Explainable Recommendation: Background, Development, Graph‑Based Structured Explanations, and Natural Language Generation Advances

Guest Speaker: Wang Xiting, PhD, Microsoft Research Asia (edited by Ma Yue, KuaiShou) – presented by DataFunTalk.

Introduction: Explainable recommendation is a new direction in recommender systems, closely related to knowledge graphs and natural language understanding.

01. Background of Explainable Recommendation

Traditional recommender systems predict a score for a user‑item pair but provide no rationale, making recommendations feel abrupt. Explainable recommendation answers the "why" by offering user‑friendly reasons (e.g., popularity, similarity, proximity), which improves trust and click‑through rates. Major companies such as Meta, Amazon, Ele.me, Last FM, and Microsoft have incorporated explanations to boost user experience.

02. Development History

Before 2015: Mostly template‑based explanations with low diversity and high manual effort.

2016 onward: Shift to free‑form explanations using knowledge graphs and natural language generation (NLG). Graph‑based explanations provide structured reasoning paths, while NLG offers fluent textual reasons.

03. Progress in Graph‑Based Structured Explanations

Knowledge graphs (KG) enrich recommendation models with abundant semantic information, improving both accuracy and interpretability.

Two main KG techniques:

KG Embedding: Represents entities and relations as vectors, integrating topology, textual, and visual features into item representations. Advantages: easy integration; drawbacks: loss of multi‑hop user‑item connections and limited interpretability.

Deep‑Learning Reasoning: Models reasoning paths on the KG. Notable works include KPRN (AAAI 2019), PGPR (Policy‑Guided Path Reasoning), and reinforcement‑learning based methods that assign rewards to paths. These approaches generate transparent, step‑by‑step explanations but face challenges such as noisy candidate paths and optimization difficulty.

Hierarchical reasoning and multi‑level KG representations are also explored to capture high‑level relationships (e.g., brand categories) and improve explanation quality.

04. Progress in Natural Language Generation (Unstructured) Explanations

Early NLG relied on hand‑crafted templates; later retrieval‑based methods selected existing sentences, which suffered from copyright and personalization issues. The latest trend uses large pre‑trained language models (e.g., BERT, GPT) fine‑tuned for recommendation text generation. Training pipelines typically involve pre‑training on Wikipedia, supervised fine‑tuning on human‑written ad copies, and reinforcement learning (e.g., masked‑sequence policy gradient) to directly optimize click‑through or relevance metrics.

These generative systems can produce diverse, personalized explanations without template constraints, though they may still generate grammatical errors or less engaging content.

05. Summary and Open Challenges

Fact‑checking: Ensuring the correctness of generated information.

Low‑resource scenarios: Generating high‑quality explanations with few or no ground‑truth examples.

Evaluation: Moving beyond subjective user studies toward reliable offline metrics.

Future work should involve users in the loop, leveraging feedback (clicks, dwell time, explicit comments) to continuously refine both recommendation scores and explanations.

Thank you for listening!