From Alibaba’s Ad Algorithms to Deep Learning: A Senior AI Engineer’s Journey

Jing Shi, an Alibaba senior algorithm expert, introduces himself as a Tsinghua graduate in machine learning and AI who has worked on advertising algorithms, including matching, estimation, ranking, and related AI technologies such as computer vision and NLP.

He outlines three parts of his talk: his career path from joining Alibaba, insights behind AI work and achievements, and interview preparation advice.

Why Alibaba?

During his student years he sought a place where research and application intersected. Discovering Alibaba’s large‑scale machine‑learning projects on a forum, he applied and was invited for an interview, eventually receiving an offer.

He chose Alibaba because of its booming retail business, rich consumer data covering the whole purchase journey, and a culture that encourages both research and practical applications.

Six Years at Alibaba: From a Slow Start to Doubling Impact

Initially he felt unprepared, spending months immersing in business data and learning to apply machine‑learning knowledge to real problems. He notes that large‑scale click‑through‑rate (CTR) estimation became a pivotal area for the industry, with massive feature sets and sample sizes.

CTR models used billions of ID features represented as sparse one‑hot vectors, leading to challenges of huge sample volume and high dimensionality. Regularization (L1) was essential to prevent over‑fitting, and feature selection was needed for efficient online prediction.

He questioned whether linear models were sufficient for such high‑dimensional data, observing that many believed non‑linear models would always outperform linear ones. However, research showed that with extremely high‑dimensional sparse features, linear models can be adequate.

To reduce manual feature engineering, he explored piecewise linear models: dividing the high‑dimensional space into regions and training an independent linear model in each region, effectively creating a segmented linear function that can approximate complex relationships.

Challenges included jointly learning the region partitions and handling massive data and feature scales. He developed a hybrid logistic regression with softmax‑based region partitioning, experimenting with dozens of variants.

The resulting system improved ad revenue and CTR by over 10% per phase, impacting billions of users daily.

Deep Learning: Current State and Insights

He describes deep learning as decoupling models from algorithms, standardizing optimization (e.g., SGD with momentum) and allowing rapid model experimentation. This modularity enables engineers to build complex models without deep algorithmic expertise.

Deep learning’s component‑based nature (e.g., CNNs, RNNs, LSTMs) lets practitioners assemble architectures suited to specific data characteristics, such as structured e‑commerce user behavior or sequential browsing patterns.

Alibaba leverages these ideas to design models that handle massive ID‑based features and heterogeneous data, publishing work like a multi‑interest user model on arXiv.

Three Stages of Technical Progress at Alibaba

1. Application: adopting state‑of‑the‑art methods to improve business outcomes.

2. Innovation: creating novel techniques and publishing top‑tier research.

3. Transformation: using technology to reshape entire business causality chains and spawn new business models.

What Kind of Technologists Does Alibaba Seek?

Alibaba looks for “extraordinary people with a common mindset” who are smart, resilient, optimistic, and self‑reflective. They should possess strong fundamentals, the ability to learn quickly, and the perseverance to tackle difficult problems.

He cites the example of Dr. Wang Jian’s dedication to Alibaba Cloud as a model of resilience and conviction.

Interview Tips for Candidates

Highlight achievements such as high‑impact projects, publications, competition rankings, or open‑source contributions. Refresh core programming skills, data structures, and algorithms (e.g., review “Introduction to Algorithms”). During interviews, proactively showcase strengths and guide the conversation toward your most compelling experiences.

Overall, his talk blends personal career reflections, technical deep‑dives into large‑scale machine learning and deep learning, and practical guidance for aspiring AI engineers.