Wide & Deep Model for Real‑Estate Purchase Intent Prediction

With the rapid development of deep learning, neural‑network‑based prediction models have become mainstream, offering high‑order non‑linear feature representation and strong fitting capabilities, yet they require large amounts of data and careful handling of high‑dimensional sparse inputs.

Google introduced the Wide & Deep model in 2016, combining a linear "wide" component that memorizes low‑order feature interactions with a deep neural network that generalizes high‑order interactions, forming a unified framework for recommendation tasks.

The wide part uses a linear model (e.g., FTRL with L1 regularization) to capture raw and crossed features, while the deep part consists of a two‑layer MLP (128→64 units) with SELU activation, processing dense embeddings of categorical features and continuous variables.

Feature engineering includes low‑frequency filtering, time‑window aggregations, and min‑max normalization, with features divided into user behavior, preference, and crossed preference groups.

Experiments on the 贝壳找房 platform show that the Wide & Deep model improves top‑x% accuracy by 31.6% over tree‑based baselines, and further gains are achieved by pre‑training embeddings for high‑dimensional sparse features.

To address the lack of temporal modeling, a GRU layer (64→32 units) is added to capture sequential user behavior, resulting in additional performance improvements.

The paper concludes with reflections on balancing model complexity and representation power, and outlines future work such as attention‑based enhancements.