MMoE Model Training and Evaluation for 58.com Recruitment Recommendation Competition

Background: 58.com’s recruitment platform is a core business where improving click‑through and conversion rates is crucial; the second AI algorithm competition provides real‑world data for participants to predict user browsing and application probabilities.

Model Overview: The article introduces the Multi‑gate Mixture‑of‑Experts (MMoE) model, which extends the traditional share‑bottom architecture by using expert networks and task‑specific gating networks to capture task differences in multi‑task recommendation.

Baseline Model: A baseline MMoE model is built with feature engineering, embedding extraction, and a pyramid‑shaped network; loss is the sum of click‑through rate (CTR) and conversion rate (CVR) losses, with hyper‑parameters such as 4 experts, DNN layers per the original paper, and a pyramid network structure.

Environment & Data Upload: The WPAI platform provides a PyTorch 1.8.0 image with common Python packages. Participants upload their code (e.g., train.py) as a ZIP and place data under /workspace/mdata (train.txt, test.txt).

Data Preprocessing: Train and test files are in LIBSVM format; sequence features are removed, inconsistent feature IDs are aligned, and a 6:2:2 split creates training, validation, and test sets. Features are normalized with min‑max scaling, and label handling distinguishes CTR and CVR samples.

Model Training: Training runs with batch_size 20480, early stopping after 1000 batches without improvement, achieving AUC 68.5 for CTR and 80.5 for CVR on the test set. Sample training logs are shown below:

开始下载训练代码,
开始解压代码文件
Archive: train.zip
creating: /workspace/train/Models/
inflating: /workspace/train/__MACOSX/._Models
inflating: /workspace/train/Models/mmoe_pytorch_ai.py
inflating: /workspace/train/process.py
inflating: /workspace/train/train.py
inflating: /workspace/train/__MACOSX/._train.py
inflating: /workspace/train/util.py
解压代码文件完成
...
Test AUC:[68.5, 80.5]
模型打包路径localModelPath=/workspace/model

Evaluation: The test set is preprocessed similarly, the trained model predicts probabilities, and results are saved to submission.csv, achieving a competition score of 0.729421.

Department Info: The AI Lab of 58.com, part of the TEG technology platform, focuses on applying AI to improve business efficiency and user experience.