A knowledge gradient policy with provable bounds for sequential decision making with stochastic binary feedbacks.

Repeated Bayesian learning sessions will inevitably collapse to priors unless reinforced with logarithmic-size sessions.

Clean small data is better than noisy big data for product representation learning via multimodal deep models.

Whole-page optimization via ordered combinatorial semi-bandits and network flow is applied to news recommendation.

Billion-scale product search via GPU-based approximate KNN in product vector space constructed via attention auto-encoder.

Scalable and robust numerical methods let you know how wave function evolves and propagates in a fractal universe.

Large-scale simulations on diblock-polymer self-organize into complex ordered metastable structures.

Pack-size and multi-pack info are inferred from product metadata to ensure consistent unit price for different products.

A Frank-Wolfe type regularized boosting algorithm with provable generalization bounds via Rademaker complexity.

The dynamics of stable phase eating metastable phase are simulated as interface moving during phase transition.

The similarity between nodes are modeled as Markov transition probability. This similarity is proved to be a quasi-metric.

Ordered phases prefer certain orientations when growing from metastable structures. There could be multiple preferred ones.