User Modeling, Adaptation, and Personalization

Social networking systems originally emerged as tools for keeping up with the daily lives of friends and strangers. They have established themselves as valuable resources and means to satisfy information needs. The challenge with information seeking through social networks is that their immense success and popularity is also a weakness. The data deluge facing users has surpassed comfortably managed levels and can impact on the quality and relevance of the information consumed. We developed a personalized model for predicting the relevance of news feed items, in order to facilitate personalized feeds. Results of a live analysis show that our approach successfully identifies and promotes relevant feed items, with the knock-on effects of increasing interaction between users and the contribution of user generated content.

We present a realistic simulation framework to examine the impact of sensor noise on the performance of user models in the museum domain. Our contributions are (1) models to simulate noisy visit trajectories as time-stamped sequences of (

x

,

y

) positional coordinates which reflect

walking

and

hovering

behaviour; (2) a discriminative inference model that distinguishes between hovering and walking on the basis of (simulated) noisy sensor observations; (3) a model that infers viewed exhibits from hovering coordinates; and (4) a model that predicts the next exhibit on the basis of inferred (rather than known) viewed exhibits. Our staged evaluation assesses the effect of these models (in combination with sensor noise) on inferential and predictive performance, thus shedding light on the reliability attributed to inferences drawn from sensor observations.

We present G

ecko

mmender

, a mobile system for personalised theme and tour recommendations in museums, based on a digital site-map representation. Star ratings provided by visitors for seen exhibits are used to predict ratings for unvisited exhibits. The predicted ratings in turn form the basis for recommendations. These recommendations are presented in one of three display modes:

StarMap

– stars on the site map,

HeatMap

– colours from green to red that indicate the interestingness of exhibits (from interesting to not interesting respectively), and

TourPlann

– directed personalised tours through the museum. G

ecko

mmender

was evaluated in a field study at Melbourne Museum (Melbourne, Australia). Our results show that (1) most participants enjoyed G

ecko

mmender

, (2) G

ecko

mmender

’s recommendations often reflected the participants’ personal interests, and (3)

HeatMap

was the most popular display mode.

We present an approach for propagation of user interests in ontology-based user models taking into account the properties declared for the concepts in the ontology. Starting from initial user feedback on an object, we calculate user interest in this particular object and its properties and further propagate user interest to other objects in the ontology, similar or related to the initial object. The similarity and relatedness of objects depends on the number of properties they have in common and their corresponding values. The approach we propose can support finer recommendation modalities, considering the user interest in the objects, as well as in singular properties of objects in the recommendation process. We tested our approach for interest propagation with a real adaptive application and obtained an improvement with respect to IS-A-propagation of interest values.

The study goal was to evaluate whether Electroencephalography (EEG) estimates of attention and cognitive workload captured as students solved math problems could be used to predict success or failure at solving the problems. Students (N = 16) solved a series of SAT math problems while wearing an EEG headset that generated estimates of sustained attention and cognitive workload each second. Students also reported on their level of frustration and the perceived difficulty of each problem. Results from a Support Vector Machine (SVM) training indicated that problem outcomes could be correctly predicted from the combination of attention and workload signals at rates better than chance. EEG data were also correlated with students’ self-report of problem difficulty. Findings suggest that relatively non-intrusive EEG technologies could be used to improve the efficacy of tutoring systems.

Users in recommender systems often express their opinions about different items by rating the items on a fixed rating scale. The rating information provided by the users is used by the recommender systems to generate personalized recommendations for them. Few recent research work on rating based recommender systems advocate the use of preference relations instead of absolute ratings in order to produce better recommendations. Use of preference relations for neighborhood based collaborative recommendation has been looked upon in recent literature. On the other hand, Matrix Factorization algorithms have been shown to perform well for recommender systems, specially when the data is sparse. In this work, we propose a matrix factorization based collaborative recommendation algorithm that considers preference relations. Experimental results show that the proposed method is able to achieve better recommendation accuracy over the compared baseline methods.

In this paper, we present an approach for modeling user trustworthiness when traffic information is exchanged between vehicles in transportation environments. Our multi-faceted approach to trust modeling combines priority-based, role-based and experience-based trust, integrated with a majority consensus model influenced by time and location, for effective route planning. The proposed representation for the user model is outlined in detail (integrating ontological and propositional elements) and the algorithm for updating trust values is presented as well. This trust modeling framework is validated in detail through an extensive simulation testbed that models vehicle route planning. We are able to show decreased average path time for vehicles when all facets of our trust model are employed in unison. Included is an interesting confirmation of the value of distinguishing direct and indirect observations of users.

In this article, we analyze and compare user behavior on two different microblogging platforms: (1) Sina Weibo which is the most popular microblogging service in China and (2) Twitter. Such a comparison has not been done before at this scale and is therefore essential for understanding user behavior on microblogging services. In our study, we analyze more than 40 million microblogging activities and investigate microblogging behavior from different angles. We (i) analyze how people access microblogs and (ii) compare the writing style of Sina Weibo and Twitter users by analyzing textual features of microposts. Based on semantics and sentiments that our user modeling framework extracts from English and Chinese posts, we study and compare (iii) the topics and (iv) sentiment polarities of posts on Sina Weibo and Twitter. Furthermore, (v) we investigate the temporal dynamics of the microblogging behavior such as the drift of user interests over time.

Our results reveal significant differences in the microblogging behavior on Sina Weibo and Twitter and deliver valuable insights for multilingual and culture-aware user modeling based on microblogging data. We also explore the correlation between some of these differences and cultural models from social science research.

In this paper, we propose a general approach to improve student modeling predictive accuracy. The approach was designed based on the assumption that student performance is sampled from multiple, rather than only one, distribution and thus should be modeled by multiple classification models. We applied k-means to identify student performances sampled from those multiple distributions, using no additional features beyond binary correctness of student responses. We trained a separate classification model for each distribution and applied the learned models to unseen students to evaluate our approach. The results showed that compared to the base classifier, our proposed approach is able to improve predictive accuracy: 4.3% absolute improvement in R

2

and 0.03 absolute improvement in AUC, which are not trivial improvements considering the current state of the art in student modeling.

Providing high quality recommendations is important for online systems to assist users who face a vast number of choices in making effective selection decisions.

Collaborative filtering

is a widely accepted technique to provide recommendations based on ratings of similar users. But it suffers from several issues like

data sparsity

and

cold start

. To address these issues, in this paper, we propose a simple but effective method, namely “Merge”, to incorporate social trust information (i.e. trusted neighbors explicitly specified by users) in providing recommendations. More specifically, ratings of a user’s trusted neighbors are merged to represent the preference of the user and to find similar other users for generating recommendations. Experimental results based on three real data sets demonstrate that our method is more effective than other approaches, both in accuracy and coverage of recommendations.

This paper explores the value of eye-tracking data to assess user learning with interactive simulations (IS). Our long-term goal is to use this data in user models that can generate adaptive support for students who do not learn well with these types of unstructured learning environments. We collected gaze data from users interacting with the CSP applet, an IS for constraint satisfaction problems. Two classifiers built upon this data achieved good accuracy in discriminating between students who learn well from the CSP applet and students who do not, providing evidence that gaze data can be a valuable source of information for building user modes for IS.

The ageing of the world population is leading to an increased number of elderly people remaining in their homes, requiring different levels of care. MIA is a user-centric project aimed at monitoring elderly people in order to help them remain safely in their homes, where the design of the system is informed by the requirements of the stakeholders. In this paper, we present the results of two user studies that ascertain the views of elderly people and their informal carers regarding the acceptability and benefits of in-home monitoring technologies: (1) concept mapping coupled with brainstorming sessions, and (2) questionnaires. We then discuss how these requirements affect the design of our monitoring system.

Social tagging systems (STS) model three types of entities (i.e. tag-user-item) and relationships between them are encoded into a 3-order tensor. Latent relationships and patterns can be discovered by applying tensor factorization techniques like Higher Order Singular Value Decomposition (HOSVD), Canonical Decomposition etc. STS accumulate large amount of sparse data that restricts factorization techniques to detect latent relations and also significantly slows down the process of a factorization. We propose to reduce tag space by exploiting clustering techniques so that the quality of the recommendations and execution time are improved and memory requirements are decreased. The clustering is motivated by the fact that many tags in a tag space are semantically similar thus the tags can be grouped. Finally, promising experimental results are presented.

Our research goal is to provide a better understanding of how users engage with online services, and how to measure this engagement. We should not speak of one main approach to measure user engagement – e.g. through one fixed set of metrics – because engagement depends on the online services at hand. Instead, we should be talking of models of user engagement. As a first step, we analysed a number of online services, and show that it is possible to derive effectively simple models of user engagement, for example, accounting for user types and temporal aspects. This paper provides initial insights into engagement patterns, allowing for a better understanding of the important characteristics of how users repeatedly interact with a service or group of services.

Conversational recommender systems allow users to learn and adapt their preferences according to concrete examples. Critiquing systems support such a conversational interaction style. Especially unit critiques offer a low cost feedback strategy for users in terms of the needed cognitive effort. In this paper we present an extension of the experience-based unit critiquing algorithm. The development of our new approach, which we call nearest neighbor compatibility critiquing, was aimed at increasing the efficiency of unit critiquing. We combine our new approach with existing critiquing strategies to ensemble-based variations and present the results of an empirical study that aimed at comparing the recommendation efficiency (in terms of the number of critiquing cycles) of ensemble-based solutions with individual critiquing algorithms.

Personalized electronic program guides help users overcome information overload in the TV and video domain by exploiting recommender systems that automatically compile lists of novel and diverse video assets, based on implicitly or explicitly defined user preferences. In this context, we assume that user preferences can be specified by

program genres

(documentary, sports, …) and that an asset can be labeled by one or more program genres, thus allowing an initial and coarse preselection of potentially interesting assets. As these assets may come from various sources, program genre labels may not be consistent among these sources, or not even be given at all, while we assume that each asset has a possibly short textual description. In this paper, we tackle this problem by considering whether those textual descriptions can be effectively used to automatically retrieve the most related TV shows for a specific program genre. More specifically, we compare a statistical approach called

logistic regression

with an enhanced version of the commonly used vector space model, called

random indexing

, where the latter is extended by means of a negation operator based on quantum logic. We also apply a new feature generation technique based on

explicit semantic analysis

for enriching the textual description associated to a TV show with additional features extracted from Wikipedia.

The task of selecting one among several items in a visual display is extremely common in daily life and is executed billions of times every day on the Web. Attention is vital for selection, but the end-to-end process of what draws and sustains attention, and how that influences selection, remains poorly understood. We study this in a complex multi-item selection setting, where participants selected one among eight news articles presented in a grid layout on a screen. By varying the position, saliency, and topic of the news items, we identify the relative importance of these visual and semantic factors in attention and selection. We present a simple model of attention that predicts many key features such as attention shifts and dwell time per item. Potential applications of our findings include optimizing visual displays to drive user attention.

Many different forms of explanation have been proposed for justifying decisions made by automated systems. However, there is no consensus on what constitutes a

good

explanation, or what information these explanations should include. In this paper, we present the results of a study into how people justify their decisions. Analysis of our results allowed us to extract the forms of explanation adopted by users to justify choices, and the situations in which these forms are used. The analysis led to the development of guidelines and patterns for explanations to be generated by automated decision systems. This paper presents the study, its results, and the guidelines and patterns we derived.

As the world moves towards the social web, criminals also adapt their activities to these environments. Online dating websites, and more generally people recommenders, are a particular target for romance scams. Criminals create fake profiles to attract users who believe they are entering a relationship. Scammers can cause extreme harm to people and to the reputation of the website. This makes it important to ensure that recommender strategies do not favour fraudulent profiles over those of legitimate users. There is therefore a clear need to gain understanding of the sensitivity of recommender algorithms to scammers. We investigate this by (1) establishing a corpus of suspicious profiles and (2) assessing the effect of these profiles on the major classes of reciprocal recommender approaches: collaborative and content-based. Our findings indicate that collaborative strategies are strongly influenced by the suspicious profiles, while a pure content-based technique is not influenced by these users.

Recommender systems are crucial components of most commercial web sites to keep users satisfied and to increase revenue. Thus, a lot of effort is made to improve recommendation accuracy. But when is the best possible performance of the recommender reached? The

magic barrier

, refers to some unknown level of prediction accuracy a recommender system can attain. The magic barrier reveals whether there is still room for improving prediction accuracy, or indicates that any further improvement is meaningless. In this work, we present a mathematical characterization of the magic barrier based on the assumption that user ratings are afflicted with inconsistencies - noise. In a case study with a commercial movie recommender, we investigate the inconsistencies of the user ratings and estimate the magic barrier in order to assess the actual quality of the recommender system.

Data-mined models often achieve good predictive power, but sometimes at the cost of interpretability. We investigate here if selecting features to increase a model’s construct validity and interpretability also can improve the model’s ability to predict the desired constructs. We do this by taking existing models and reducing the feature set to increase construct validity. We then compare the existing and new models on their predictive capabilities within a held-out test set in two ways. First, we analyze the models’ overall predictive performance. Second, we determine how much student interaction data is necessary to make accurate predictions. We find that these reduced models with higher construct validity not only achieve better agreement overall, but also achieve better prediction with less data. This work is conducted in the context of developing models to assess students’ inquiry skill at designing controlled experiments and testing stated hypotheses within a science inquiry microworld.

Reliable personality prediction can have direct impact on many adaptive systems, such as targeted advertising, interface personalization and content customization. We propose an algorithm to infer a user’s personality profile more reliably by fusing analytical predictions from multiple sources including behavioral traces, textual data, and social networking information. We applied and validated our approach using a real data set obtained from 1,040

World of Warcraft

players. Besides behavioral and social networking information, we found that text analysis of character names yields the strongest personality cues.

The long-term goal of our research is to design information visualization systems that adapt to the specific needs, characteristics, and context of each individual viewer. In order to successfully perform such adaptation, it is crucial to first identify characteristics that influence an individual user’s effectiveness, efficiency, and satisfaction with a particular information visualization type. In this paper, we present a study that focuses on investigating the impact of four user characteristics (perceptual speed, verbal working memory, visual working memory, and user expertise) on the effectiveness of two common data visualization techniques: bar graphs and radar graphs. Our results show that certain user characteristics do in fact have a significant effect on task efficiency, user preference, and ease of use. We conclude with a discussion of how our findings could be effectively used for an adaptive visualization system.

In recent years, there has been increased interest and research on identifying the various ways that students can deviate from expected or desired patterns while using educational software. This includes research on gaming the system, player transformation, haphazard inquiry, and failure to use key features of the learning system. Detection of these sorts of behaviors has helped researchers to better understand these behaviors, thus allowing software designers to develop interventions that can remediate them and/or reduce their negative impacts on user outcomes. In this paper, we present a first detector of what we term WTF (“Without Thinking Fastidiously”) behavior, based on data from the Phase Change microworld in the Science ASSISTments environment. In WTF behavior, the student is interacting with the software, but their actions appear to have no relationship to the intended learning task. We discuss the detector development process, validate the detectors with human labels of the behavior, and discuss implications for understanding how and why students conduct inquiry without thinking fastidiously while learning in science inquiry microworlds.

To keep a learner motivated, an intelligent tutoring system may need to adapt its feedback to the learner’s characteristics. We are particularly interested in adaptation of performance feedback to the learner’s personality. Following on from an earlier study that investigated the effect of generalized self-efficacy, this study examines how feedback may need to be adapted to the trait Conscientiousness from the Five Factor Model. We used a User-as-Wizard approach, with participants taking the role of the adaptive feedback generator. Participants were presented with a fictional student with a validated polarized level of Conscientiousness, along with a set of marks the student had achieved in a test. They provided feedback to the learner in the form of a short statement. We examined the level to which participants bent the truth as adaptation to the learner’s conscientiousness. The study suggests that adaptation to conscientiousness may be needed: using a positive slant for highly conscientious students with failing grades.

In this paper we describe an initial attempt to build multi-faceted user models from raw Twitter data. The key contribution is to describe a technique for categorising users and their social ties according to a collection of curated topical categories and in this way resolve much of the preference noise that is inherent within user conversations. We go on to analyse and evaluate this approach on a data set of over 240,000 Twitter users and discuss the applications of these novel user models.

User ratings are a valuable source of information for recommender systems: often, personalized suggestions are generated by predicting the user’s preference for an item, based on ratings users explicitly provided for other items. In past experiments that were carried out by us in the gastronomy domain, results showed that rating scales have their own “personality” exerting an influence on user ratings. In this paper, we aim at deepening our knowledge of the effect of rating scale personality on user ratings by taking into account new empirical settings and a different domain (a museum), and partially different rating scales. We compare the results of these new experiments with our previous ones. Our aim is to further validate in a different application context, and domain, and with different rating scales, the fact that rating scales have their own personality which affects users’ rating behavior.

Modeling learners is a fundamental part of intelligent tutoring systems. It allows tutors to provide personalized feedback and to assess the learners’ mastery over a task domain. One aspect often overlooked is the modeling of erroneous behaviors that can be used to provide error specific feedback. This is especially true for model-tracing tutors that usually require erroneous procedural knowledge associated to each of the possible error. This process can be automated thanks to a task independent model describing the learners’ erroneous behaviors. The model proposed in this paper is inspired by the Sierra theory of procedural error and is developed for ASTUS, an authoring framework for model-tracing tutors.

Touch is a new and significantly different method of interacting with a computer and it is being adapted at a rapidly increasing rate with the introduction of the tablet computer. We log the characteristics of a student’s touch interaction while solving math problems on a tablet. By correlating this data to high and low effort problem solving conditions we demonstrate the ability to predict student effort level. The technique is context free, thus can potentially be applied to any computer tablet application.

In recommendation systems a variation of the cold start problem is a situation where the target user has few-to-none item ratings belonging to the target domain (e.g., movies) to base recommendations on. One way to overcome this is by basing recommendations on items from different domains, for example recommending movies based on the target user’s book item ratings. This technique is called cross-domain recommendation. When basing recommendations on a source domain that is different from the target domain a question arises, from which domain should items be chosen? Is there a source domain that is a better predictor for each target domain? Do books better predict a users’ taste in movies or perhaps it’s their music preferences? In this study we present initial results of work in progress that ranks and maps between pairs of domains based on the ability to create recommendations in domain one using ratings of items from the other domain. The recommendations are made using cross domain collaborative filtering, and evaluated on the social networking profiles of 2148 users. Initial results show that information that is freely available in social networks can be used for cross domain recommendation and that there are differences between the source domains with respect to the quality of the recommendations.

The recent exponential growth in mobile applications and the growing reliance on and awareness of ‘user models’ by end-users have led to the need to rethink the functional and end-user requirements of existing user modelling systems. This paper has two goals. Firstly, leveraging a functioning user modelling ecosystem that provides anywhere and anytime access to desktop-, web-, and mobile- applications, this paper identifies a current opportunity (and need) to enhance user interaction with existing user modelling frameworks, by extending beyond the stereotypical cloud-based user modelling approach to encompass also a client-based service and an accompanying synchronisation module. Secondly, we draw on an analysis of previous work and a small user study, to establish the need for a user-centred design focus for user modelling frameworks. We also identify functionality that end-users (rather than developers) need and want from a user modelling ecosystem.

This paper introduces the idea of adaptive score reports that can be used to provide educational stakeholders with a personalized experience aimed at facilitating student understanding and use of assessment information. These reports can also provide additional learning opportunities for users based on assessment results. An interactive score report for students is used to illustrate opportunities for adaptation.

Matrix factorization is a general technique that can extract latent factors from data. Recent studies applied matrix factorization to the problem of establishing which skills are required by question items, and for assessing student skills mastery from student performance data. A number of generic algorithms, such as Non-negative Matrix Factorization and Tensor factorization, are used in these studies to perform the factorization, but few have looked at optimizing these algorithms to the specific characteristics of student performance data. In this thesis, we explore how one such characteristic can lead to better factorization: the fact that items are learnt in a constrained order and allow such inferences as if a difficult item is succeeded, an easier one should also be succeeded. In particular, we want to address this question: can a partial order knowledge structure (POKS) be used to guide matrix factorization algorithms and lead to faster or better solutions to latent skills modelling?

Current Game-based Learning (GBL) applications often lack features that have been commonplace in conventional e-learning. One of these is the ability to provide players with a personalised experience. The author’s dissertation aims at further establishing adaptivity in GBL as a practical feature through three contributions: An adaptive educational game that fulfills key demands for GBL; an architecture that adapts at runtime according to authored references; and a controlled trial to evaluate effects of adaptivity in a statistically sound way.

This paper aims to present our generic model for user assistance. The approach we propose allows a designer to describe the assistance requested for an application using a common formalism; this description is then used by a generic assistant. A set of epiphyte assistants is then involved to perform assistance actions specified in the assistance description made by the designer.

Recommender systems (RSs) are heavily used in e-commerce to provide users with high quality, personalized recommendations from a large number of choices. Collaborative filtering (CF) is a widely used technique to generate recommendations [1]. The main research problems we desire to address are the two severe issues that original CF inherently suffers from:

Data sparsity

arises from the phenomenon that users in general rate only a limited number of items;

Cold start

refers to the difficulty in bootstrapping the RSs for new users or new items.

The principle of CF is to aggregate the ratings of like-minded users. However, the reported matrix of user-item ratings is usually very sparse (up to 99%) due to users’ lack of knowledge or incentives to rate items. In addition, for the new users or new items, in general, they report or receive only a few or no ratings. Both issues will prevent the CF from providing effective recommendations, because users’ preference is hard to extract. Although many algorithms have been proposed to date, these issues have not been well-addressed yet.

Due to the open ended nature of the interaction with exploratory environments (EE) for learning, it is not trivial to add mechanisms for providing adaptive support to users. Our goal is to devise and evaluate a data mining approach for providing adaptive interventions that help users to achieve better task performance during the interaction with an EE.

Web Design has been evolving with Web-based systems becoming more complex and structured due to the delivery of personalised information adapted to end-users. Although information presented can be useful and well formatted, people have little mental workload available for dealing with unusable systems. Subjective mental workload assessments tools are usually adopted to measure the impact of Web-tasks upon end-users thanks to their ease of use and are aimed at supporting design practices. The Nasa Task Load Index subjective procedure has been taken as a reference technique for measuring mental workload, but it has a background in aircraft cockpits, supervisory and process control environments. We argue that the tool is not fully appropriate for dealing with Web-information tasks, characterised by a wide spectrum of contexts of use, cognitive factors and individual user differences such as skill, background, emotional state and motivation. Furthermore, in this model, inputs are averaged without considering their mutual interactions and relations. We propose to see human mental workload as non-monotonic concept and to model it via argumentation theory. The evaluation strategy includes coparisons with the NASA-TLX in terms of statistical correlation, sensitivity, diagnosticity, selectivity and reliability.

The PhD research presented in this paper addresses some of the problems involved in creating a context-aware personalized service. Our main interest is in the steps of defining, detecting, acquiring and using real and relevant context of users. Our goals are to: collect and publish a context-rich movie recommender database, add theoretical requirements for contextual information in existing definitions of context, develop a methodology for relevant-context detection and inspect the impact of relevant and irrelevant context on the rating prediction using the matrix-factorization algorithm. This paper presents the work done so far and future plans with open issues.

The creation of accurate user profiles of interest across heterogeneous websites is a fundamental step for personalisation, recommendations and analysis of social networks. The opportunities offered by the Web of Data and Semantic Web technologies introduce new interesting challenges. In particular, the main benefits for user profiling techniques are given by the extensive amount of already available and structured information and the solution to the “cold start” problem. On the other hand it is difficult to manage a massive “open corpus” such as the Web of Data and select only the relevant features and sources from an heterogeneous collection of datasets. Hence we propose semantic technologies for interlinking social websites and provenance management on the Web of Data to retrieve accurate information about data producers. The goal is to build comprehensive user profiles based on qualitative and quantitative measures about user activities across social sites.

In this paper, we present our vision and some initial experiments on how to anticipate significance, similarity or polarity of various types of (preferably implicit) user feedback and how to form individual user preference for recommendation. Throughout the corporate web, we can observe the same patterns or actions in user behavior (e.g. page-view, amount of scrolling, rating or purchasing). Recorded user behavior – user feedback – is often used as base for personalized recommendation, but the connection between the feedback and user preference is often unclear or noisy.

Our goal is to analyze user behavior in order to understand its relation to the user preference. We report on some initial experiments on a real-world e-commerce application. We describe our new models and methods how to combine various feedback types and how to learn user preferences.

In this paper, we propose an interoperable ubiquitous user model which illustrates different aspects of the individual’s interests, preferences and personality. It is constructed by mining socially enhanced online traces of the user and aggregating the partially obtained profiles. Those traces include actions performed and relationships established in the social web accounts in addition to the local machine traces such as bookmarks and web history. Moreover, we claim that mining the content in a context-aware approach and computing fuzziness values during the process results in a more reliable user profile.

The Web is an important resource for a programmer: as much as 20% of a programmer’s time is spent on the Web [2]. When a programmer searches for information on the Web, two distinct information needs arise depending on the programmer’s previous knowledge of a library’s Application Programming Interfaces (APIs):

learning

how to invoke a software library versus reminding the programmers themselves the details deemed not worth remembering [2].