Technical Insights and Solution Strategies from the Tencent Advertising Algorithm Competition – Video Ad Track

On June 16, 2021 the second round of the Tencent Advertising Algorithm Competition began, attracting 4,335 participants from thousands of universities and enterprises, and partnering with the ACM Multimedia conference as a grand challenge for video advertising.

The competition also opened a paper submission track; papers must be submitted in PDF using the ACM article template, limited to four pages plus references, and sent to [email protected] by July 11, 23:59.

A top contestant, nicknamed “Darwin”, shares his solution process for the video ad track. In the first two weeks he focused on reproducing the baseline (0.735 AUC) and found RMSProp gave modest gains (≈0.768 AUC) compared to Adam.

Mid‑competition, he switched from TensorFlow/Keras to Paddle to gain flexibility, and concentrated on model reproduction and modality selection, ensuring proper activation, regularization, and normalization; all modalities except the cover image contributed positively.

In the later stage, he emphasized improving video and text features using pretrained models (e.g., ViT for video, OCR/ASR for text), which raised performance to 0.795‑0.803 AUC. Model ensembling via multi‑fold cross‑validation added another 0.7‑0.8 % improvement.

He notes that attempts with ResNet3D for video features yielded negligible gains and slowed inference, so they were abandoned.

The final architecture concatenates modality‑wise NextVLAD‑processed global features, applies SE fusion, and uses a single fully‑connected layer for classification; training uses Adam with cosine annealing, and dropout rates are set above 0.9 for each modality.

Future work will explore loss functions, data augmentation, hierarchical labels, and end‑to‑end training to break the current performance plateau.