While machine learning has achieved dramatic successes recently, a numerous researches have risen concerns against its deficiency such as lack of robustness and fairness. This tendency can be observed even for the cutting edge model architectures and training algorithms. Why does this unexpectancy happen? In this talk, we focus on the discrepancy between objective functions that learning algor thms optimize and evaluation criteria that ultimately define the goodness of learned models. By clearly distinguishing them, it enables us to verify whether the learning algorithms do achieve our desired properties and design suitable learning criteria. Specifically, I will introduce our recent work on adversarial robust classification and similarity learning.
In decision tree learning, we split an instance space into two subspaces based on a of instances at each node. For numerical or ordinal categorical features, this can be done by an "optimal" hyperplane that separates the domain space of those features. For nominal categorical features, however, it is not obvious to define a "hyperplane" of the domain space. Lucena (Lucena, AISTAT 2020) pointed out that the domain space of nominal categorical features may be structured and exploited this structural information to learn decision trees. In this method, at each internal node, we need to find an "optimal" bipartition of a graph whose vertices are labeled by either +1 or ?1 such that both sides are connected and the misclassification is minimized. In this paper, we formalize this problem as Connected Bipartition and investigate its computational complexity.
This report decribes a method for computing a strength of an argument on Weighted Bipolar Argumentation Framework. We propose a method on WBAF with support and attack relation between arguments, and extend it to the one with a set-support. In our method, first, the strength regarding support relations is computed, and the arguments connected by support relations are combined into a meta-argument. Then, the strength regarding attack relations is computed on a meta-WBAF. We describe the method and show its characteristics.
In this paper, we developed an envy-free, transparent and scalable mechanism for the high school course allocation problem. We experimentally compared true preference to two proposed scalable preference representations: the list and point representation. Results from this experiment suggested that the list representation is a scalable method in a real-life setting.
In this talk, we present our developments for applying the clustering algorithm that we call "data polishing". The applications are marriage support recommendation system, analysis for social media text data to detect the changes of behaviors of users, analysis for catalyst experimental data to understand the mechanism of the catalyst. Each application includes non-trivial difficulty that cannot be solved by direct applications of the algorithm. In this talk we show our contributions including the re-designs of the problems that were needed in the applications.
One way to measure the efficiency of enumeration algorithms is to evaluate it with respect to the input size and the number of solutions. Since the number of solutions can be typically exponentially larger than the size of the input, the running time of an enumeration algorithm is huge even if we manage to design an extremely efficient enumeration algorithm. However, in realworld problems, it is not always necessary to enumerate all the solutions and it may be required to enumerate sufficiently many "good" solutions. In such a situation, top-k enumeration algorithms are of great importance in many areas. On the theoretical side, Lawler (Lawler, Management Science 1972) developed a general framework for designing efficient top-k enumeration algorithms. The outline of Lawler's framework is a kind of best-first search. By using this framework, we can design a top-k enumeration algorithm that works in space linear with respect to k, where k is the number of solutions we wish to enumerate. However, since k may be exponentially larger than the input size n, the algorithm needs a huge amount of memory if k is large. In this paper, we modify Lawler's framework and obtain a space-efficient framework that runs in space depending only on a polynomial in n.
In this paper, we consider a situation in which an intelligent agent observes data for learning in environments. We also discuss a way to properly represent observed data in such a situation, and show that efficient data processing can be approximated by a neural net with a generalized convolutional structure.
Recently, many cars and road infrastructures have collected traffic data. Furthermore, traffic flow prediction using these data has been the focus of many studies. Traffic flow prediction is useful in avoiding traffic jams and suggesting an efficient route. However, large-scale traffic flow prediction takes much execution time. This paper proposes a method of partitioning a road network for distributed processing for large-scale traffic flow prediction. Our method consists two steps : (1) Partitioning the process of training models; (2) Selecting input data for each model. Our experimental evaluation shows that the method successfully reduces execution time. Too much input data does not improve prediction accuracy. Moreover, some input data is unrelated to distance between roads.
Designing de novo molecules with required property is an indispensable task in drug discovery and materials science. However, because of combinatorial problems in chemical diversity, it is still challenging to find optimized molecules. Here we propose a new approach to generate molecular graphs via decomposing existing molecules by frequent subgraph mining and combining subgraphs by Monte Carlo tree search (MCTS). Our experiments show that our method can discover better molecules in terms of the penalized log P and the drug-likeness compared to the state-of-the-art molecular graph generation methods.