Haiyang Jia

Learning dynamic Bayesian network with immune evolutionary algorithm

Dynamic Bayesian networks (DBNs) are directed graphical models of stochastic processes. How to learn the structure of DBNs from data is a hot problem of research. In this paper, the author presents an immune evolutionary algorithm for learning the network structure of DBNs. The results of contrast experiment prove that the constringency rate is more rapid than EGA-DBN algorithms.

Cardinal direction relations in 3D space

The existing 3D direction models approximate spatial objects either as a point or as a minimal bounding block, which decrease the descriptive capability and precision. Considering the influence of objects shape, this paper extends the planar cardinal direction (CD) into 3D space and obtains a new model called TCD (three-dimensional cardinal direction). Base on the smallest cubic TCD relations and original relations, explain the correlations between basic TCD relations and block algebra. Then according to the results in block algebra, two novel ways to compute the inverse and composition of basic TCD relations are proposed. And an O(n4) algorithm to check the consistency of a set of basic TCD constraints over simple blocks is given.

Composing cardinal direction relations basing on interval algebra

Direction relations between extended spatial objects are important commonsense knowledge. Skiadopoulos proposed a formal model for representing direction relations between compound regions (the finite union of simple regions), known as SK-model. It perhaps is currently one of most cognitive plausible models for qualitative direction information, and has attracted interests from artificial intelligence and geographic information system. Originating from Allen first using composition table to process time interval constraints; composing has become the key technique in qualitative spatial reasoning to check the consistency. Due to the massive number of basic directions in SK-model, its composition becomes extraordinary complex. This paper proposed a novel algorithm for the composition. Basing the concepts of smallest rectangular directions and its original directions, it transforms the composition of basic cardinal direction relations into the composition of interval relations corresponding to Allens interval algebra. Comparing with existing methods, this algorithm has quite good dimensional extendibility, that is, it can be easily transferred to the tridimensional space with a few modifications.

Learning Markov equivalence classes of Bayesian Network with immune genetic algorithm

Bayesian Networks is a popular tool for representing uncertainty knowledge in artificial intelligence fields. Learning BNs from data is helpful to understand the casual relation between variables. But Learning BNs is a NP hard problem. This paper presents an immune genetic algorithm for learning Markov equivalence classes, which combining dependency analysis and search-scoring approach together. Experiments show that the immune operators can constrain the search space and improve the computational performance.

A hybrid approach for learning Markov equivalence classes of Bayesian network

Bayesian Networks is a popular tool for representing uncertainty knowledge in artificial intelligence fields. Learning BNs from data is helpful to understand the casual relation between the variable. But Learning BNs is a NP hard problem. This paper presents a novel hybrid algorithm for learning Markov Equivalence Classes, which combining dependency analysis and search-scoring approach together. The algorithm uses the constraint to perform a mapping from skeleton to MEC. Experiments show that the search space was constrained efficiently and the computational performance was improved.

Inversing Cardinal Direction Relations

Focusing on the inversing operation of cardinal directions, the current generative method does not always work correctly. According to the given definitions of smallest rectangular direction and original directions, the correlations between cardinal directions and interval algebra are built. Basing on the above analysis, the algorithm to compute the inverse direction and its proof are given. The results show that out of 511*511 possible pairs of basic cardinal directions, there are 757 pairs of inverse relations over simple regions and 1621 pairs over compound regions. Finally the discussion points out the reason why existing generative method works incorrectly and indicates that our method has a quite good dimensional expansibility.

An Effective Clustering Method Using a Discrete Particle Swarm Optimization Algorithm-Based Hybrid Approach

The purpose of this paper is to present and evaluate an improved Naive Bayes algorithm for clustering. Many researchers search for parameter values using EM algorithm. It is well-known that EM approach has a drawback - local optimal solution, so we propose a novel hybrid algorithm of the discrete particle swarm optimization (DPSO) and the EM approach to improve the global search performance. We evaluate this hybrid approach on 4 real-world data sets from UCI repository. In a number of experiments and comparisons, the hybrid DPSO+EM algorithm exhibits a more effective and outperforms the EM approach

Research on Distribution Alignment and Semantic Consistency in the Adversarial Domain Adaptation

Domain adaptation is an effective method solving the learning tasks lack of labeled data. In recent years, the adversarial domain adaptation (ADA) has achieved attractive results in a series transfer learning tasks. ADA reduces the distribution discrepancy between the source and the target by extracting the domain invariant features. However, the lack of constraints on the transferable features leads to poor results even negative transfers. A novel ADA method is proposed to solve this problem which contains two main improvements: the conditional distribution alignment and the semantic consistency regularization. The experiment demonstrate that the proposed method has promising improvement in the classification accuracy on the benchmark dataset. The code and data can be downloaded from https://github.com/kiradiso/EADA.

Causal Discovery with Bayesian Networks Inductive Transfer

Bayesian networks (BNs) is a dominate model for representing causal knowledge with uncertainty. Causal discovery with BNs requiring large amount of training data for learning BNs structure. When confronted with small sample scenario the learning task is a big challenge. Transfer learning motivated by the fact that people can intelligently apply knowledge learned previously to solve new problems faster or with better solutions, the paper defines a transferable conditional independence test formula which exploit the knowledge accumulated from data in auxiliary domains to facilitate learning task in the target domain, a BNs inductive transfer algorithm were proposed, which learning the Markov equivalence class of BNs. Empirical experiment was deployed, the results demonstrate the effectiveness of the inductive transfer.

Learning the structure of Dynamic Bayesian Network with domain knowledge

Dynamic Bayesian Network (DBN) is a graphical model for representing temporal stochastic processes. Learning the structure of DBN is a fundamental step for parameter learning, inference and application. For large scale problem, the structure learning is intractable. In some domains the training data is very limited and noisy, so learning the DBN structure only with training data is impractical. Domain knowledge may improve both the efficiency and the accuracy of the learning algorithm. But usually, the domain knowledge is uncertainty, unclear and even with conflict. This paper presents a novel algorithm for learning the structure of DBN, which consider the data and domain knowledge simultaneously, empirical experiment shows that the proposed algorithm improved the efficiency and the accuracy of the DBN structure learning.