Digital transaction platforms have been growing significantly in recent years. New businesses spawned by them have captured the attention of scholars and practitioners alike. However, high platform failure risks have plagued the industry, and the literature has only given this issue scant treatment. Previous research on firm failures mainly relies on the firm's 10-K report data that describes their operations and financial performance; however, this approach does not apply in the context of digital transaction platforms, most of which are not public firms and therefore not required to issue 10-K reports publicly. That is, there is a research gap between the literature on firm failures and the need for understanding digital transaction platform failures. In the big data era, online contents have become an important part of the platform economy, yet few research studies on platform failures from an online content perspective. To address this research gap, this dissertation focuses on platform failure from an online contents perspective and develops several deep learning-based methods in evaluating the impacts of online contents on platform failures in order to predict platform failure risks.

The first study develops a data-driven analytic based on deep learning techniques to mine customer sentiments towards digital transaction platforms using customer-generated contents; in order to deal with unique issues in assessing digital transaction platform failure risks, we develop alternative measures on customer sentiment by means of multiple scales, such as positive, neutral, negative, and extreme negative, which has been shown to be a useful innovation. In this study, we apply deep learning techniques to extract customer sentiments towards platforms and use the mining results to assess platform status and trends. More specifically, we apply a domain-related corpus, the word-embedding technique (Word2vec), and the deep learning algorithm (bidirectional long-short term memory) in the analysis of customer sentiments towards platforms based on customer-generated contents. We compare the performance of this method with that of lexicon-based methods in measuring customer sentiments to demonstrate its efficacy. The results show a significant improvement in accuracy by capturing context-aware information from colloquial customer contents.

The second study applies the customer-sentiments-mining method from the first study to investigate the impact of customer sentiments on platform failures and the extent to which platform failure risks can be predicted. Drawing on the resource-based view (RBV), customer sentiments are critical social resources for the platform and determine firm survival. Previous research on firm failures mainly relies on the firm's 10-K report data that describes their operations and financial performance; however, this approach does not apply in the context of digital transaction platforms, most of which are not public firms and therefore not required to issue 10-K reports publicly. We model the relationship between customer sentiments and platform failures by the time-dependent Cox model. The result shows significant links between customer sentiments and platform failures. A deviance analysis is adopted to validate that those customer sentiments are significant predictors for platform failures compared with platform attributes and political factors. We further calculate the accuracy improvement by incorporating customer sentiments in platform failure risk prediction and show that the predictive power of customer sentiments varies across different types of customer sentiments and prediction periods.

The third study investigates the problem of platform failure prediction through the use of various types of online contents from key players of digital transaction platforms such as customers, platforms, and third parties. Although the online contents are studies in some contexts, such as customer purchase behavior and firm equity, a study to examine how online contents generated from various actors are related to platform failure is still lacking. To remedy this gap, we propose an interpretable deep learning approach for predicting platform failures associated with online contents. Using the predictive XGBoost algorithm, we find that customer-generated contents and third-party-generated contents are strong predictors of platform failures. Furthermore, we build causal survival forests to facilitate the interpretation of our predictive results. The results show that the presence of essays and third-party generated contents has a positive relationship with platform failure. In contrast, the presence of reviews has a negative relationship with platform failure.