User experience (UX) testing platforms capture many data types related to user feedback and behavior, including clickstream, survey responses, screen recordings of participants performing tasks, and participants’ think-aloud audio. Analyzing these multimodal data channels to extract insights remains a time-consuming, manual process for UX researchers. This paper presents a large language model (LLM) approach for generating insights from multimodal UX testing data. By unifying verbal, behavioral, and design data streams into a novel natural language representation, we construct LLM prompts that generate insights combining information across all data types. Each insight can be traced back to behavioral and verbal evidence, allowing users to quickly verify accuracy. We evaluate LLM-generated insight summaries by deploying them in a popular remote UX testing platform, and present evidence that they help UX researchers more efficiently identify key findings from UX tests.

Organizations increasingly rely on behavioral analytics tools like Google Analytics to monitor their digital experiences. Making sense of the data these tools capture, however, requires manual event tagging and filtering — often a tedious process. Prior approaches have trained machine learning models to automatically tag interaction data, but draw from fixed digital experience vocabularies which cannot be easily augmented or customized. This paper introduces a novel machine learning interaction pattern that generates customized tag predictions for organizations. The approach employs a general user experience word embedding to bootstrap an initial set of predictions, which can then be refined and customized by users to adapt the underlying vector space, iteratively improving the quality of future predictions. The paper presents a needfinding study that grounds the design choices of the system, and describes a real-world deployment as part of UserTesting.com that demonstrates the efficacy of the approach..

Virtual assistants like Google Assistant and Siri often interface with external apps when they cannot directly perform a task. Currently, developers must manually expose the capabilities of their apps to virtual assistants, using App Actions on Android or Shortcuts on iOS. This paper presents savant, a system that automatically generates task shortcuts for virtual assistants by mapping user tasks to relevant UI screens in apps. For a given natural language task (e.g., "send money to Joe"), savant leverages text and semantic information contained within UIs to identify relevant screens, and intent modeling to parse and map entities (e.g., "Joe") to required UI inputs. Therefore, savant allows virtual assistants to interface with apps and handle new tasks without requiring any developer effort. To evaluate savant, we performed a user study to identify common tasks users perform with virtual assistants. We then demonstrate that savant can find relevant app screens for those tasks and autocomplete the UI inputs.

In the last two decades, Emoji have become a mainstay of digital communication, allowing ordinary people to convey ideas, concepts, and emotions with just a few Unicode characters. While emoji are most often used to supplement text in digital communication, they comprise a powerful and expressive vocabulary in their own right. In this paper, we study the affordances of 'emoji-first' communication, in which sequences of emoji are used to describe concepts without textual accompaniment. To investigate the properties of emoji-first communication, we built and released Opico, a social media mobile app that allows users to create reactions --- sequences of between one and five emoji --- and share them with a network of friends. We then leveraged Opico to collect a repository of more than 3700 emoji reactions from more than 1000 registered users, each tied to one of 2441 physical places.

Sorting items by user rating is a fundamental interaction pattern of the modern Web, used to rank products (Amazon), posts (Reddit), businesses (Yelp), movies (YouTube), and more. To implement this pattern, designers must take in a distribution of ratings for each item and define a sensible total ordering over them. This is a challenging problem, since each distribution is drawn from a distinct sample population, rendering the most straightforward method of sorting — comparing averages — unreliable when the samples are small or of different sizes. Several statistical orderings for binary ratings have been proposed in the literature (e.g., based on the Wilson score, or Laplace smoothing), each attempting to account for the uncertainty introduced by sampling. In this paper, we study this uncertainty through the lens of human perception, and ask 'How do people sort by ratings?' In an online study, we collected 48,000 item-ranking pairs from 4,000 crowd workers along with 4,800 rationales, and analyzed the results to understand how users make decisions when comparing rated items. Our results shed light on the cognitive models users employ to choose between rating distributions, which sorts of comparisons are most contentious, and how the presentation of rating information affects users' preferences.

Recently, researchers have developed black-box approaches to mine design and interaction data from mobile apps. Although the data captured during this interaction mining is descriptive, it does not expose the design semantics of UIs: what elements on the screen mean and how they are used. This paper introduces an automatic approach for generating semantic annotations for mobile app UIs. Through an iterative open coding of 73k UI elements and 720 screens, we contribute a lexical database of 25 types of UI components, 197 text button concepts, and 135 icon classes shared across apps. We use this labeled data to learn code-based patterns to detect UI components and to train a convolutional neural network that distinguishes between icon classes with 94% accuracy. To demonstrate the efficacy of our approach at scale, we compute semantic annotations for the 72k unique UIs in the Rico dataset, assigning labels for 78% of the total visible, non-redundant elements.

Outfits in online fashion data are composed of items of many different types (e.g. top, bottom, shoes) that share some stylistic relationship with one another. A representation for building outfits requires a method that can learn both notions of similarity (for example, when two tops are interchangeable) and compatibility (items of possibly different type that can go together in an outfit). This paper presents an approach to learning an image embedding that respects item type, and jointly learns notions of item similarity and compatibility in an end-to-end model. To evaluate the learned representation, we crawled 68,306 outfits created by users on the Polyvore website. Our approach obtains 3-5% improvement over the state-of-the-art on outfit compatibility prediction and fill-in-the-blank tasks using our dataset, as well as an established smaller dataset, while supporting a variety of useful queries.

Advances in computer vision and machine learning are changing the way people dress, and buy clothes. Given the vast space of fashion problems, where can data-driven technologies provide the most value? To understand consumer pain points and opportunities for technological interventions, this paper presents the results from two independent need-finding studies that explore the gold-standard of personalized shopping: interacting with a personal stylist. Through interviews with five personal stylists, we study the range of problems they address and their in-person processes for working with clients. In a separate study, we investigate how styling experiences map to online settings by building and releasing a chatbot that connects users to one-on-one sessions with a stylist, acquiring more than 70 organic users in three weeks. These conversations reveal that in-person and online styling sessions share similar goals, but online sessions often involve smaller problems that can be resolved more quickly. Based on these explorations, we propose future personalized, online interactions that address consumer trust and uncertainty, and discuss opportunities for automation.

To evaluate the performance of mobile app designs, designers and researchers employ techniques such as A/B, usability, and analytics-driven testing. While these are all useful strategies for evaluating known designs, comparing many divergent solutions to identify the most performant remains a costly and difficult problem. This paper introduces a design performance testing approach that leverages existing app implementations and crowd workers to enable comparative testing at scale. This approach is manifest in ZIPT, a zero-integration performance testing platform that allows designers to collect detailed design and interaction data over any Android app — including apps they do not own and did not build. Designers can deploy scripted tests via ZIPT to collect aggregate user performance metrics (e.g., completion rate, time on task) and qualitative feedback over third-party apps. Through case studies, we demonstrate that designers can use ZIPT's aggregate data and visualizations to understand the relative performance of interaction patterns found in the wild, and identify usability issues in existing Android apps.

Data-driven models help mobile app designers understand best practices and trends, and can be used to make predictions about design performance and support the creation of adaptive UIs. This paper presents Rico, the largest repository of mobile app designs to date, created to support five classes of data-driven applications: design search, UI layout generation, UI code generation, user interaction modeling, and user perception prediction. To create Rico, we built a system that combines crowdsourcing and automation to scalably mine design and interaction data from Android apps at runtime. The Rico dataset contains design data from more than 9.7k Android apps spanning 27 categories. It exposes visual, textual, structural, and interactive design properties of more than 72k unique UI screens. To demonstrate the kinds of applications that Rico enables, we present results from training an autoencoder for UI layout similarity, which supports query-by-example search over UIs.

Design plays an important role in adoption of apps. App design, however, is a complex process with multiple design activities. To enable data-driven app design applications, we present interaction mining --- capturing both static (UI layouts, visual details) and dynamic (user flows, motion details) components of an app's design. We present ERICA, a system that takes a scalable, human-computer approach to interaction mining existing Android apps without the need to modify them in any way. As users interact with apps through ERICA, it detects UI changes, seamlessly records multiple data-streams in the background, and unifies them into a user interaction trace. Using ERICA we collected interaction traces from over a thousand popular Android apps. Leveraging this trace data, we built machine learning classifiers to detect elements and layouts indicative of 23 common user flows. User flows are an important component of UX design and consists of a sequence of UI states that represent semantically meaningful tasks such as searching or composing. With these classifiers, we identified and indexed more than 3000 flow examples, and released the largest online search engine of user flows in Android apps.

The outfits people wear contain latent fashion concepts capturing styles, seasons, events, and environments. Fashion theorists have proposed that these concepts are shaped by design elements such as color, material, and silhouette. A dress may be "bohemian" because of its pattern, material, trim, or some combination of them: it is not always clear how low-level elements translate to high-level styles. In this paper, we use polylingual topic modeling to learn latent fashion concepts jointly in two languages capturing these elements and styles. Using this latent topic formation we can translate between these two languages through topic space, exposing the elements of fashion style. We train the polylingual topic model (PLTM) on a set of more than half a million outfits collected from Polyvore, a popular fashion-based social network. We present novel, data-driven fashion applications that allow users to express their needs in natural language just as they would to a real stylist and produce tailored item recommendations for these style needs.