Capturing Human Behavior in Quantitative Models

Businesses operating across both online and offline platforms face numerous challenges in offering relevant products or personalized content to users by accounting for their unique preferences and choice behavior. Moreover, with recent advances in artificial intelligence (AI) technologies, it has also become essential to evaluate and update traditional frameworks to meet user needs. In this thesis, we explore some of these key issues and propose innovative solutions to overcome them. First, we develop efficient and data-driven algorithms to solve the assortment optimization problem for arbitrary feasible regions with no compact structure. Our algorithms utilize binary search and maximum inner product search (MIPS) techniques to identify near-optimal solutions for large-scale instances efficiently under the simpler multinomial logit choice model. For an arbitrary collection of assortments, our algorithms can find a solution in time that is sub-linear in the number of assortments, and for the simpler case of cardinality constraints - linear in the number of items (existing methods are quadratic or worse). Empirical validations using a real world dataset (in addition to experiments using semi-synthetic data based on the Billion Prices dataset and several retail transaction datasets) show that our algorithms are competitive even when the number of items is ~ 10^5 (10 times larger instances than previously studied). Second, we focus our attention on developing richer choice models tailored for social media settings. Specifically, we address the engagement maximization problem on platforms like X/Twitter by explicitly modeling choice behavior of users. We treat the engagement prediction task as a multi-label classification problem based on tweet topics derived from clustering the tweets. Next, we design a deep learning model to predict user engagement with specific tweet topics by incorporating engagement histories and investigate the impact of tweet recommendations on user engagement outcomes by solving a tweet optimization problem. Finally, we demonstrate the effectiveness of our proposed model by comparing its performance to popular methods from the existing literature. Third, we investigate the challenge of quantifying the personality traits of language models, considering their growing integration into various applications, including chatbots and automatic summarizers. Our analysis is grounded in the observation that language models inherently capture biases present in their training data. In this study, we introduce an innovative approach that leverages the Big Five personality assessment questionnaire, commonly used for evaluating human personalities, to quantify the personalities of these models. Additionally, we propose a few approaches for modifying the personas of these language models and present results validating them. Finally, we discuss the challenges associated with the extensive engineering required for the processing of weblogs for creating datasets required for training the recommendation models and the difficulties in their interpretation by non-experts. We propose leveraging screenshots of users' internet browsing activities for comprehending users' preferences. Specifically, we introduce a sophisticated and interactive recommendation framework, InteraRec, that captures high-frequency screenshots of web pages as users navigate through a website. Leveraging state-of-the-art multimodal large language models (MLLMs), it extracts valuable insights into user preferences from these screenshots by generating a textual summary based on predefined keywords. Subsequently, an LLM-integrated optimization setup utilizes this summary to generate tailored recommendations. Moreover, we employ a re-ranking approach leveraging InteraRec to improve the performance of the existing session-based recommendation models. Through extensive experiments, we validate the efficacy of InteraRec on a newly curated screenshot dataset based on user interactions on the Amazon website.