Building Human Values into Recommender Systems: An Interdisciplinary Synthesis

Stray, Jonathan; Halevy, Alon; Assar, Parisa; Hadfield-Menell, Dylan; Boutilier, Craig; Ashar, Amar; Beattie, Lex; Ekstrand, Michael D.; Leibowicz, Claire; Sehat, Connie Moon; Johansen, Sara; Kerlin, Lianne; Vickrey, David; Singh, Spandana; Vrijenhoek, Sanne; Zhang, Amy; Andrus, McKane; Helberger, Natali; Proutskova, Polina; Mitra, Tanushree; Vasan, Nina

doi:10.48550/arxiv.2207.10192

Cited by 8 publications

(6 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…also highlight a host of validity concerns impacting RAIs, including many discussed in Section 2.2. Recent work has also surfaced validity issues in content moderation [37] and recommender systems [65,89]. Despite this growing awareness, we currently lack a holistic understanding of validity threats to prediction targets in human-AI decision-making.…”

Section: Measurement and Validity In Algorithmic Systemsmentioning

confidence: 99%

Ground(less) Truth: A Causal Framework for Proxy Labels in Human-Algorithm Decision-Making

Guerdan¹,

Coston²,

Wu³

et al. 2023

Preprint

View full text Add to dashboard Cite

A growing literature on human-AI decision-making investigates strategies for combining human judgment with statistical models to improve decision-making. Research in this area often evaluates proposed improvements to models, interfaces, or workflows by demonstrating improved predictive performance on "ground truth" labels. However, this practice overlooks a key difference between human judgments and model predictions. Whereas humans reason about broader phenomena of interest in a decision -including latent constructs that are not directly observable, such as disease status, the "toxicity" of online comments, or future "job performance"predictive models target proxy labels that are readily available in existing datasets. Predictive models' reliance on simplistic proxies makes them vulnerable to various sources of statistical bias. In this paper, we identify five sources of target variable bias that can impact the validity of proxy labels in human-AI decision-making tasks. We develop a causal framework to disentangle the relationship between each bias and clarify which are of concern in specific human-AI decision-making tasks. We demonstrate how our framework can be used to articulate implicit assumptions made in prior modeling work, and we recommend evaluation strategies for verifying whether these assumptions hold in practice. We then leverage our framework to re-examine the designs of prior human subjects experiments that investigate human-AI decision-making, finding that only a small fraction of studies examine factors related to target variable bias. We conclude by discussing opportunities to better address target variable bias in future research.

show abstract

Section: Measurement and Validity In Algorithmic Systemsmentioning

confidence: 99%

Ground(less) Truth: A Causal Framework for Proxy Labels in Human-Algorithm Decision-Making

Guerdan¹,

Coston²,

Wu³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…A recommender is a personalized information filter that selects, for each individual, a small set of items out of a much larger pool [5]. Recommenders differ from search engines in that they typically produce personalized results, and they are often invoked without a specific user query, such as a news recommender presenting significant new events without requiring the user to search for them explicitly.…”

Section: Recommender System Biasesmentioning

confidence: 99%

“…Mitigating such misalignment requires measuring it, and one of the most straightforward sources of additional information is user surveys. A wide variety of surveys are used to evaluate and train commercial recommender systems [5]. Some of these surveys merely ask the user to rate a previous recommendation, but some attempt to measure more abstract constructs, such as whether an item is "worth your time" or contributes to "meaningful social interactions" [46].…”

Section: Use Of Survey Datamentioning

confidence: 99%

“…When the recommender then selects items for a user, the predicted response for that user is used as a factor in content ranking. This approach is used by YouTube [47,48], Facebook [46,49], and elsewhere [5]. While generating "fake" survey responses may seem a strange approach, conventional recommenders already depend on predicting engagement-future user reactions-which is not drastically different from predicting a survey response.…”

Section: Use Of Survey Datamentioning

confidence: 99%

See 1 more Smart Citation

The AI Learns to Lie to Please You: Preventing Biased Feedback Loops in Machine-Assisted Intelligence Analysis

Stray

2023

Analytics

View full text Add to dashboard Cite

Researchers are starting to design AI-powered systems to automatically select and summarize the reports most relevant to each analyst, which raises the issue of bias in the information presented. This article focuses on the selection of relevant reports without an explicit query, a task known as recommendation. Drawing on previous work documenting the existence of human-machine feedback loops in recommender systems, this article reviews potential biases and mitigations in the context of intelligence analysis. Such loops can arise when behavioral “engagement” signals such as clicks or user ratings are used to infer the value of displayed information. Even worse, there can be feedback loops in the collection of intelligence information because users may also be responsible for tasking collection. Avoiding misalignment feedback loops requires an alternate, ongoing, non-engagement signal of information quality. Existing evaluation scales for intelligence product quality and rigor, such as the IC Rating Scale, could provide ground-truth feedback. This sparse data can be used in two ways: for human supervision of average performance and to build models that predict human survey ratings for use at recommendation time. Both techniques are widely used today by social media platforms. Open problems include the design of an ideal human evaluation method, the cost of skilled human labor, and the sparsity of the resulting data.

show abstract

“…In general, the problem of determining whether engagement means an item is genuinely valuable or merely attention-getting requires the collection of some sort of additional feedback, and there are many ways to do this including providing new user controls and directly asking a subset of users with surveys. Better conflict is one of many values we might want social media to support, and the methods to measure and operationalize these values are developing rapidly (Stray et al 2022).…”

Section: Strategy 1: Reduce Engagement Incentives To Divisivenessmentioning

confidence: 99%