C3.ai

Risks and Remedies for Black Box Artificial Intelligence

There is no denying that artificial intelligence (AI) and machine learning technologies will shape our future. Advanced AI systems can solve the most complex business problems and create significant economic benefits for organizations. But coupling such systems with poor interpretability - also referred to as explainability - can create a recipe for disaster. In this article, we outline the issues that can be addressed through implementing interpretable AI, the challenges to AI interpretability, and how the C3 AI Suite supports AI interpretability.

Quantiply

Dynamic Customer Segmentation

Common practice for a business to better understand their customer base is to categorize their customers into discrete peer groups based on categories such as their activits, income levels, spending behaviors, products being used, demographics, or geography. This is typically accomplished through Customer Segmentation, however simple implementations of this process can be riddle with flaws as this process tends to be static, data is treated as the same, and differences are considered walls. We developed an adaptive customer segmentation approach to make this process more dynamic and aggregate many different data sources. This solution allowed for better representations of customer dynamics without losing the power of interpretability or simplicity of customer segmentation.

AMLSim

The interest and exploration around AI solutions for financial crimes has gained much interest in the recent years. However, this hasn’t come as qucikly as other fields for which AI has become rapidly deployed. AI, specifically Machine Learning, is bounded by the data being used and within financial crimes the data is very sparse and hard to come across (for good reason as it is extremely private). Through expanding the work done by researchers at MIT and IBM, where they build an Anit-Money Laundering simulator, we enhanced our AI capabilities through incorporating much needed behavioral patterns to train our models for better accountability within financial crimes.

Explanable Graph Embeddings

Financial activities can be described through a graphical approach allowing for many state-of-the-art techniques to help extract meaningful information. We explored Graph Embeddings in our modeling efforts to improve the amount of information gained from networks. However, embeddings actually create more abstraction rather than explanation. Our work focused on algorithmic approaches to better explain these embeddings and provide better accountability for our models. Our work was submitted to AAAI-Make 2020

PhD Work

Distinguishing member and leader roles

One major distinction within communities is between leaders and members. My work is continuing on the already existing literature of online social roles through examining social roles within enterprise communities. Through a mixed method approach of analyzing content and behaviors of self defined roles, I have been able to develop models in order to distinguish such roles. Work here is to be presented at CSCW 2018: Living in the Present: Understanding Long-Term Content Referencing in Enterprise Online Communities

Ecology of Social Roles

Social roles typically are defined in relational terms, with this in mind it is needed to understand the social influence that roles have on each other. This work is expanding upon simple role differences by examining a community as an ecology through social network analysis and graphical modeling. I recently expanded this work to examine substructures called Graphlets that exist within the larger community network. I believe such substructures can measure the collaborative nature of small-groups.

Success of Social Role Behaviors

To understand the effect of social roles on a community level, we measure the community quality, which is typically referred to as online community success. There are varying degrees to which online community success has been defined but I have chosen to primarily use an empirically agreed upon surveyed measure of member satisfaction. Much of my work has been on understanding how role behaviors either have an improving or declining effect on member satisfaction. Some of this work was presented at ICWSM 17: ‘Just the Facts’: Exploring the Relationship between Emotional Language and Member Satisfaction in Enterprise Online Communities.

Role dynamics over time

There is has been much literature on the changes of roles and communities over time, however little quantitative evidence has been provided. My final goal of my thesis is to understand social role changes over time and how member-leader responsibilities are fluid through the life-cycle of a community. I’ve taken this approach both linguistically and graphically.

Older Projects

2018

Exploring the topical relationship of user recorded events in a psychology study on Digital Memory, Reflection and Behavior Change. I used topic modeling to assist in classifying the type of events users reported in the study and found significant differences in both what events user reported and the type of language they used in descriptions of the events related to pre and post study measures of their well-being.

2015

Classifying argument style within political debate forums. This research is a shift from examining cooperative communities, to observing conflict that arises within online communities. Previous research on argument style has been focused on traditional rhetoric forms. However, from observations made within online debate forums it is seen that typically, people are arguing with an emotional appeal instead of structural. My research has developed a fair classifier to identify a binary distinction between fact or emotional based arguments. Further observations have found that arguments are not restricted to only the use of one style and future work is to include a continuous scale between the two styles. This work was incoporated in my thesis work through examining the impact of different language types in an Enterprise Community context.

2014

Analytics of large scale online communities to predict online individual and social behavior. Various descriptive lifecycle models characterize community development, but there has been little quantitative exploration of how communities change over time, specifically how they organize and structure extensive long-term content. I’ve explored whether content is organized, who organizes it, and which social media tools they use. Technical implications include the need for more dedicated support for curation, in particular to better exploit linking tools and to encourage members to take more responsibility for organization. The ultimate goal is to identify successful individual and community practices and metrics to increase online community health. This work has been incorporated into my work distinguishing social roles.