DeepFM Ranking Model

DeepFM (Deep Factorization Machine, proposed by Huawei Noah's Ark Lab and CAS in 2017) is a landmark CTR ranking model that seamlessly integrates low-order explicit feature interactions with high-order deep non-linear representations; it features two parallel branches sharing identical embedding weights: 1) FM Component (Factorization Machine capturing 1st-order linear terms and 2nd-order dot-product interactions $\sum_{i<j} \langle v_i, v_j \rangle x_i x_j$ in $O(kd)$ linear time); 2) Deep Component (multi-layer feed-forward DNN learning high-order non-linear combinations); final CTR is given by $\hat{y} = \text{sigmoid}(y_{\text{FM}} + y_{\text{DNN}})$.

Industrial baseline for CTR prediction in e-commerce, search ads, and feed recommendation systems over sparse categorical features.

Google's Wide & Deep required labor-intensive manual feature engineering for the 'Wide' component; DeepFM eliminates manual cross-feature engineering by jointly training explicit 2nd-order FM interactions with deep networks on shared embeddings end-to-end.

Derive the mathematical proof reducing the FM 2nd-order interaction from $O(d^2)$ down to $O(kd)$ linear complexity?

Contrast DeepFM vs Wide & Deep vs FNN regarding embedding weight sharing and pre-training dependency?

How should continuous numerical features and high-cardinality categorical features be preprocessed and embedded differently in DeepFM?

Analyze DeepFM's log-loss optimization and GPU memory throughput optimization during high-concurrency online inference?

Trace the evolution from 2nd-order DeepFM to arbitrary-order explicit cross models like DCN and xDeepFM (Compressed Interaction Network)?