I am a soon-to-be graduate of the University of California: San Diego.
I am studying Cognitive Science
and strive to contribute to the development of human-centered, data-driven
technologies. I'm most inspired at the intersection of mathematical
formalism and human behavior, especially surrounding language and music.
Lastly, I'm keenly aware of the importance of ethical machine learning
and data science, and uphold myself to high standards here. I also
enjoy running, reading, backpacking, and gardening.
The best way to reach me is at firstname.lastname@example.org where x = alexliebscher0
I was a data science intern for TextRecruit during the summer of 2018. I discovered usage and performance insights from the company's large collections of customer and user data.
I focused my time on the 40+
million messages the company
had transmitted through their platform. I built a sentiment analysis model which improved F1 scores over 1.7x compared to other out-of-the-box solutions. I wished to model the flow of text message conversations with the company's
artificial intelligence chatbot to
determine retention and confusion rates, along with drop off points. I engineered two classification models to tag inbound user text messages and outbound chatbot messages as one of 30 or 34 classes, respectively. These were applied to
conversations to generalize the structure and present insights on weaknesses in the platform's NLP capabilities. I also provided KPIs about chatbot usage, and customer activity and growth, which helped company leadership validate business
Maintained and upkept the yards and gardens of large residential
and commercial properties. Developed a keen sense of detail, an indefatigable
work ethic, and the character to place myself in uncomfortable situations
and drive myself toward the company’s idea of success.
La Jolla, CA | September 2016 - June 2020
Courses of Interest
Computer Science and Math
Supervised Machine Learning
Race, Gender, and Artificial Intelligence
Unsupervised Machine Learning
Mathematical Statistics I & II
Public Rhetoric & Practical Communication
Computational Models of Cognition
Vector & Multivariable Calculus
Data Structures & Object-Oriented Design
Cognitive Consequences of Technology
Intro to CompSci: Java I & II
Neuroanatomy and Physiology
Sensation and Perception
Research Methods and Statistical Analysis
Effects of Battle and Journey Metaphors on Charitable Donations for Cancer Patients
Fall 2018 - Present
Patients with cancer often describe their experience metaphorically as a battle
(“my fight against cancer”) or as a journey (“my path through cancer treatment”).
Experimental work has demonstrated that these metaphors can influence people's
reasoning and emotional inferences about experiences with cancer (Hendricks, Demjen, Semino, & Boroditsky, 2018; Hauser & Schwarz, 2019).
However, it is currently unknown how the use of these metaphorical frames translate
into behavioral changes, such as the likelihood and magnitude of charitable giving.
Using hand-labeled data from more than 5,000 GoFundMe cancer-related campaigns in a
regression framework, we asked whether or not a campaign’s usage of metaphor predicts
several measures of donation behavior beyond what other control variables predict
(e.g. shares on Facebook). We found that both metaphor families (battle or journey)
have a positive effect on campaign success and donation behavior.
To establish whether these relationships are causally meaningful, we designed an
online experiment simulating the experience of donating to a crowdfunding campaign.
We manipulate the metaphorical framing and recipient gender in the campaign. We will
be measuring real donations to the campaigns from participants to determine if an
effect of metaphor exists on charitable donations.
Tools: Python (pandas), R (lme4, ggplot, pwr), Github & git
As an avid music fan, I love discussing music and bonding with people
over a mutual interest in a band or genre. However, it's difficult to
keep track of all the music I've listened to and enjoyed (and not enjoyed).
For this reason, I designed and developed
a locally stored and run browser application for recording, logging, and
rating my music collection. Through trial and error, I taught myself
the React-Redux ecosystem for the front-end, constructed a REST API in Flask
on the back-end, and hold all my music data on a local MongoDB instance. I've
incorporated a system for comparing and ranking albums, so we may definitevly
answer questions like, "What're your top three albums?" A recommendation
heuristic is also built-in to suggest which album I should listen to next. Lastly,
I've made it open source and maintain it on GitHub.
The introduction of contextual word embedding (CWE) models has led
to improvements on a wide variety of tasks. Yet, the black-box
nature of deep learning language models may be inhibiting further
progress. Tenney et al (2019) introduced a novel edge probing framework
to explore the syntactic and semantic information encoded within
contextual embeddings. They assessed the degree to which these types
of information are encoded in the embeddings through a series of
traditional linguistic tasks. Here,
I expand this framework and study how nonliteral meaning may be
also encoded within these embeddings. Nonliteral meaning
is often highly abstract, conceptual, and cultural.
I find that contextual embeddings
do encode some level of nonliteral meaning, as distinguished by
our probing of metaphor and metonymy detection tasks.
For any classification problem, choosing a proper classifier and its parameters is critical for success. This paper is my attempt at pushing for methodical supervised machine learning. I
evaluate the performance of seven classification
algorithms across four data sets. For thoroughness, each classifier was tested over three independent trials, where each trial was subject to three partitions of the data. For each partition of the data, cross validation was
For each CV fold, an optimal set of hyper-parameters was found for the classifier using Bayesian Search. Contrary to traditional grid search, this method improves performance, takes advantage of the underlying parameter space with specific
priors, and reduces redundant and insignificant searches. Performance overall proved that Random Forests, Gradient Boosted Trees, and RBF-SVM achieve the highest results. K-Nearest Neighbors may also be a viable solution but should be
treated with care and precision.
For this course project, I was interested in building a model with the ability to "fill in the blanks" when presented with a statement such as "Her hair is as red as ____".
For this, I created my own dataset
and annotated certain
components of each simile. These similes were embedded in a vector space with Word2Vec. I then went through a model selection process in which I tried to minimize the model complexity and data complexity, while maximizing the prediction
output entropy. I reduced the word embedding space (for better generalization), tuned the number of components of the Gaussian Mixture Model (representing the number of semantic topic groups), decreased the probability of the model
randomness, and maximized the entropy of the bin counts of predicted GMM components (to prevent most similes from being clustered under one component). Given partial data (an incomplete simile), the model permutates over all possible
combinations of latent components and vocabulary, searching for the maximum log probability solution. This solution is the "blank" in which to fill the simile. Some results are intriguing, creative, and plausible, whereas others are (fun
I love talking about music: learning about new artists and songs, discussing albums, and debating the lyrics of music. I might argue that my taste in music defines my friend-circle to some degree. When I meet
new people, often times one
of the first questions I'll ask is, "So what kind of music do you listen to?" or "Who are your top three artists right now?" Music can tell you a lot about a person, and I take this to heart. Hence, this project is an attempt to
me a little better.