Data-driven design for industrial applications

Lin, Kangcheng

Permalink

https://hdl.handle.net/2142/132517 Copy

Description

Title

Data-driven design for industrial applications

Author(s)

Lin, Kangcheng

Issue Date

2025-11-24

Director of Research (if dissertation) or Advisor (if thesis)

Kim, Harrison

Doctoral Committee Chair(s)

Kim, Harrison

Committee Member(s)

• Sowers, Richard
• Wang, Pingfeng
• Kontou, Eleftheria

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Industrial Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Product Design
• Machine Learning
• Deep Learning

Abstract

In an era where vast troves of consumer data are continuously generated online, data-driven design has become essential for developing products that truly meet user needs. This dissertation leverages emerging data sources and analytics to inform industrial product design, addressing limitations of traditional methods (e.g. surveys and focus groups) that are often slow, costly, and narrow in scope. By mining large-scale user-generated content – from textual reviews to video testimonials – and applying advanced machine learning techniques, the research develops a comprehensive framework for integrating data-driven insights into the design process. The work is centered around four key challenges in modern product design: (1) extracting important product attributes and customer needs from unstructured feedback, (2) modeling consumer behavior and purchase decisions, (3) responding adaptively to major market disruptions (such as COVID-19), and (4) harnessing generative AI for early-stage design evaluation. Through six interrelated studies, the dissertation demonstrates how these challenges can be met using a spectrum of methodologies including topic modeling, sentiment analysis, discrete choice modeling, longitudinal trend analysis, mining of video reviews, and fine-tuned large language models. First, an automated data-mining pipeline is developed to transform online product reviews into actionable design insights. Using topic modeling (Latent Dirichlet Allocation) and aspect-based sentiment analysis, this method performs an Importance–Performance Analysis that pinpoints which product attributes customers find most important and how well current offerings perform on those attributes. The analysis, applied at scale, identifies features that are both highly valued and underperforming, thus highlighting strategic targets for improvement. This initial study demonstrates a scalable approach to discover latent customer priorities from unstructured text, providing designers with evidence-based guidance on where to focus development efforts. Next, the dissertation extends the analysis of customer feedback by considering the nonlinear nature of user satisfaction. A study inspired by the Kano model employs an explainable neural network to classify product features into “must-have,” “performance,” or “attractive” categories. This classification reveals nuanced relationships between product attributes and customer satisfaction – for example, improving a “must-have” feature yields fundamental gains in satisfaction, whereas enhancing an “attractive” feature might delight users but with diminishing returns. By elucidating that not all product improvements translate equally into customer satisfaction, this contribution refines the strategic prioritization of design requirements. It underscores the importance of understanding which features will have the most meaningful impact on the user experience, beyond what simple importance rankings can show. The research then moves from understanding preferences to predicting consumer choice. In a dedicated study on consumer behavior modeling, a data-driven discrete choice model is constructed to quantify how product attributes and online reputation factors jointly influence purchase decisions. Since consumers’ actual choice sets are not directly observable from sales data, this approach first reconstructs plausible choice sets by clustering products with similar specifications and market profiles. Within these sets, a multinomial logit model is applied, revealing how both intrinsic features (e.g. functional attributes, performance metrics) and extrinsic signals (e.g. star ratings, number of reviews) affect the probability of a product being chosen. The findings highlight that social proof factors – high ratings and positive reviews – significantly sway decision-making alongside traditional engineering features. This insight calls attention to the role of online perception in product success, emphasizing that designers and managers must account for user feedback dynamics and product reputation when evaluating design alternatives. Another critical dimension addressed in this dissertation is adaptability to market disruptions. A longitudinal study examines how consumer preferences and product usage priorities shifted in response to the COVID-19 pandemic, a significant external disruption to many industries. By analyzing time-stamped review data before and after the onset of COVID-19, the research detects marked changes in sentiment and importance across product features. For instance, in personal electronics, features enabling remote work and virtual connectivity rose sharply in importance during the pandemic, while concerns about price and durability also evolved with changing consumer circumstances. Using topic modeling and trend analysis, the study provides a data-driven lens for detecting emergent user needs: it shows designers which attributes gained or lost prominence due to the new context. These insights inform product design strategies under disruptive events, guiding companies to realign feature priorities and innovate solutions that better serve emerging customer requirements. The results demonstrate that an agile, data-driven approach can successfully capture the “voice of the customer” amid upheaval, enabling timely design adaptations that maintain product relevance and competitiveness. To broaden the scope of data-driven design, the dissertation also explores user-generated video reviews as a rich, underutilized source of product insight. In one study, a novel four-stage video review mining methodology is presented: it transcribes spoken content from online videos, preprocesses and filters the text, uses topic modeling to extract frequently discussed features, and then analyzes sentiment and feature importance. Applied to consumer tech products, this approach validates that YouTube reviews and similar videos can be systematically mined to uncover what users care about most. The study reveals that video feedback, which often contains in-depth demos and honest opinions, can surface product attributes (and pain points) that might not be as evident in written reviews. By successfully extracting preferences from non-traditional data modalities, this contribution expands the toolbox of data-driven design: it underscores that designers can tap into multimedia feedback channels to gain a more holistic understanding of user experience. Building on the analysis of video data, a further study investigates product ecosystem effects through the lens of customer reviews. By analyzing transcripts and comments from videos, this work examines how a core product and its complementary products or accessories are discussed together, shedding light on issues of interoperability and ecosystem integration. For example, in the context of a camera system, video reviews might reveal how the camera body and various lenses or attachments work in unison and which combinations users find most effective or problematic. This analysis of interconnected user feedback uncovers deeper insights into how the value of a product is influenced by its broader ecosystem of use. The findings emphasize that in modern markets, a product’s success is often tied not just to its standalone features, but also to how well it works with other products in its environment. By mining sentiments about cross-product usage, designers can identify opportunities to improve overall user satisfaction through better compatibility, standardization, or ecosystem design. This study’s contribution lies in demonstrating a method to capture those complex, system-level design insights from messy, real-world data, highlighting an increasingly important aspect of product design strategy. Finally, the dissertation looks toward the future of design evaluation by leveraging generative Artificial Intelligence. The last study introduces a novel approach in which a large language model (LLM) is fine-tuned on a corpus of product reviews to act as a surrogate for human users in the early stages of design. By training the LLM on domain-specific feedback (including tailoring to distinct customer segments or personas), the research enables the model to generate realistic narratives and evaluations of conceptual product ideas that have not yet been released. These AI-generated “virtual reviews” mimic the tone and content of genuine user feedback, providing an instantaneous, confidential evaluation tool for designers. This approach allows product teams to solicit constructive critiques of prototypes or concept sketches without organizing actual user focus groups, which can be time-consuming and risk leaking intellectual property. In experiments, the fine-tuned LLM is able to highlight potential strengths and weaknesses of new product concepts in a manner similar to real customers – for example, predicting that a certain design feature might attract praise for its innovation but also criticism for usability. This pioneering study demonstrates how generative AI can serve as a creative partner in design, augmenting the human-centered design process with machine-generated insights and enabling more rapid, iterative refinement of product ideas. In summary, “Data-Driven Design for Industrial Applications” brings together a suite of advanced data analytics and AI techniques to enhance product design decision-making from multiple angles. The six studies collectively show a progression from mining what customers say about existing products, to understanding why they behave as they do, to anticipating how they might respond to new circumstances or ideas. Importantly, each study contributes a distinct piece to the overarching goal of making product development more customer-centric, proactive, and evidence-based. By identifying critical attributes from unstructured feedback, the research helps designers focus on features that truly matter. By modeling consumer choice behavior, it quantifies the influence of product improvements and reputation factors on market success. By analyzing disruptions like COVID-19, it provides a playbook for adapting designs when user priorities shift unexpectedly. By exploiting rich media sources and generative models, it pushes the boundary of how designers can gather and even simulate user feedback in the digital age. Taken together, these contributions illustrate that harnessing big data and AI – from classical machine learning to state-of-the-art LLMs – can fundamentally improve the agility and efficacy of the product design process. This dissertation thus charts a path for industry to leverage data-driven methods not only to understand the customer of today, but to better predict and meet the needs of the customer of tomorrow.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132517

Copyright and License Information

Copyright 2025 Kangcheng Lin

Owning Collections