Abstract Mixed‐initiative interaction is a key feature of Conversational Information Seeking (CIS), where both users and agents actively participate in dialogue. Given the potential risks associated with agent‐initiative interactions, a CIS process model is needed to guide when and how the agent should take the initiative. Conversation analysis was employed to derive a speech act framework from the 37,268 utterances in the Wizard of Search Engine dataset. Lag sequential analysis was then used to identify adjacency pairs of speech acts with significant transition probabilities. This study identified six user acts, including querying, assessing, elaborating, reformulating, instructing, and deciding, and seven agent acts, including answering, inquiring, checking, eliciting, soliciting, informing, and offering. These acts were connected based on their adjacency, resulting in the CIS process model. This model comprises one fundamental querying‐answering sub‐process and four optional sub‐processes: need negotiation, process collaboration, result evaluation, and query elicitation. The robustness of this model was evaluated on another CIS dataset, ConvSearch, with new behavioral patterns emerging. This study identifies the multi‐stage, iterative, and dynamic nature of mixed‐initiative user–agent interaction in CIS, offers methodological insights into CIS data collection, annotation, and analysis, and provides guidance for the development and evaluation of CIS agents.

Abstract In combating filter bubbles, an undesirable consequence of personalized recommendations, prior research has focused on improving algorithms to increase the diversity of the content recommended. Following a user‐centered approach firmly grounded in information science, this study is dedicated to optimizing interaction patterns with algorithmic affordances, aiming to augment the diversity of the content consumed and induce favorable attitude changes. A controlled experiment was conducted on a mock personalized recommender system that provided both information and interactivity affordances, exemplified by stance labels and stance‐based filters, respectively. A total of 142 participants were recruited to browse recommendations generated by the system on a specific controversial topic, and the selectivity of their information consumption behavior and the change in their attitude extremity were measured. It was found that both types of affordances were effective in reducing users' behavioral selectivity. While stance labels inhibited the consumption of pro‐attitudinal information, stance‐based filters facilitated the consumption of counter‐attitudinal information. Furthermore, the affordances could immediately mitigate the attitude extremity of those with a higher level of algorithmic literacy. The findings not only enrich the growing body of literature on filter bubbles but also offer valuable implications for the affordance design practices of personalized recommender systems.