Donghang Yin

QQ Kandian(match&pre-rank):

• Match: Introduced VQVAE techs to news recommendation which models the item clustering via user behaviors, bringing +0.9% in duration and +0.6% in PV.
• Pre-ranking: developed a meta-learning-based two-tower model which can rapidly adopt each user's preferences with a few consumed items and achieved +1.4% in PV and +1.2% in duration.

QQ Kandian(rank, director):

• Utilized a multimodal adversarial network to learn modality-invariant and modality-specific representations and enhanced the duration by 1.93%.
• Researched a learnable feature selection method based on variational dropout and effectively selected top features in ranking model, reducing 40% online response time and machine usage.
• Developed a multi-stream user interest model which utilized both positive and negative user behaviors, bringing +1.6% PV, +1.4% duration.
• Optimized the duration estimate problem via a series of binary classification sub-problems which can be seen as a ordinal regression problem, bringing +1.1% in duration.

Wechat moments display ads(director):

• Optimized two-tower model structure which made short paths for specific features to participate final interaction, increasing GAUC by 0.003 and GMV by 1.5%.
• Designed a multi-task self-supervised learning framework for two-tower pre-rank model to enhance ads representation and finally achieved +0.004 in GAUC and +1.8% in GMV.

Search Ads in Alimama:

• Optimized the parameter-server based online learning algorithm xftrl and achieved +1.8% on CTR and +2.0% on rpm. Furthermore, the parameters in the linear model were innovatively used in the dnn model to speeded up the convergence and achieved another +1.2% on CTR and +1.5% on RPM.
• Developed the online deep learning system in algorithm aspect. The training procedure was upgraded from offline batch learning setting to online incremental setting and several aspects such as differentiated update policy and optimizer modification were concerned .
• First introduced pixel-level image information to the search ads CTR prediction on taobao app and increased CTR/RPM by 2%/1.8% respectively.

Tmall International:

• Developed trigger importance modeling in item retrieval which differentiated user historic behaviors according to their potential. By doing so, UVCTR increased by 1.7% and PVCTR increased by 2.3%.
• Used long term purchase history and short term view history in taobao app to build user interest and facilitate model performance in tmall international which built within taobao app, bringing 5.4% lift in UVCTR and 7% lift in PVCTR.

• Worked on display ads optimization on baidu mobile app including feature engineering, audience targeting, flow distributor, ctr prediction etc.

• Built the whole machine learning pipeline on advertising system from scratch including data collection, model training and model deployment.
• Conducted research on several CTR prediction models including LR, FM, FFM, GBDT+LR, FM+FTRL and deployed them successively which in total increased 20% on CTR and 15% on ecpm.
• Used PID(Proportion Integration Differentiation) controller in the real-time bidding environment to balance the ads bidding pace.