (2017). "Attention Is All You Need."
Advances in Neural Information Processing Systems, 30. arXiv:1706.03762
Summary: The foundational transformer paper. While not eCommerce-specific, it introduced the self-attention mechanism that powers SASRec, BERT4Rec, and all modern sequential recommendation models.
Why it matters: Understanding the transformer architecture is prerequisite knowledge for modern recommendation system development.
Brown, T., et al. (OpenAI). (2020). "Language Models are Few-Shot Learners."
Advances in Neural Information Processing Systems, 33. arXiv:2005.14165
Summary: Introduces GPT-3 (175 billion parameters) and demonstrates that large language models can perform many tasks with few or no task-specific training examples (few-shot and zero-shot learning). The foundational paper for the current LLM revolution.
Why it matters: The conceptual basis for using commercial LLMs (GPT-4, Claude) for eCommerce tasks like product description generation and customer service without training custom models.
Lewis, P., et al. (Facebook AI). (2020). "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks."
Advances in Neural Information Processing Systems, 33. arXiv:2005.11401
Summary: Introduces RAG — combining a retrieval system (dense vector search) with a generation model (LLM) to produce accurate, knowledge-grounded responses. The LLM generates text conditioned on retrieved passages, reducing hallucination.
Why it matters: RAG is the architecture for product-accurate shopping assistants. Without RAG, LLMs hallucinate product details; with RAG, they generate grounded, accurate responses from the actual product catalogue.
Ouyang, L., et al. (OpenAI). (2022). "Training language models to follow instructions with human feedback."
Advances in Neural Information Processing Systems, 35. arXiv:2203.02155
Summary: Introduces InstructGPT (the precursor to ChatGPT), trained with Reinforcement Learning from Human Feedback (RLHF). RLHF aligns LLMs with human preferences, making them much more useful and safe for customer-facing applications.
Why it matters: Explains why commercial LLMs (GPT-4, Claude) reliably follow instructions and avoid harmful outputs in ways that earlier language models did not — critical for customer-facing eCommerce chatbots.
He, K., Zhang, X., Ren, S., & Sun, J. (2016). "Deep Residual Learning for Image Recognition."
Proceedings of the IEEE CVPR, pp. 770–778. arXiv:1512.03385
Summary: Introduces ResNet (Residual Networks) — deep neural networks using skip connections to enable training of 100+ layer networks. ResNet-50 and ResNet-101 remain widely used baselines for product image classification and visual search.
Why it matters: The foundational architecture for product image understanding. Visual search engines, try-on systems, and defect detection all build on residual network concepts.
Tan, M., & Le, Q.V. (2019). "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks."
Proceedings of the 36th International Conference on Machine Learning (ICML). arXiv:1905.11946
Summary: Demonstrates that compound scaling of CNN width, depth, and resolution produces dramatically more efficient models. EfficientNet-B4 achieves ResNet-50 accuracy with 5× fewer parameters.
Why it matters: EfficientNet is the go-to backbone for product image classification and visual search in resource-constrained environments (edge deployment, mobile).
Dosovitskiy, A., et al. (Google Brain). (2020). "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale."
ICLR 2021. arXiv:2010.11929
Summary: Adapts the transformer architecture for image recognition by treating image patches as tokens. ViT models achieve state-of-the-art on image classification when pre-trained on large datasets.
Why it matters: ViT-based models (especially CLIP from OpenAI) power the best visual search systems for eCommerce, enabling text-to-image and image-to-image product matching with strong semantic understanding.
Salinas, D., Flunkert, V., Gasthaus, J., & Januschowski, T. (Amazon). (2020). "DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks."
International Journal of Forecasting, 36(3), 1181–1191. DOI: 10.1016/j.ijforecast.2019.07.001
Summary: Amazon's probabilistic demand forecasting model using LSTMs trained jointly across many time series. Key innovation: training a single model across all SKUs learns global patterns that improve individual SKU forecasts, particularly for new or sparse-history items.
Why it matters: DeepAR is available as a built-in algorithm in AWS Forecast and is practical for Irish retailers using AWS. Particularly effective for the cold-start forecasting problem (new products with little history).
Lim, B., Arik, S.O., Loeff, N., & Pfister, T. (Google). (2021). "Temporal Fusion Transformers for Interpretable Multi-Horizon Time Series Forecasting."
International Journal of Forecasting, 37(4), 1748–1764. DOI: 10.1016/j.ijforecast.2021.03.012
Summary: The Temporal Fusion Transformer (TFT) combines multi-head attention, gated skip connections, and variable selection networks for multi-horizon forecasting. Provides interpretability (which features matter most) alongside high accuracy.
Why it matters: State-of-the-art benchmark performance on retail forecasting datasets. The interpretability is valuable for explaining forecasts to merchandising teams — not just accurate, but explainable.
Oreshkin, B.N., Carpov, D., Chapados, N., & Bengio, Y. (Element AI). (2020). "N-BEATS: Neural basis expansion analysis for interpretable time series forecasting."
ICLR 2020. arXiv:1905.10437
Summary: A pure deep learning approach to time series forecasting that decomposes each time series into trend and seasonality basis functions. Outperforms statistical methods (ARIMA, ETS) on M4 forecasting competition.
Why it matters: Often competitive with TFT while being simpler to train and interpret. Good choice for teams wanting deep learning forecasting without complex feature engineering.
European Parliament and Council. (2024). Regulation (EU) 2024/1689 on Artificial Intelligence.
Official Journal of the European Union, L 2024/1689.
Available at: eur-lex.europa.eu
Summary: The foundational regulatory text. Establishes a risk-based framework for AI regulation in the EU, prohibiting certain AI practices, imposing requirements on high-risk AI, and requiring transparency for limited-risk AI (including chatbots).
Why it matters: Every Irish eCommerce business using AI must understand the AI Act. This is the authoritative source.
Bommasani, R., et al. (Stanford Center for Research on Foundation Models). (2023). "Holistic Evaluation of Language Models."
arXiv:2211.09110.
Summary: Introduces HELM — a framework for comprehensively evaluating LLMs across accuracy, calibration, robustness, fairness, bias, and toxicity. Relevant for AI Act conformity assessments of LLM-powered eCommerce applications.
Why it matters: Provides a practical framework for the kind of comprehensive AI system evaluation that high-risk AI conformity assessments require.
Ali, M., Sapiezynski, P., Bogen, M., Korolova, A., Ravel, A., & Madio, A. (2019). "Discrimination through optimization: How Facebook's ad delivery can lead to biased outcomes."
Proceedings of the ACM on Human-Computer Interaction, 3(CSCW). DOI: 10.1145/3359301
Summary: Documents how algorithmic ad delivery can perpetuate demographic discrimination even without explicit discriminatory targeting. Relevant for eCommerce personalisation systems that might inadvertently discriminate in product recommendations or pricing.
Why it matters: The AI Act's non-discrimination requirements and GDPR's fairness principles create legal obligations to audit recommendation and pricing algorithms for discriminatory effects. This paper shows why the risk is real.
Ecommerce Europe. (2023). European eCommerce Report 2023.
Available at: ecommerce-europe.eu
Summary: Annual comprehensive report on European eCommerce market size, trends, and regional comparisons. Covers Ireland's eCommerce market share, cross-border commerce flows, payment methods, and consumer behaviour.
Why it matters: The statistical foundation for any analysis of Irish eCommerce market opportunity and AI ROI benchmarking.
Information Commissioner's Office (UK ICO). (2020). Explaining decisions made with AI.
Available at: ico.org.uk (Guidance)
Summary: While UK-based, this guidance closely tracks the EU DPC's interpretation of GDPR Article 22 for automated decision-making. Provides practical guidance on when automated decision-making requires human oversight and how to explain AI-generated decisions.
Why it matters: Irish eCommerce businesses processing EU customer data must understand GDPR Article 22 implications for personalisation and automated pricing decisions.
For eCommerce technology leaders: Start with the recommendation papers (Category 1) and LLM papers (Category 2). These define the technology frontier your systems should be heading toward.
For data scientists and ML engineers: All five categories are relevant. Priority: Koren 2009 (matrix factorisation foundation), He 2017 (neural CF), Kang & McAuley 2018 (sequential transformers), and DeepAR (for forecasting).
For compliance and legal teams: Category 5 (EU AI Act and regulation) plus the GDPR guidance are primary references.
For business and merchandising: YouTube's two-stage recommendation paper (Covington 2016) provides the most accessible high-level view of how production recommendation systems work.
Michael English curates this research index as part of IMPT.io's technology development and knowledge sharing programme. Papers selected for relevance to Irish and EU eCommerce practitioners.
impt.io | Clonmel, Co. Tipperary, Ireland
Keywords: AI eCommerce research Ireland, recommendation systems academic papers, LLM eCommerce literature, machine learning retail papers EU, EU AI Act research, computer vision eCommerce papers, Michael English AI eCommerce research