跳转至

DeepFM: shared embeddings for factorization machines and deep networks

简体中文

Core mechanism. Jointly trains second-order FM interactions and a deep branch. Private industrial features and online systems are outside scope.

Paper information

Field Details
Paper arXiv 1703.04247
Organization Huawei Noah's Ark Lab
First public release 2017-03-13 (arXiv v1)
Authors' code The paper did not provide official or author code when checked on 2026-08-09.
Adapter / method deepfm
Local reproduction code src/auto_research/reproductions/deepfm/

Original paper

Motivation and method

CTR ranking needs both low-order feature combinations and high-order nonlinear interactions. Wide & Deep needs manual wide features; DeepFM shares field embeddings across both paths end to end. The model adds first- and second-order FM terms to an MLP output, reducing manual crosses and duplicate parameters.

flowchart LR
  X["Sparse fields"] --> E["Shared embeddings"]
  E --> F["Second-order FM interactions"]
  E --> D["Deep MLP"]
  F --> S["Add logits"]
  D --> S

Key figure from the paper

DeepFM: shared embeddings for factorization machines and deep networks — Figure 1

Figure 1: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equation

\[ \hat y=\sigma\left(w_0+\sum_i w_i x_i+ \sum_{i<j}\langle v_i,v_j\rangle x_ix_j+y_{\mathrm{DNN}}\right). \]

Paper results

On Criteo and enterprise data, the paper compares AUC/log loss and reports better results than FM and deep-only models. This approved classic exception does not lower the online-evidence threshold for new industrial papers.

Local reproduction

Local comparison: The embedding MLP baseline and shared-embedding FM + deep treatment differ in the FM branch. NDCG@10 improves +23.58% relative; see metrics/movielens-100k-seeds42-44.json.

  • Data: complete downloaded MovieLens-100K, with a fixed compact local training subset.
  • Matched baseline: an embedding MLP using the same embeddings, training steps, and negative sampling.
  • Method: real second-order FM terms plus the deep branch, not a hand-built scoring proxy.
  • Run: auto-research reproduce --paper deepfm --dataset-dir data.

Across three seeds, NDCG@10 moves 0.02743→0.03390 and Hit@10 0.05556→0.06667. Results have separate JSON and Markdown artifacts; checkpoints are not committed.