Machine Learning (algorithm)
Probabilistic Graphic Model and Casual Inference: Research and Applications
Bayesian networks (BN), the most well-known probabilistic graphical model, have been widely used to model causal relations between random variables in ML, AI and DM fields. With the development of deep learning technology, casual models attract huge attention once again due to their advantages in building ``explainable’’ models to break through the weakness of deep models. By specifying the conditional probability distribution of each node in BN given its parents, BN provides a light-weight and explicit factorization of the joint probability distribution.
In the real world, the casual relations are often complex and not directly visible. Randomized controlled trials, which are in general considered the experimental “gold standard” for uncovering causal relationships, are often very costly or even infeasible. Structure learning, which aims at learning suitable structure of the causality networks purely from observation data, has therefore been a hot research topic in the recent time. We have seen many promising applications of structure learning, such as gene expression data analysis, error identification, model factorization and system optimization, to name but a few.
Nowadays structure learning methods include combinatorial optimization algorithms and continuous optimization algorithms. Combinatorial optimization algorithms are difficult to attain high result quality and high time efficiency at the same time. Continuous optimization algorithms are also very slow. Meanwhile, both of them can scale on at most thousands of nodes, which is far from enough in real-world applications. Therefore, in this project, we want to explore feasible structure learning algorithms for large-scale Bayesian networks with up to millions of nodes. Taking this tool in hands, we could further unlock the possibility to apply Bayesian networks in some promising applications, including but not limited to AI for reasoning, AI for system and AI for recommendation.
- A well-designed large-scale structure learning algorithm which is accurate, fast and scalable.
- An open-source structure learning system with an optimized implementation of the algorithm.
- Applying structure learning to solve some long-standing problems in new perspectives, such as DB system optimization.
- Applying our method to some business applications, such as error identification and explainable recommendation system.
- Publish 1-2 top-tier conferences and apply for 1-2 patents.
Related Research Topics
- Large-scale structure learning algorithm and system design
- The integration of Bayesian networks and deep models
- Using structure learning to optimize DB systems and building recommendation systems
- Incremental and distributed structure learning algorithms
- Rule learning and logical inference besides Bayesian networks
Suggested Collaboration Method
AIR (Alibaba Innovative Research), one-year collaboration project.