Award
  • Winners of 2015 Taiwanese Association for Artificial Intelligence (TAAI) Master Thesis and Doctoral Dissertation Award (Download from here).
    • Award  Master Thesis Title / Doctoral Dissertation Title  Student Name  Advisor Name Affiliation / Country
    Doctoral Dissertation Inferring Missing Sensor Values and Recommending Sensor Deployment Locations Based on Urban Big Data: A Case Study on Air Quality Hsun-Ping Hsieh Shou-De Lin Graduate Institute of Networking and Multimedia, Department of Computer Science and Information Engineering College of Electrical Engineering and Computer Science
    National Taiwan University, Taiwan
    Doctoral Dissertation A Study of Efficient Algorithms for Mining High Utility Patterns in Transactional and Event Databases Cheng-Wei Wu Vincent S. M. Tseng Department of Computer Science and Information Engineering
    National Cheng Kung University, Taiwan
    Master Thesis Warm Start for Parameter Selection of Linear Classifier Bo-Yu Chu Chih-Jen Lin Department of Computer Science and Information Engineering College of Electrical Engineering and Computer Science
    National Taiwan University, Taiwan
    Master Thesis Deep Semi-Supervised Learning for Domain Adaptation Hung-Yu Chen Jen-Tzung Chien Institute of Communications Engineering College of Electrical and Computer Engineering
    National Chiao Tung University, Taiwan
    Master Thesis Improving Ranking Consistency for Web Search by Leveraging a Knowledge Base and Search Logs Jyun-Yu Jiang Pu-Jen Chen Department of Computer Science and Information Engineering College of Electrical Engineering and Computer Science
    National Taiwan University, Taiwan
Award Paper Title and Authors First Page of Paper
Best Paper Award
A Practical Divide-and-Conquer Approach for Preference-Based Learning to Rank 
Han-Jay Yang and Hsuan-Tien Lin

Abstract—Preference-based learning to rank (LTR) is a model that learns the underlying pairwise preference with soft binary classification, and then ranks test instances based on pairwise preference predictions. The model can be viewed as an alternative to the popular score-based LTR model, which learns a scoring function and ranks test instances based on their scores directly. Many existing works on preference-based LTR address the step of ranking test instances as the problem of weighted minimum feed- back arcset on tournament graph. The problem is somehow NP- hard to solve and existing algorithms cannot efficiently produce a decent solution. We propose a practical algorithm to speed up the ranking step while maintaining ranking accuracy. The algorithm employs a divide-and-conquer strategy that mimics merge-sort, and its time complexity is relatively low when compared to other preference-based LTR algorithms. Empirical results demonstrate that the accuracy of the proposed algorithm is competitive to state-of-the-art score-based LTR algorithms. 
 PDF
Merit Paper Awards
Mining Closed+ High Utility Itemsets without Candidate Generation 
Cheng-Wei Wu, Philippe Fournier-Viger, Jia-Yuan Gu and Vincent S. Tseng

Abstract—High utility itemsets (HUIs) mining refers to discovering sets of items that not only co-occur but also carry high utilities (e.g., high profits). HUI mining receives extensive attentions in recent years due to the wide applications in various domains like commerce and biomedicine. However, huge number ofHUIs might be produced to users, which degrades the efficiency of the mining process. A promising solution to this problem is to mine closed+ high utility itemset (CHUI), a compact and lossless representation of HUIs. Nevertheless, existing algorithms incur the problem of producing a large amount of candidates, which degrades the mining performance in terms of time and space. In this paper, a novel algorithm named CHUI- Miner (Closed+ High Utility Itemset mining without candidates) for mining CHUIs is proposed, which directly computes the utility of itemsets without producing candidates. To our best knowledge, this is the first work addressing the issue of mining CHUIs without candidate generation. Experimental results show that CHUI-Miner is several orders of magnitude faster than the state-of-the-art algorithms. 
 PDF
Merit Paper Awards
Variants of Heuristic Rule Generation from Multiple Patterns in Michigan-style Fuzzy Genetics-based Machine Learning 
Yusuke Nojima, Kazuhiro Watanabe and Hisao Ishibuchi

Abstract—In the design of rule-based classifiers, a single rule is often generated from a single pattern in a heuristic manner. Since the generated rule is likely to be over-specialized to the pattern, its conditions are often randomly replaced with don’t care. However, the generalized rule with don’t care conditions does not always have high classification ability. This is because the replacement is randomly performed without utilizing any information about other patterns. In our previous studies, we proposed an idea of generating a fuzzy classification rule from multiple patterns. In this paper, we propose its six variants. Each variant has a different criterion for choosing multiple patterns from which a single rule is generated. The proposed variants are used to generate fuzzy classification rules in Michigan-style fuzzy genetics-based machine learning. The usefulness of each variant is evaluated as a heuristic fuzzy rule generation method through computational experiments on 20 benchmark data sets. 
 PDF
Merit Paper Awards
Multiobjective Orienteering Problem with Time Windows: An Ant Colony Optimization Algorithm 
Yu-Han Chen, Wei-Ju Sun and Tsung-Che Chiang

Abstract—The orienteering problem with time windows(OPTW) deals with the problem about selecting a set of points of  interest and then determining the route to visit them under the time window constraints. In the classical OPTW each candidate point of interest is associated with a profit value, and the objective is to maximize the total profit. In this study, we extend the problem and allow each point to have multiple profit values, which could reflect different aspects of consideration. We propose an ant colony optimization (ACO) algorithm to solve the multi objective OPTW (MOOPTW) with the goal of finding the set of Pareto optimal solutions. To our best knowledge, this is the first study to address the MOOPTW with comprehensive numerical experiments. Our algorithm is a decomposition-based one, which decomposes the multiobjective optimization problem into single-objective sub-problems. Pheromone matrices are associated with sub-problems. We also incorporate path-relinking and propose several strategies. We apply our algorithm to solve 76 public benchmark instances and offer the list of non-dominated solutions to facilitate performance comparison in future researches.
 PDF

Award Paper Title and Authors First Page of Paper
Best Paper Award
PDF
Merit Paper Awards

PDF
Merit Paper Awards
PDF
Merit Paper Awards
PDF