312 - Bot or Not: A Novel Machine Learning Strategy for Identifying Fraudulent Survey Responses

Publication Number: 312.4165

Kacie Kidd, West Virginia University School of Medicine, Morgantown, WV, United States; Suzanne E. Vogler, West Virginia University, Morgantown, WV, United States; Maryam Khodaverdi, WVCTSI, Morgantown, WV, United States; Gina Sequeira, Seattle Children's, Seattle, WA, United States; Nicole F. Kahn, University of Washington School of Medicine, Seattle, WA, United States; Nadia Dowshen, Childrens Hospital of Philadelphia, Philadelphia, PA, United States; Elizabeth Miller, University of Pittsburgh School of Medicine, Pittsburgh, PA, United States; Brad Price, West Virginia University, Morgantown, WV, United States; Alfgeir Logi. Kristjansson, West Virginia University, Morgantown, WV, United States

Background: Online research survey predominance has led to increased malicious respondents, particularly in the setting of incentivized surveys. While some strategies exist to protect survey data, there has been a marked rise in the accessibility of artificial intelligence (AI) platforms allowing bot masters to deploy never-ending volumes of human-like bot respondents in the hopes of receiving payment. Novel approaches for survey data protection are needed to combat this increasing threat, particularly when studying hard-to-reach populations.

Objective: We aimed to develop a machine learning approach informed by qualitative data using natural language processing to rapidly delineate bot responses modeled off of data from a serially bot-swarmed online survey dataset.

Design/Methods: Data from a survey instrument aimed at assessing participant experiences was compiled following a series of suspected bot swarms. Traditional data cleaning techniques – including those specific for bot detection such as average completion time and consistency assessments – were applied to the dataset followed by expert review by the research team and a community partner group to determine if responses were more likely human or bot. Using machine learning and natural language processing, we tokenized and normalized text responses to a single qualitative survey item via Term Frequency-Inverse Document Frequency. The model deployed is a Support Vector Machine (SVM) Classifier utilizing Adaptive Synthetic Sampling (ADASYN) to generate synthetic data for underrepresented classes. We assessed the subsequent model’s performance with and without dimensionality reduction. Tuning parameters were selected via 5-fold cross-validation.

Results: Survey data included 4,781 responses with approximately 5% suspected to be human respondents per expert review. Models performed similarly with and without dimensionality reduction so such decomposition was used to reduce computational effort and simplify the model. The fitted model predicted human versus bot responses with over 87% accuracy compared to direct expert review and was able to identify over 80% of actual human responses.

Conclusion(s): The novel approach described serves as a potential strategy for confronting a new era in online survey research wherein seemingly human AI bot respondents have the potential to disrupt data collection and interpretation, or worse, lead to false conclusions. Use of a machine learning approach paired with natural language processing, as employed in this study may allow rapid reliable separation of human from AI bot and be central to the future of online survey research.

Figure 1
Figure 1: Histograms depicting the distribution of words in suspected bot (A) vs. human (B) responses demonstrating a lower frequency and larger variety of word use in suspected human responses.

Figure 2
Figure 2: Word clouds depicting the frequency of words in suspected bot (A) vs. human (B) responses demonstrating a lower frequency and larger variety of word use in suspected human responses.