Practical Applications and Challenges of Machine Learning and AI at QCon Beijing 2019

On May 6, 2019, QCon Beijing hosted a technical session titled “Algorithm and AI Application Practices” where four experts from Beike shared insights on machine learning for user profiling, deep learning for house‑quality scoring, intelligent customer service, and AI‑driven floor‑plan generation.

1. Machine Learning in User Profiling and Challenges

Beike serves over 30 million monthly users with more than 1.8 billion listings and 200 k agents. To improve user experience, the platform builds a multi‑dimensional user‑profile system consisting of factual, mining, and predictive tags.

Factual tags record observable behaviors (e.g., number of viewed listings). Mining tags infer motivations behind actions, while predictive tags forecast future behaviors.

The profiling system is organized into three layers: business (recording user actions), data (collecting and structuring logs), and computation (modeling and analysis). Four major challenges are identified: unified event standards across front‑ and back‑ends, stable storage and computation at massive scale, unified data abstraction for multi‑dimensional modeling, and evaluation of profiling accuracy.

Beike addresses these by aggregating data from multiple sources to enrich user records, clustering GPS data to infer work and residence locations, and applying SMOTE sampling for predictive tagging. The overall tag‑framework consists of four layers: data collection, data processing (ID mapping, cleaning), tag calculation (statistics, abstraction, modeling), and storage.

2. Deep Learning for House‑Quality Scoring and Algorithm Optimization

Beike’s AI‑Find‑House feature ranks listings by predicted transaction probability. The target Y is defined as recent transaction‑related actions; X includes static property features. A two‑week prediction window was found optimal.

Model evolution progressed through three stages: (1) an initial XGBoost model, (2) a DNN+RNN hybrid model, and (3) continuous business‑driven optimization.

The current DNN+LSTM architecture (5 DNN layers + 1 LSTM layer) reduces feature engineering to 21 dimensions while leveraging batch normalization and dropout.

To explain scores to agents, Beike visualizes five dimensions (static attributes, owner intent, cost‑effectiveness, market heat, customer interest) via radar charts, guiding agents to improve listing quality.

3. Intelligent Customer Service System Construction and Algorithm Iteration

The NLP‑driven intelligent客服 system reduces human support costs by automating knowledge‑base queries. Its architecture comprises four modules: data (knowledge base and graph generation), access layer (error correction, sentiment analysis), central‑control & task layers (core routing), and an evaluation metric system.

Core NLU identifies intents, routing requests to QABot (knowledge Q&A), TaskBot (tool usage), or ChatBot (small talk). Retrieval combines keyword‑based BM25 and semantic Faiss vector search. Intent filtering removes irrelevant candidates, and a DSSM model performs deep semantic matching.

4. AI for Automatic Floor‑Plan Generation

Beike leverages a self‑developed device to capture 2D/3D data and applies AI to generate 3D models and floor plans. Two approaches are discussed:

FloorNet : a multi‑branch network (PointNet, FCN‑based floorplan, image branch) processing point clouds (originally 50 k points, later increased to 200 k) to produce vectorized floor plans.

GAN‑based method : uses Pix2Pix‑style generative adversarial networks to translate point‑cloud density maps into raster floor plans, followed by vectorization.

Comparative analysis shows FloorNet yields higher accuracy but slower speed, while GAN is faster but less precise. Improvement directions include increasing point‑cloud density, refining loss functions, and simplifying post‑processing.

5. Recap of the Session

The session highlighted Beike’s end‑to‑end AI pipeline: massive data collection, multi‑layer tagging, deep‑learning model evolution, intelligent客服, and advanced floor‑plan reconstruction, illustrating how AI can empower real‑estate services.