User Preference Mining and Modeling Practices at Beike

1. Background

User preference is a concrete description of a user's intrinsic needs derived from historical behavior and data; mining these preferences helps abstract user demand from noisy information and guides search, recommendation, push, precise marketing, and fine‑grained operations.

1.2 Challenges of Preference Mining

Accurate expression – how to measure preference accuracy and handle multi‑modal (multi‑peak) preferences.

Interpretability – the output must be understandable by downstream rules, algorithms, operations staff, and agents.

High‑dimensional preferences – attributes such as geographic location are high‑dimensional and non‑ordinal, making extraction difficult.

2. Common Preference Mining Approaches

Two major families are used: statistical methods and model‑based methods.

2.1 Statistical Preference Mining

The idea is to weight user actions (with decay) and normalize the sum to obtain a preference score.

Key points:

How to assign weights to different behaviors (e.g., conversion vs. browsing).

How to define decay coefficients (exponential or step decay) to reflect the time‑sensitivity of user needs.

Challenges include difficulty proving optimal behavior weights, subjectivity in choosing decay forms, and limited optimization capability.

2.2 Model‑Based Preference Calculation

Supervised models (XGBoost, DNN, LSTM, GRU) are employed to predict the probability of heavy future actions on specific attribute items.

For low‑dimensional, equal‑length preferences, multi‑class classification works; for high‑cardinality attributes (e.g., location), embedding‑based top‑N recommendation and pair‑wise training are used.

Embedding reduces high‑dimensional sparse features to dense vectors; deep interest networks (DIN) add an activation unit to make the model output interpretable.

3. Preference Mining Practice at Beike

Beike links users, agents, and listings; mining user preferences improves personalized recommendation, targeted push, and helps agents understand user needs.

3.1 Multi‑class Preference Mining

Problem definition: predict the attribute dimension of heavy future actions based on past behavior.

Sample construction considers the balance between online and offline behavior data.

Optimization target: estimate multi‑hot or one‑hot preference probabilities.

Offline posterior metrics use inner‑product or cross‑entropy between predicted and actual behavior vectors.

Tree models and DNNs are built with features such as weighted behavior counts, decay factors, and city embeddings.

3.1.3 Incorporating User Behavior Sequences

LSTM models treat preference mining as a multivariate time‑series prediction problem, segmenting behavior into periods (day or week) and feeding the sequence into an LSTM.

City information is embedded and concatenated with the LSTM output before a fully‑connected layer produces probability scores.

3.2 Binary Preference Mining (Seq4Rec)

To handle high‑cardinality attributes, the multi‑class problem is transformed into a binary classification task using a recall set.

Seq4Rec builds user sequences of item‑action pairs, concatenates them with recall‑item embeddings, and feeds them into an LSTM‑based architecture.

3.3 Deep Interest Network (DIN) and Activation Unit

DIN introduces an activation unit that computes element‑wise differences between user sequence embeddings and target item embeddings, allowing the model to focus on the most relevant historical interactions.

Attention weights are learned via fully‑connected layers, and the final preference vector is obtained by weighted aggregation.

4 Conclusion

The article summarizes Beike's preference modeling pipeline, including statistical weighting, XGBoost/DNN, LSTM, Seq4Rec, and DIN with attention, and reports improvements in offline inner‑product and cross‑entropy metrics, with ongoing optimization efforts.

5 References

Beyond User Embedding Matrix: Learning to Hash for Modeling Large‑Scale Users in Recommendation

Deep Interest Network for Click‑Through Rate Prediction

Real‑time Personalization using Embeddings for Search Ranking at Airbnb