Hyung-il Ahn

Affective-Cognitive learning and decision making: a motivational reward framework for affective agents

In this paper we present a new computational framework of affective-cognitive learning and decision making for affective agents, inspired by human learning and recent neuroscience and psychology. In the proposed framework ‘internal reward from cognition and emotion’ and ‘external reward from the external world’ serve as motivation in learning and decision making. We construct this model, integrating affect and cognition, with the aim of enabling machines to make smarter and more human-like decisions for better human-machine interactions.

A CRM system for social media: challenges and experiences

The social Customer Relationship Management (CRM) landscape is attracting significant attention from customers and enterprises alike as a sustainable channel for tracking, managing and improving customer relations. Enterprises are taking a hard look at this open, unmediated platform because the community effect generated on this channel can have a telling effect on their brand image, potential market opportunity and customer loyalty. In this work we present our experiences in building a system that mines conversations on social platforms to identify and prioritize those posts and messages that are relevant to enterprises. The system presented in this work aims to empower an agent or a representative in an enterprise to monitor, track and respond to customer communication while also encouraging community participation.

Mixture of gaussian processes for combining multiple modalities

This paper describes a unified approach, based on Gaussian Processes, for achieving sensor fusion under the problematic conditions of missing channels and noisy labels. Under the proposed approach, Gaussian Processes generate separate class labels corresponding to each individual modality. The final classification is based upon a hidden random variable, which probabilistically combines the sensors. Given both labeled and test data, the inference on unknown variables, parameters and class labels for the test data is performed using the variational bound and Expectation Propagation. We apply this method to the challenge of classifying a students interest level using observations from the face and postures, together with information from the task the students are performing. Classification with the proposed new approach achieves accuracy of over 83%, significantly outperforming the classification using individual modalities and other common classifier combination schemes.

Effects of guided and unguided style learning on user attention in a virtual environment

In this paper, we investigated the effects of guided and unguided style VR learning on user attention and retained knowledge. We conducted a study where users performed guided or unguided style learning in the virtual environment while user attention was measured through an eye tracking system and physiological sensors. The virtual environment contained the five specific events associated with different stimuli, but the guided task was designed to provide the specific goals whereas the unguided task asked the user to actively search for the interesting items. The results showed that the unguided task followed by the guided task made a considerable learning effect by giving a preview to the user. In addition, tactile feedback, sudden view point change, unique appearance and behavior, and sound stimuli played an important factor in increasing human attention states that also induced enhancing human memory about VR experience.

Measuring Affective-Cognitive Experience and Predicting Market Success

We present a new affective-behavioral-cognitive (ABC) framework to measure the usual cognitive self-report information and behavioral information, together with affective information while a customer makes repeated selections in a random-outcome two-option decision task to obtain their preferred product. The affective information consists of human-labeled facial expression valence taken from two contexts: one where the facial valence is associated with affective wanting, and the other with affective liking. The new “affective wanting” measure is made by setting up a condition where the person shows desire to receive one of two products, and we measure if the face looks satisfied or disappointed when each of the products arrives. The “affective liking” measure captures facial expressions after sampling a product. The ABC framework is tested in a real-world beverage taste experiment, comparing two similar products that actually went to market, where we know the market outcomes. We find that the affective measure provides significant improvement over the cognitive measure, increasing the discriminability between the two similar products, making it easier to tell which is most preferred using a small number of people. We also find that the new facial valence “affective wanting” measure provides a significant boost in discrimination and accuracy.

Switching vector autoregressive models with higher-order regime dynamics Application to prognostics and health management

Regime switching vector autoregressive (RSVAR) models are typically used to model changing dependency structures of multivariate time series. These changing regimes are represented by using a first-order Markov process where the transition distribution reflects the probabilities of moving to one of the other regime in the subsequent time step. Instead of representing the state of the system at different points in time, we extend this framework by using an explicit time representation that allows us to query against probability distributions of when particular regime changes take place. In contrast to continuous time based approaches such as continuous time Bayesian networks or continuous time Markov processes, we do not rely on intensity matrices that describe trajectories of consecutive states. Here we define regime changes as events and understand time as context of an event. This allows us to integrate dependencies at different time granularities while being able to perform inference in a decomposed way. As a consequence, we can efficiently consider higher-order effects stretching across a large number of consecutive regimes. The underlying assumption is that timely evolution of variables between regime switches is completely captured by the VAR model or possibly a set of VAR models with varying measuring rates and that there is a representative set of multiple time series exhibiting similar higher-order regime dynamics. In this paper we show how such dynamics can be learned integrative with learning RSVAR model parameters and how the regime dynamics can be considered in the RSVAR inference procedures. We demonstrate the benefits of our approach based on a simple scenario. Further, an application to a typical prognostics scenario is presented, leading to the highest score in the IEEE PHM 2014 Data Challenge for the industrial track.

Survival analysis for HDLSS data with time dependent variables: Lessons from predictive maintenance at a mining service provider

In gene expression analysis it is often the goal to predict survival given a high-dimensional space of covariates. In corresponding literature models are described that deal with low sample size which is a typical feature of such studies. This is also the case in asset management services where downtime of assets is very costly and thereby replacements are scheduled long before the actual risk of failure increases. Although sometimes good surrogates of the true failure probability are available, it is in practice often the case that a number of weak predictors exist which needed to be filtered from a large set of candidates. Although the challenge is similar to gene expression analysis, a crucial difference is that covariates in condition monitoring are dynamic whereas genes are not. The result is that in gene expression analysis any data in between failure can be omitted, which leads to a potentially high bias in variable selection for condition monitoring. The authors are not aware of any survival models that deal with high dimensional low sample size (HDLSS) data in case of time-dependent covariates. In this paper we evaluate the performance of different modeling techniques in case of HDLSS survival data including the definition of a discrete time model where survival is modeled as a locally independent, binary outcome variable. We thereby study the trade-off between omitting measurements between times of failure and disregarding temporal dependencies. The analysis is based on a real life case study where 39 components of 50 mining haul trucks were monitored in operations over almost 6 years.

Sales Prediction with Social Media Analysis

Social media has been valuable sources to predict the future outcomes of some events such as box-office movie revenues or political elections. This paper focuses on periodic forecasting problems of product sales based on social media analysis and time-series analysis. In particular, we present a predictive model of monthly automobile sales using sentiment and topical keyword frequencies related to the target brand over time on social media. Our predictive model illustrates how different time scale-based predictors derived from sentiment and topical keyword frequencies can improve the prediction of the future sales.

TEM: a novel perspective to modeling content onmicroblogs

In recent times, microblogging sites like Facebook and Twitter have gained a lot of popularity. Millions of users world wide have been using these sites to post content that interests them and also to voice their opinions on several current events. In this paper, we present a novel non-parametric probabilistic model - Temporally driven Theme Event Model (TEM) for analyzing the content on microblogs. We also describe an online inference procedure for this model that enables its usage on large scale data. Experimentation carried out on real world data extracted from Facebook and Twitter demonstrates the efficacy of the proposed approach.

Action planning with commonsense knowledge

Understanding other peoples goals is an essential part of interpersonal interactions. This capability enables a person to naturally predict another persons future actions in a situation and produce appropriate joint or shared actions. In like manner, a human-like planning agent (or sociable robot) should be able to understand the users action goal and come up with subgoal-based plans to achieve the goal. In this paper we focus on how the agent can automatically construct the subgoal-based action hierarchy corresponding to the users high-level goal. As a first step, we implement an action-planning engine based on ConceptNet, and indicate the drawbacks of using ConceptNet for this purpose. Also, we present the structure of a new goal-oriented commonsense-reasoning knowledgebase for the agents action-goal representation and action planning.