Optimizing Advertising Feature Evaluation Process with the Opal Machine Learning Platform

Background : In iQIYI's advertising system, feature quality is a key factor for model performance. The existing feature evaluation process faced two main challenges: (1) a human‑resource bottleneck – the whole workflow required algorithm engineers, leading to high entry barriers, low efficiency, and inconsistent standards; (2) increasing process complexity – the growing number of advertising scenarios resulted in divergent evaluation procedures.

To improve efficiency, the advertising algorithm team collaborated with the big‑data team and leveraged the Opal machine‑learning platform. By standardizing the evaluation methodology on Opal, the iteration speed of feature evaluation increased threefold, reducing the cycle from five days to 1.5 days.

Typical Feature Evaluation Workflow (four stages) :

Data processing – data cleaning, transformation, feature construction and selection.

Sample generation – sample splitting, merging and balancing.

Model evaluation – model selection, training, and feature‑importance assessment.

Result summarization – analysis of evaluation outcomes and decision on feature rollout.

The previous workflow suffered from high evaluation thresholds, lack of a unified pipeline, and reliance on personal experience.

After migration to Opal (starting end of 2022), the new workflow introduced a complete, unified pipeline executed entirely on the platform. The optimized process is illustrated in the accompanying diagrams.

Key Steps of the Optimized Process :

Feature Production & Pre‑Evaluation

Jupyter‑Lab: notebooks are used to write scripts for large‑scale data processing and analysis, with automatic versioning.

Feature & Sample Processing: a graphical interface allows writing SQL and invoking feature operators for preprocessing, cleaning, and stitching new features.

Feature Analysis Module

Category‑feature BIAS pre‑screening: monitors model bias for each categorical feature and validates information‑gain contribution.

Offline Model Evaluation

New training templates: parameterized templates trigger distributed training tasks on iQIYI’s internal platform.

Pipeline debugging: specific nodes can be launched and retried upon failure.

Metric collection & visualization: evaluation results are sent back to Opal for standardized display.

Feature Management & Tracking

Online model feature collection: extracts features used by deployed models via model metadata.

Feature rollout tracking: provides feature queries, model queries, and routine monitoring of feature samples.

Summary & Outlook

The Opal platform has significantly lowered the barrier and cost of feature evaluation, allowing algorithm engineers to focus on iterative feature optimization and delivering cost‑effective business improvements. As feature iteration accelerates, the need for robust feature measurement grows. Future work includes building a comprehensive feature evaluation and rating system and further optimizing the feature rollout pipeline (online A/B testing, model training, and updates) on Opal.