How Cloud Trade Fairs Use AI to Power Smart Recommendations

Intelligent recommendation is now ubiquitous in internet products, combining static user profiles (gender, age, interests) with dynamic behavior data (clicks, likes, comments, collections) to uncover deep user interests.

Unlike news or e‑commerce recommendation, the exhibition scenario must satisfy the needs of exhibitors, buyers, and personal users, especially for the "Never‑ending" cloud trade fair that extends its influence from a week to a whole year.

With nearly ten thousand registered exhibitors across 200+ sub‑industries, the system faces three key challenges: delivering precise first‑view recommendations without historical data, handling high concurrency during the offline exhibition period, and maintaining relevance for subsequent views.

Technical Foundations

The solution adopts a four‑module recommendation system serving both the website and mobile app, providing personalized recommendations for exhibitors, booths, products, and projects.

Cold‑start is addressed by leveraging user registration info, interest tags, and browsing preferences, while a knowledge base from JD Intelligent Cloud’s emergency resource platform enriches the matching process.

Cache Architecture

To meet high‑performance demands, a multi‑level cache combines Caffeine (a Java 8‑based high‑hit‑rate cache using Window‑TinyLFU) and Redis. Caffeine’s algorithm improves on LRU by protecting true hot data and using a 4‑bit CountMinSketch for efficient frequency tracking.

Data consistency across nodes relies on Redis’s publish/subscribe mechanism, achieving eventual consistency.

User and Item Profiling

User profiles are built from registration data, interest tags, and browsing habits. Item profiles include exhibitor, booth, product, and project information, with cross‑linking between them to enrich the model.

Keyword matching normalizes industry names and transaction types, supporting cold‑start scenarios by ranking based on information popularity.

Recommendation Pipeline

The system uses a two‑stage approach: recall and ranking. Recall employs collaborative filtering and matrix factorization to generate candidate sets, while ranking applies deep learning models such as Wide&Deep and DeepFM for accurate ordering.

Deep Interest Network (DIN) Model

The DIN model incorporates an attention mechanism that weights candidate items against a user's recent N behaviors, producing a weighted embedding that reflects the user's current interests.

Its architecture consists of an Embedding Layer, Pooling Layer, Concat Layer, MLP, and a Logloss loss function.

Attention computes similarity between query and key, normalizes weights, and aggregates values, enabling the model to focus on the most relevant historical actions for each candidate.

Conclusion

The intelligent recommendation component successfully delivers personalized suggestions to tens of thousands of users, handling cold‑start, high concurrency, and multi‑modal data through a combination of profiling, caching, and advanced deep learning techniques.