DSAI: Unbiased and Interpretable Latent Feature Extraction for Data-Centric AI
Hyowon Cho, Soonwon Ka, Daechul Park, Jaewook Kang, Minjoon Seo, Bokyung Son · 2024
Read DSAI: Unbiased and Interpretable Latent Feature Extraction for Data-Centric AI by Hyowon Cho, Soonwon Ka, Daechul Park, Jaewook Kang, Minjoon Seo, Bokyung Son (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06303.txt.
Introduction
Hyowon Cho and five co-authors build a data-first AI that pulls latent features from datasets with no lean on past knowledge. Large models tend to name features from what they were taught, not from what the data show. That lean can warp the features they claim to find. Their 2024 framework, DSAI, works in stages. It scores each feature it finds with a clear prominence measure.
On synthetic sets where the true features are known, the framework found most expert features and stayed true to the data. On real sets it surfaced sound patterns with little expert help. One use is classification that a person can read and check. In a neat twist, the title of the study itself was picked from a list by the rules of the framework. The team pitches it as a cure for knowledge-led readings.
Catalog record
| CREATOR | Hyowon Cho, Soonwon Ka, Daechul Park, Jaewook Kang, Minjoon Seo, Bokyung Son |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 76 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-06303
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DSAI: Unbiased and Interpretable Latent Feature Extraction for Data-Centric AI Hyowon Cho, Soonwon Ka, Daechul Park, Jaewook Kang, Minjoon Seo, Bokyung Son Abstract Large language models (LLMs) often struggle to objectively identify latent characteristics in large datasets due to their reliance on pre-trained knowledge rather than actual data patterns. To address this data grounding issue, we propose Data Scientist AI (DSAI), a framework that enables unbiased and interpretable feature extraction through a multi-stage pipeline with quantifiable prominence metrics for evaluating extracted features. On synthetic datasets with known ground-truth features, DSAI demonstrates high recall in identifying expert-defined features while faithfully reflecting the underlying data. Applications on real-world datasets illustrate the framework’s practical utility in uncovering meaningful patterns with minimal expert oversight, supporting use cases such as interpretable classification 1 1 1 The title of our paper is chosen from multiple candidates based on DSAI-generated criteria. . 1 Introduction The ability to analyze large-scale datasets is a cornerstone of deriving actionable business insights. Traditionally, this task has been managed by human data scientists , but it faces several key challenges: (1) the large volume of data makes it difficult to review all information comprehensively, (2) human analysis can often include subjective bias, and (3) collaboration with domain experts is often required, leading to high operational costs. Large language models (LLMs) have emerged as powerful tools for identifying patterns within massive datasets, leveraging their ability to process and generate language in context Touvron et al. (2023) ; Dubey et al. (2024) ; Achiam et al. (2023) ; OpenAI (2024) ; Lam et al. (2024) ; Wan et al. (2024) . However, their application to data analysis is limited by critical shortcomings. First, LLMs often struggle to identify latent characteristic patterns in large datasets due to inherent data grounding issues , where outputs rely on pre-trained knowledge rather than the specific nuances of the input data Kossen et al. (2024) ; Kenthapadi et al. (2024) ; Wu et al. (2024) . Second, the difficulty in verifying LLM-generated responses and the lack of quantitative evaluation methods require expert oversight, which can be prohibitively expensive at scale. To address these limitations, we propose Data Scientist AI (DSAI) , a framework that systematically applies LLMs to extract and refine latent features from data. Unlike direct feature extraction approaches, DSAI adopts a bottom-up approach, starting with detailed analysis of individual data points, aggregating their characteristics, and deriving actionable features. The process is guided by defined perspectives which provide LLMs with a consistent framework for interpreting data points while minimizing subjective bias. The DSAI pipeline operates in five stages: #1 Perspective Generation identifies data-driven perspectives from a small subset of data. #2 Perspective-Value Matching assigns values to individual data points by evaluating them against these perspectives. #3 Clustering groups values with shared characteristics to reduce redundancy. #4 Verbalization converts extracted features into a compact criterion form. #5 Prominence-based Selection determines which features to use based on a prominence intensity metric that quantifies the discriminative power of each extracted feature. Throughout these stages, the LLM remains task-agnostic – we do not reveal the specific domain or the “correct answer” context during feature generation. This design minimizes bias and ensures that the identified features are grounded in the data rather than the model’s background knowledge. We validate our framework on datasets curated for our experiments, including a research titles dataset Wang et al. (2018) and an advertising slogans dataset Jin et al. (2023) , each with expert-defined criteria that serve as ground truth for evaluation. We then demonstrate DSAI’s practical value on three real-world datasets: news headlines with click-through rate (CTR) labels Wu et al. (2020) , a spam detection dataset Kim (2016) , and Reddit comments with community engagement metrics Magnan (2019) . Our main contributions are as follows: • Minimizing Bias : We ensure that LLMs focus on latent characteristics present in the data, thus mitigating the tendency to rely on domain-specific prior knowledge (addressing the data grounding issue). • Prominence Metric : We introduce a quantitative metric for feature prominence, which serves as a proxy for the discriminative power of each extracted feature. • Interpretability : We improve interpretability through feature-to-source traceability, allowing users to trace each extracted feature back to the data points that support it. • Efficiency : We enable thorough examination of large datasets with minimal human labor by systematically guiding LLMs through the analysis process. 2 Related Works 2.1 Latent Feature Extraction with LLM Recent advancements in LLMs have demonstrated their effectiveness in extracting latent features, particularly in identifying perspectives and matching values in data Peng et al. (2023) . Studies using LLM-based clustering techniques have shown promising results in extracting high-level concepts Lam et al. (2024) ; Wan et al. (2024) ; Viswanathan et al. (2024) ; Pham et al. (2024) , demonstrating their utility for analyzing large datasets Wang et al. (2023) ; Kwon et al. (2023) . Research has also shown that decomposing complex tasks into multiple stages or aspects enhances performance Saha et al. (2023) ; Liu et al. (2024) . Our framework combines perspective generation, multi-stage feature construction, and clustering using LLMs to conduct comprehensive latent feature extraction. 2.2 Bias of LLMs LLMs face challenges in adapting to new patterns due to pre-existing knowledge biases. Kossen et al. (2024) show that in-context learning (ICL) struggles to overcome these biases even with explicit prompts or many-shot examples. Using external knowledge bases also has limited effectiveness in reducing hallucinations and reliance on internal biases Kenthapadi et al. (2024) ; Lee et al. (2023) . Advanced LLMs fail to align with provided context in about 40% of predictions when the given context conflicts with their prior knowledge Wu et al. (2024) . These findings emphasize the need for robust techniques to mitigate bias and enhance grounding, especially for applications where data-driven conclusions are crucial. 3 Challenges in LLM-Driven Data Analysis In this section, we examine the behavior of a state-of-the-art LLM when tasked with directly generating features from data. These exploratory experiments reveal the data-grounding challenges that motivate our DSAI approach. 3.1 Setting Dataset Annotation and Sampling We use two expert-annotated text datasets where the “ground truth” latent features are known (defined by domain experts). The first dataset consists of research paper titles and the second contains advertising slogans. In each dataset, experts Nair and Gibbert (2016) ; Tullu (2019) ; Kohli et al. (2007) ; Padrakali and Chitra Chellam (2017) have outlined a set of specific criteria or features that good examples should exhibit. We annotated each sample for the presence or absence of each criterion. This yields a detailed profile of which features are present in each data point. We then designated positive examples as those that satisfy many of the expert criteria (top-scoring "high-quality” samples) and negative examples as those that violate or lack many of the criteria (bottom-scoring "low-quality” samples). This dataset construction emphasizes concrete feature differences rather than a subjective quality judgment. Model All experiments in this section use GPT-4o OpenAI (2024) as the LLM, which is shown to have strong capabilities in understanding nuanced text and performing annotation-like tasks Tan et al. (2024) . The model was prompted in a zero-shot manner to generate dataset features