Building a collective experience

International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.4, No.1, January 2014

BUILDING A COLLECTIVE-EXPERIENCE
ENGINE FOR EXPERIENCE-TRANSFER
AMONGST WEB USERS
Jeremy Hall and Yasushi Kiyoki
Graduate School of Media and Governance, Keio University 5322,
Endo, Fujisawa, Kanagawa, Japan 252-8520

ABSTRACT
This paper describes the Collective Experience Engine (CEE), a system for indexing ExperientialKnowledge about Web knowledge-sources (websites), and performing relative-experience calculations
between participants of the CEE. The CEE provides an in-browser interface to query the collective
experience of others participating in the CEE. This interface accepts a list of URLs, to which the CEE adds
additional information based on the Queryee's previously indexed Experiential-Knowledge. The core of the
CEE is its Experiential-Context Conversation (ECConversation) functionality, whereby an collection of a
person’s Web Experiential-Knowledge can be utilized to allow a real-world conversation-like exchange of
information to take place, including adjusting information-flow based on the Queryee's experiential
background and knowledge, and providing additional experientially-related knowledge integrated into the
answer from multiple selected 'experience donors'. A relative-experience calculation ensures that
information is retrieved only from relative-experts, to ensure sufficient additional information exists, but
that such information isn't too advanced for the Queryee to process. This paper gives an overview of the
CEE, and the underlying algorithms and data structures, and describes a system architecture and
implementation details.

KEYWORDS
Data Mining, Knowledge Representation, Collaboration, Web Application, Experiential Calculations

1. INTRODUCTION
The Web has had several evolutions of Knowledge-Expertise expression on the Web. Before the
Web itself, people used bulletin board systems to gather and trade knowledge. The Usenet
network was another such system. These systems and their successors such as Wikis, and Q&A
domain-specific sites such as Stackoverflow provide effective ways of users pooling knowledge
on specific, known subjects.
While current methods of finding information on the Web focus on ways to link keywords to
explanation or discussion, it is still up to the user to find many potentially disparate sources of
information, understand how they fit together, draw conclusions about the reliability and utility of
various disparate information sources, and to do so with the potential handicap of not knowing
the most efficient or proper keywords or wording for finding such sources.
DOI : 10.5121/ijdkp.2014.4101

1


This paper describes the Collective-Experience Engine (CEE), created to enable direct querying
and visualization of the collective and untapped Experiential-Knowledge stored in the brains of
all Web users (see Figure 1). Users of the CEE can learn from the experience of others, rather
than having to build this experience on their own via extensive searching and surveying of
websites. They can then make informed decisions about information sources on the Web, and
have confidence in those decisions by relying on the knowledge and experience of other people
participating in the system.

Figure 1. This figure gives an overview of the Collective Experience Engine, describing the data
flow and processes within the CEE. A participant in the CEE uses their Web browser both for
standard Web viewing, and for interacting with the CEE. While browsing, their interactions are
sent to the CEE to be encoded as Experiential-Knowledge. When a Queryee submits a query, a
special interface is used to select and then submit URLs describing the type of knowledge-sources
they wish to learn about. The CEE inspects the Queryee's Experiential-Knowledge, and combines
it with the explicitly selected URLs. Next, several 'Experiential-Context Conversations' occur
between data-representations of the Queryee, and potential-donors of Experiential-Knowledge,
who themselves are merely participants in the CEE. Once experience-donors have been selected,
their Contextual-Answers are generated, which includes personal Experiential-Knowledge from
2


each donor, which was calculated to be related to the Queryee's original Contextual-Question.
Finally the Contextual-Answers are returned as a custom visualization, presented to the Queryee
in their Web browser.
The main feature of the CEE is the concept of Experiential-Context Conversations
(ECConversations), whereby experiential-context negotiation occurs between proxies of people,
rather than requiring explicit communications. An ECConversation is a model of a real-world
interaction between people, where both communicating parties adjust the content of their
communication as a function of their own, and the other person's experience and knowledge
regarding the subject of their conversation. By modelling real-world Experiential-Knowledge
transfer, we can automate and scale it to allow participants in the CEE to share their own and
utilize others' Experiential-Knowledge.
Within the CEE, we explicitly define a model of individual experience, which is lacking in
current knowledge discovery Web systems. With such a model, we can develop new visualization
and computational functionality for creating new query methods. The CEE is functionally an
overlay on top of existing Web-based knowledge-retrieval systems. The CEE reveals the utility
and the context in which websites are useful, from the point of view of several people
simultaneously. This allows people searching for information on an unfamiliar topic to make a
decision about the quality and reliability of unfamiliar websites that they are viewing, without
having to survey huge numbers of websites, or become experts in a topic themselves.
The CEE's ECConversation-algorithms determine how to combine the Experiential-Knowledge
of each user. A query is a combination of a person's explicitly-selected URLs for describing the
type of knowledge-sources they wish to understand, and a contextually-related portion of that
person's captured Experiential-Knowledge. A person's Experiential-Knowledge is based on
compiled Web-browsing behavior (not just viewed content, but actions in the browser itself such
as tab-switching), and encodes the unique experience-signature of a person on the Web. The
algorithms determine the appropriateness of which Experiential-Knowledge to include in the
initial query, as not all experientially-related knowledge may be appropriate given a person's
explicitly-selected URLs. The algorithms also determine who to select as ExperientialKnowledge donors from the rest of the participants in the system, and which additional
Experiential-Knowledge from those donors is appropriate to include in the final results. Given
two people that have chosen an identical set of URLs to submit as part of the initial query, due to
the different experiences of those two people, and how the delta in Experiential-Knowledge
affects the discovery of donors and which of those donors' Experiential-Knowledge to include,
the results returned to two people who select identical initial query parameters are going to differ.
This paper gives an overview of the query process, and describes the ECConversation-algorithms
in detail. A system architecture and prototype implementation is proposed, and the resultsvisualization process is described.

2. BACKGROUND: EVALUATING SOURCES OF KNOWLEDGE ON THE
WEB
The ability to evaluate the reliability and potential utility of a website is greatly impacted by
domain-knowledge regarding the website's topic. Low-knowledge browsers of websites depend
on surface cues such as typology keywords highlighted by web-search engines, and thus fail to
3


interpret when a website is only tangentially related to their target knowledge [1]. People familiar
with a topic use metacognitive functions to evaluate knowledge-sources. This means that in
addition to the content and semantics of a website, other information such as the author, date, and
the document type of the information is evaluated as well. Finally, past experience aids greatly in
rapidly evaluating new knowledge-sources, as it enables people 'easily link prior knowledge to
task requirements and to information found on the Web' [2]. Research into the difference in Webresearch techniques between novices and experts shows that experts '[...]more often activate their
prior knowledge[...]', and yet they '[...]show little differences in the way they search the internet...'
[3]. This indicates that by providing people with more prior knowledge and understanding of the
framework in which an expert evaluates their knowledge, we could rapidly improve the Websearch process for users unfamiliar with a topic they are searching.
P. Gerjets et al explains how searching on the Web has supplanted interaction with experts--for
instance when diagnosing computer problems or seeking medical advice. The environment of the
Web contains a large variety of complex information-domains, however the variability in terms of
the quality and reliability of the information is substantial, possibly owning to a mix of both
experts and laypeople providing information. 'As a result, Web users are required to
appropriately evaluate diverse, potentially diffuse, or even contradictory sources of
information[...]Web search is often related to personal decisions under uncertainty in domains
characterized by fragile and conflicting evidence[...]it has been shown that searchers usually face
difficulties in appropriately evaluating information during Web search' [4]. In order to overcome
such search obstacles, users must put in a large amount of manual effort to understand and
integrate knowledge about a search-topic, and perform a survey of websites to understand
information variability and reliability.
In H. v. Oostendorp & S. R. Goldman, the importance of the formation of a mental 'Document
Model', a mental model of the relationships between content and its sources, is described [5].
According to Bhavnani et al. [6], even the most common of topics on the Web is a 'fairly complex
task', where people '[...]must first visit more than one general page to get an overview[...]and then
visit specialized pages to get an in-depth understanding about specific concepts[...]'. The paper
goes on to mention that because such '[...]search procedures are similar to what search experts
have been observed to use, and because they are difficult to acquire just from using search
engines like Google [...] motivate the design of new approaches to search systems that explicitly
provide such guidance'.
In Britt and Aglinskas [7], they describe how 'Experts attend to many features of sources, and
some, such as style, may be too subtle [for novices]'. Also, the lack of 'gatekeepers of credibility,
such as editors and publishers' are placing greater emphasis on people's need 'to filter and
evaluate information sources'. Experts are better able to deal with bias, as they use multiple
criteria when evaluating sources, and not just base evaluations on content and the amount of
information provided [8]. 'The fact that the amount of immediately available information is nearly
unlimited on the WWW underlines the need for a reasonable selection of information[...]
laypersons need to activate prior knowledge in order to integrate information from multiple texts
and thereby build semantic connections between information from different sources. Finally, to
gain knowledge about the sources, laypersons have to evaluate sources in terms of quality and
credibility. This involves finding out about the author as well as his or her credentials, intentions,
possible affiliations, and sponsors.' Bråten [9] also describes how experts overcome bias better
than novices, especially when the content is written to be more easily understood, and thus gains
a greater value than is necessarily warranted in the novice's eyes.
4


In developing the CEE, we have created a system to allow novices and people unfamiliar with a
topic to understand how a relative-expert would collect information on said topic. Even if the
novice doesn't understand the reasoning behind a relative-expert's choices, they get the benefit of
the extra semantic, source, and other meta-cognitive knowledge that an expert has used to create
their own mental 'Document Model' of related knowledge-sources. Not only should utilization of
the CEE allow relative-novices to immediately gain confidence and the use of better-sources of
knowledge, but by providing an overview of good usage of knowledge-sources for a given topic,
the CEE gives people the ability to more accurately identify the reliability and usefulness of a
website on their own.

3. OVERVIEW OF THE COLLECTIVE-EXPERIENCE ENGINE
Performing a query with the Collective-Experience Engine is an easy process, whereby a person
defines a list of URLs to describe the type of information-sources they wish to better understand.
Selecting URLs is made easy with an in-browser interface we've developed, where people can
perform the add function from within a web browser window just by right-clicking on a page, or a
link to a page whose URL they wish to add.
The CEE passively-collects Experiential-Knowledge, extracting the relations among URLs that
that person has previously visited, and their utility to the user, in order to build a body of
knowledge to use on behalf of the CEE participant. This passive collection occurs as a person
utilizes a Web browser with a special browser plug-in installed.
When a query is submitted, the CEE first constructs a Contextual-Question, which is the
Queryee's contribution to an ECConversation. The Contextual-Question not only encodes the
initially selected list of URLs, but also includes contextually-relevant Experiential-Knowledge
(see Figure 2). URLs based on relevant Experiential-Knowledge are chosen based on two factors:
•

An experiential-relationship existing in the person's encoded Experiential-Knowledge
from or to an explicitly selected URL.

•

An appropriate relatedness-factor, calculated by comparing the relatedness of the
explicitly-selected URLs to each other, and then to each experientially-related URL. This
is meant to prevent contextually-unrelated experiential-relations from being included in
the final Contextual-Question.

5

ining

Figure 2. This figure describes in detail the process of formulating a Contextual-Question. The
Question.
Contextual-Question is created as part of the query process, and is used to find potential
query-process,
experience-donors, and then to describe the type of Experiential Knowledge that should be
donors,
Experiential-Knowledge
donated to the Queryee. First, the Queryee explicitly selects URLs, which describe the type of
explicitly
knowledge-sources the Queryee wishes to learn more about. The CEE then extracts important
sources
keywords from each URL's website, calculates the variance pattern of the URLs' keywords, and
then selects experientially-related URLs whose keywords are relevant within the same tolerance
related
as the explicitly selected URLs'. The combination of explicitly selected, and contextually
explicitly-selected,
experientially-related URLs make up a Contextual
s
Contextual-Question

6

ining

Figure 3. This figure describes the process of using the Contextual Question to identify
Contextual-Question
experiential-donor candidates. Any participant in the CEE is a potential donor of Experiential
donor
ExperientialKnowledge. A Contextual-Question first has the utility of each of its included URLs calculated,
Question
included
and the variance calculated from the resulting utility vector. URLs whose utility is larger than a
utility-vector.
calculated variance-range are used as the Core Context. All participants within the CEE whose
range
Core-Context.
Experiential-Knowledge contains the entire set of URLs within the calculated Core-Context
Knowledge
Core
become an potential experience-donor.
Next, Experiential-Knowledge donor candidates are selected from within the CEE in a process
Knowledge donor-candidates
called 'Donor Discovery' (see Figure 3). Each person is checked regarding whether they have
experience related to the URLs contained in the Contextual Question. People that are found to be
Contextual-Question.
experiential-matches then form the coun
matches
counterpart to the Contextual-Question, which we call a
Question,
Contextual-Answer. This answer not only contains knowledge about the URLs contained in the
Answer.
Contextual-Question, but also about that person's experientially related URLs that are checked to
Question,
experientially-related
be appropriately contextually-related to the URLs listed in the Contextual
related
Contextual-Question.

7

ining

Figure 4. This figure describes the concept of Experiential Context Conversations or
.
Experiential-Context
ECConversations, which allow a context
context-based conversation between two people to take place
using data-representations of those people in place of actual explicit conversation. Similar to how
representations
a real conversation would take place, the background experience and knowledge of the relative
relativenovice are included in the question, and then the answer is tailored to the questions context, while
still providing additional experientially
experientially-related knowledge from the experience-donor
donor.
By encoding and adjusting for experience in each query, and for each potential match, a two
two-way
Experiential-Context Conversation is taking place, similar to how a conversation between two
onversation
people speaking to each other occurs. This is the reasoning behind the new keyword
'ECConversation' (see Figure 4).
).

8

ining

Figure 5. This figure describes the Contextual Result Space's (CRS) interface, whose main
purpose is the visualization of the Contextual
Contextual-Question and Contextual-Answers. The CRS uses
Answers.
circles to represent websites, and assigns them size, mass, and attractive forces to convey their
relationships, and their relative
relative-utility from their owners’ point of view. Each person whose
knowledge is represented in the CRS is assigned a different color, and thus the CRS provides a
different
means to not only understand the value of knowledge sources from others' points of view, but also
knowledge-sources
to compare and contrasts those points of view.
As part of finding the appropriate people from which to retrieve Exper
Experiential-Knowledge as query
Knowledge
results, relative-experience is calculated. This calculation is performed between a Contextual
experience
ContextualQuestion, and a Contextual-Answer, which means only experience related to the relevant
Answer,
knowledge-domain is considered. In our model, usage of a URL equates to experience at that
domain
usage
URL. Only people whose experience falls within a calculated range have their Contextual
ContextualAnswers included in the results. Finally, a ranking function which calculates how close the ideal
mix of experience at the desired information
esired information-sources, and relative-experience is performed, and
experience
the top few Contextual-Answers are returned. The need for a relative expertise calculation is
Answers
relative-expertise
based on two premises:

9


•

Experiential-Knowledge from people with similar or less expertise than the person
performing the query is less likely provide usable and trustworthy insight when compared
to a person with relative-expertise.

•

Experiential-Knowledge from people with too much expertise may contain difficult-tounderstand or difficult-to-utilize insights, due to the large gap in experience.

The results are returned in the form of a custom visualization we've created to express
Experiential-Knowledge, called the Contextual Result Space (CRS). The CRS is an interactive
visualization, which can express many simultaneous points of view, and allows for intuitive
understanding of how those points of view relate amongst each other, including that of the person
who performed the query. The goal of the results, and the subsequent visualization isn't to make a
ranked list of 'best' URLs for the user, but to impart knowledge of how others use and value
URLs on the Web, so that the person performing the query can make an informed decision on the
value of that knowledge for themselves.

3.1. The Contextual Result Space
To express the complexity of an individual's Experiential-Knowledge, we have developed a
custom visualization based on the RDCS [10] we call a Contextual-Result Space or CRS. The
CRS represents Experiential-Knowledge of a website as a circle, adjusts the size (and
consequently mass) of the circle based on the utility of the website, and models links in the
Experiential-Knowledge graph as an attractive force between circles (see Figure 5). The final
visualization-result is a result of a dynamic simulation of the interactions of the circlesrepresenting-websites with each other.
The visualization receives Experiential-Knowledge as results from several donors, and combines
it into one visual representation which allows intuitive comparison of knowledge between the
Queryee's Contextual-Question, and the resulting Contextual-Answers. The CRS extends the
functionality of the RDCS to better fit the realities of the Collaborative Experience Engine.
The CRS creates a space into which the Contextual-Question and top-ranked Contextual-Answers
are combined. The basic rules which differ from the RDCSS are as follows:
•

Each circle represents the Experiential-Knowledge of a person at a given URL, with
respect to the other URLs present within the CRS. This means that any compiled
experience which references a website that isn't also being represented in the CRS isn't
counted. In this way, the utility of a website is properly conveyed within the context of
the CRS being visualized.

•

A given website can only be represented by one person's experience. This to keep the
visualization clear and intuitive. If data from two or more people include experience
about the same website, then the CRS chooses only one based on the ranking of the
results, with the Queryee's Contextual-Question data being ranked the highest, and then
the donors' Contextual-Answer data following in ranked order. This gives the most
experientially-related knowledge the highest priority for knowledge-expression within the
visualization.

10


•

By default, the visualization of the Queryee's Contextual-Question data is also included,
and takes the highest priority. This is so that the other Experiential-Knowledge can be
understood in terms of it relating to the Queryee's experience. However, due to the
priority rule, the visualization won't show any of the experts' experience for the websites
that are already known to the Queryee. For this reason, there is a toggle that tells the CRS
to remove the Contextual-Question experience from the visualization, so that the relativeexperts' experience can be more completely visualized.

The visualization allows the Queryee to understand the importance and usability of knowledgesources related to their query, from the point of view of relative experts.

4. CEE USAGE SCOPE
The Collective Experience Engine is a tool for learning about new or unfamiliar topics, and for
when one must rely on new or unfamiliar data-sources to learn about those topics. It can be used
in conjunction with any other tool or resource which provides information-sources-as-results. The
most common of these is probably keyword-based search engines, however social networks,
blogs, and expert-knowledge websites are all good candidates for being used in conjunction with
the CEE, as they often contain references to additional potentially-useful knowledge-sources.
The CEE relies on the ability to capture a large portion of a person's experience in order to be
effective. The Web is a good candidate for the CEE because all experiences between a person and
the Web can be captured in that person's Web browser. Also, the list of all possible experiences
dealing with Websites is limited to a few actions in the browser, such as loading a website,
switching tabs, moving the mouse, scrolling, or spending time reading Web content.

5. ALGORITHMS AND THE DATA MODEL
This section describes the basic data structures and algorithms used within the CEE. A weighteddirected graph is used as a model for a person's Experiential-Knowledge. The primary algorithms
include:
•

A Contextual-Question formulator, which takes a query as input, and generates a
Contextual-Question vector on-the-fly for searching against other peoples' ExperientialKnowledge.

•

A Relative-Experience extraction and scaling algorithm. This algorithm allows for
experience with respect to a given Knowledge-Domain to be calculated, and utilized as a
query parameter. This works by allowing Contextual-Question vectors to be searched
effectively, despite the original data having differing values due to relative differences in
experience.

•

The Contextual Result Space visualization algorithm, which allows for the comparison
and contrast of a number of people's relevant Experiential-Knowledge simultaneously.
This algorithm regulates the conversion of graph-data into placement, behavior, and
visual aspects within the visualization that is utilized as the query results.

11


5.1. Modeling Experiential-Knowledge
We use an Experiential-Knowledge weighted-directed graph (graphEK) to model per-user Web
browsing activity within the CEE. Nodes in the graph represent a URL, often a Website (To
denote a specific person's graphEK, we utilize the notation graphEK-iduser). The following
information is represented in the graph:
5.1.2. URL Experience
All URLs experienced by a person are represented by the nodes within a person's graphEK. The
labeling function for the nodes in the graph uses a idurl and an iduser as inputs nodelabel =
flabel(idurl,iduser). This means that while each user has their own independent nodes for the same
URL, and thus conceptually each user has an independent graphEK, all users' nodes have a
globally unique identifier, and can be easily stored within a single namespace. Despite having
globally unique labels, users' graphs can be easily merged for visualization purposes due to the
use of common idurl ids.
5.1.3. URL View Transition
When a person changes the URL they are viewing--for instance when switching browser tabs, or
opening a new window--a directed link is used to represent this activity. For instance, viewing
URLa, then URLb results in the link linkab being created. Thus, link.a is the referrer or origin of a
transition, and link.b is the target or endpoint of an activity.
5.1.4. URL Knowledge
A person's Knowledge of a URL is stored in the adjacency graph of nodelabel , which we call
adjlabel. The directed links stored in adjlabel all include the time they were created (link.then).
5.1.5. URL Utility
A simple explanation of the Utility of a nodelabel (utilitylabel) is that it is represented by its degree
(degreelabel) multiplied by the combined weight of the node's adjacency graph adjlabel.

As a practical matter, the URL Utility isn't such a straightforward calculation. Firstly, the weight
of a link isn't a simple property, but rather an aggregation function which we must calculate. We
must also factor in both the time-decay of the links in adjlabel, and the Knowledge-Domain that is
being used as a context with which to understand the Experiential-Knowledge at a URL. A key
for a link in adjlabel is an {a,b} pair, and a then property, which defines when the link was created
in the graphEK. When factoring in time-decay, we use the following process to determine a decayfactor:

12


Equation 2 describes a function which will take a then in UTC milliseconds, and return a
weighting factor that will have no effect if then is the current time NOW, and will drop to
approximately 0.067 in a year’s time. We divide by 1000 in order to convert the units to seconds,
and thus make the final units more consistent on a human scale. This equation favors recent
knowledge more strongly than past knowledge. In the future, giving the user the ability to adjust
the time-decay factor may be beneficial.
The utility calculation is affected by a selected Knowledge-Domain because the purpose of the
Knowledge-Domain is to only allow Experiential-Knowledge relevant to the context of the query
to be represented. A node in the graphEK might have many links in its adjacency graph that are
irrelevant to the query at hand. To factor in the Knowledge-Domain, we simply perform an
intersection between the set of idurl ids in the Knowledge-Domain, and the set of ݈݅݊݇௔௕ ∈
݆ܽ݀௟௔௕௘௟ , using b to represent each link as we see in Equation 3:

The validLinks set contains linkab links with a link.a and a link.b property where b is the target,
and a is the referrer of a relationship between two nodelabel nodes.
The final utillabel value with time-decay and Knowledge-Domain filtering is shown in Equation 4:

5.2. ECConversation Algorithms
This section describes the algorithms necessary for the ECConversation functionality of the CEE.
This includes how to formulate a Contextual-Question, using the Contextual-Question to find
experience-donors for having an ECConversation with, and then how those donors generate a
Contextual-Answer as a response.
The ECConversation Algorithms use two main factors in generating the data structure for
querying: keyword relatedness, and utility-spread. The algorithm ensures that items extracted
from a person's graphEK are related in terms of content to the explicitly selected query URLs'
(idquery) ids by comparing the important keywords from those URLs' content. Later, ContextualQuestion data structures (vecCQ) are compared using fuzzy matching, by understanding the
variance of utilities within the selected URLs, and using that variance as a range on a per- idurl id
basis.
When a person submits a query, the system utilizes that person's graphEK to generate a
Contextual-Question. A Contextual-Question is the data-structure used to ask the question 'What
do other people know about information-sources regarding type-X, which are related to my
previous experience?', and is used compare and search Experiential-Knowledge within the CEE.
5.2.1. Contextual-Question Formulation
The first thing that happens after a person submits a query, is the potential surrounding idurl ids
are gathered (idpotentialSur). The idpotentialSur ids are selected by being experientially-related to the
idquery ids, within the graphEK:
13


In order to understand whether a member of idpotentialSur is contextually-related enough to the idquery
id set, we must first calculate the mean and σ of the idquery URLs' keyword vectors. We use
important keywords extracted from the source URLs represented by the idurl ids stored in idquery to
do this. There are several potential algorithms with which to extract and calculate important
keywords from a URL. We chose to use a modified form of tf-idf (term frequency-inverse
document frequency). The function to extract a vector of scored keywords from a give URL with
respect to a document set documentset is ݂௞௘௬௦௖௢௥௘ (݅݀௨௥௟ , ݀‫ݐ݊݁݉ݑܿ݋‬௦௘௧ ). The documentset is a set
of ids, either idpotentialSur, idquery, or a combination of the two. This function returns a vector whose
keys are keywords, and whose values are the scores of those keywords.
First we create a documentSMvec vector, containing the mean of all the vectors in the idquery ids
documentset.

Next, the mean and σ distance for individual ݂௞௘௬௦௖௢௥௘ vectors from the idquery and the mean
documentSMvec vector. The function ݂ௗ௜௦௧ is a standard distance formula calculation for measuring
the distance between two vectors:

Now that both the meanquery and σquery distance values are calculated, we can select members from
idpotentialSur that are within ±1.5 σquery from meanquery, and combine the resulting idsur set with idquery
to get our vecCQ:

Now that we know which idurl ids the Contextual-Question is composed of, we can calculate the
utilization statistics, so that we can do fuzzy matching when querying other people's ExperientialData. We calculate the mean Contextual-Question idurl utilization meanCQUtil and its standard
deviation (ߪCQUtil):

Now that we know which idurl ids the Contextual-Question is composed of, we can calculate the
utilization statistics, so that we can do fuzzy matching when querying other people's Experiential14


Data. We calculate the mean Contextual-Question idurl utilization meanCQUtil and its standard
deviation (ߪCQUtil):

The final Contextual-Question contains vecCQ, meanCQUtil, and ߪCQUtil.
5.2.2. Experience-Donor Search
Once we've generated the Contextual-Question, we use it to discover experiential-donors. As
there is no keyword-based data remaining in the Contextual-Question, this search is done
completely based on Experiential-Knowledge of URLs. The initial operation is simple--we form a
subset of vecCQ, based on subtracting 1.5 sigmas of ߪ஼ொ௎௧௜௟ from each key's value in vecCQ. This
removes any idurl ids that aren't absolutely required in order to describe the Contextual-Question,
due to having a relatively-low utilization, and creates a base-requirement set (vecCQBase) for
matching the person's query:

Using the vecCQBase set, we can then select the set of people within the CEE who are potential
donor-candidates for providing Experiential-Knowledge to the person performing the query.
Assume the CEE (veccollective) is a set of ݃‫݌ܽݎ‬ℎா௄ି௜ௗೠೞ೐ೝ . In order to find the set of initial donorcandidates (vecinit):

5.2.4. Ranking and Result Limits and the Relative-Experience Algorithm
With too many results, the visualization would quickly become overwhelming. This means that
the ability to rank, and then only select a limited number of experiential-donors' contributions to
the results is important. The ranking algorithm sorts based on utilization-knowledge to create a
vecrank containing iduser keys and a utilization-score value where the greater the utilization-score,
the greater the rank. Remember that vecCQMatches contains graphEK graph members:

15


In the case of performing a relative-experience calculation, we want the relative-ideal utility value
to be based on a utility value an order of magnitude greater than the user, rather than the user’s
raw utility value. Thus, Equation 16 must be altered if such a calculation is desired:

5.2.5. Generating Results
The purpose of the results is to describe to a person how other people utilize information-sources
within the Contextual-Query's context. This means that we wish to return not just the matches to
the Contextual-Question, but also the surrounding Experiential-Knowledge from the experiencedonors' graphEK. To do this, we perform identical operations to when the Contextual-Question
was originally generated. The main differences are that instead of the idquery ids set, we instead
substitute in the vecCQ set, and of course the iduser user id is that of the donor's graphEK. The
calculations can stop after Equation 10 is complete, and the resulting set is called idQResult instead
of vecCQ. These idQResult sets of idurl ids are then returned to the person who performed the query,
and visualized.

6. SYSTEM ARCHITECTURE AND IMPLEMENTATION
Our software architecture is client-server. The client is a Chrome browser plug-in, which is able
to capture user-actions as they browse the web. These actions are fed to a server which compiles
them into an Experiential-Knowledge graph. The data structures are represented in a NoSQL keyvalue based server called Redis. Communications are handled via a synchronization-based
protocol provided by Firebase. The visualization is provided by a custom canvas-based HTML5
web application.
There are two main roles of the system:
•
•

Passively mining web-behavior data, and converting it into the Experiential-Knowledge
graph.
Providing query functionality.

6.1. Web-Behavior Mining
While there are many web-behavior possibilities to capture and encode as ExperientialKnowledge data, our system focuses on loading URLs, and switching between tabs in the web
browser. As the user performs these actions, the client plug-in captures them, and sends a
message to the server with two pieces of information: the target URL, and the referrer URL (if
there was one). For instance, switching between tabs causes the URL which was loaded in the
previously-active tab to be the referrer, and the newly-activated tab's URL to be the target. The
server then adds the Experiential-Knowledge to the person's graph.
As mentioned before, the Redis database uses a key-value system to store data, and provides data
types such as strings, lists and sets. A node in the Collective-Brain is stored as two sorted-sets
(for storing the links in the graph) and a per- idurl id sorted-set for storing the list of keywords and
their scores for each URL represented in a person's graphEK. For the sorted-sets representing the
links, the key for a sorted-set is a concatenation of iduser,idurl[in|out]. In the implementation, links
16


are broken into incoming and outgoing lists, hence the "in" or "out" at the end of each node. The
score of each entry in the sorted-sets is the time, and the value is an idurl id. Because there is a full
set of incoming and outgoing data for each idurl, there are no dependency-lookups or chains, and
thus data-retrieval is quick.
As an optimization, to allow for quicker matching amongst people, we also have a set that stores
all of the idurl ids that are represented within a person's graphEK. This allows us to perform a test
to quickly determine if a given user has the entire set of core-URLs from the ContextualQuestion, to qualify them as a potential-donor of experiential-knowledge for a given query.
6.2. Query Functionality
The Query interface and visualization is provided by a browser plug-in. Hooks in the browser's
API allow us to modify the GUI to provide right-click menu options for adding URLs to a query.
A button in the browser's chrome provides a person with the ability to see all of the currentlyadded URLs in a visual manner, and then submit the query. Results are visualized within the
browser itself using a canvas-based HTML5 render, the same as the RDCS [10].

7. FUTURE WORK AND CONCLUSION
In the future, we plan on capturing more behavioral aspects of a user, including the amount of
time spent viewing content, and capturing content embedded in a webpage. By saving each
element embedded within a webpage individually, we can more accurately capture the
experiential knowledge of a person, because we can discover relationships wherein the same
content is embedded in multiple different webpages. Finally, the ability parse media in addition to
text would further improve the accuracy of the CEE.
Another future improvement is the addition of query modifiers. For instance, allowing the
Queryee to specify from what type of people (friends, experts, groups, a specific person) to query
information. Perhaps a group would like to curate their own Experiential-Knowledge, or the
Queryee would just like to ask friends since those people are considered more trustworthy.
Another possible query modifier is the range of acceptable relative-experience.
Finally, a continuously working version of the CEE, where the Contextual Result Space updates
as a user browses might provide a better experience, as the feedback would be quicker, and more
tied to the browsing experience.
We have designed a system called the Collective Experience Engine which captures and
distributes the experience of participants in the CEE for the collective benefit of the whole. The
CEE overcomes the problem of insufficient knowledge about a topic causing uncertainty and
poor choice when selecting knowledge-sources to learn from. The query process is designed to be
easy and intuitive, automatically tailoring the query to the knowledge and experiential
background of the person who submitted it. The use of Experiential-Context Conversations
ensures that the content of the results are properly tailored to the context of the person submitting
the query, and the relative-experience calculation ensure that the results are properly tuned to
their experience level. The CEE targets a definite need on the Web for better dissemination of
understanding of various knowledge-sources, and the ability to understand such sources from the
point of view of people with more experience than oneself.
17


REFERENCES
[1]

Susan R. Goldman, "Choosing and using multiple information sources: Some new findings and
emergent issues ," Learning and Instruction , vol. 21, no. 2, pp. 238-242, 2011, Special Section I:
Solving information-based problems: Evaluating sources and information, Special Section II:
Stretching the limits in help-seeking research: Theoretical, methodological, and technological
advances. [Online]. http://www.sciencedirect.com/science/article/pii/S0959475210000204
[2 ]Saskia Brand-Gruwel and Marc Stadtler, "Solving information-based problems: Evaluating sources
and information ," Learning and Instruction , vol. 21, no. 2, pp. 175-179, 2011, Special Section I:
[3] Saskia Brand-Gruwel, Iwan Wopereis, and Yvonne Vermetten, "Information problem solving by
experts and novices: analysis of a complex cognitive skill," Computers in Human Behavior, vol. 21,
no. 3, pp. 487-508, 2005.
[4] Peter Gerjets, Yvonne Kammerer, and Benita Werner, "Measuring spontaneous and instructed
evaluation processes during Web search: Integrating concurrent thinking-aloud protocols and eyetracking data ," Learning and Instruction , vol. 21, no. 2, pp. 220-231, 2011, Special Section I:
[5] Charles A Perfetti, Jean-François Rouet, and M Anne Britt, "Toward a theory of documents
representation," The construction of mental representations during reading, pp. 99-122, 1999.
[6] Suresh K and Jacob, Renju T and Nardine, Jennifer and Peck, Bhavnani, "Exploring the distribution
of online healthcare information," in CHI'03 Extended Abstracts on Human Factors in Computing
Systems.: ACM, 2003, pp. 816-817.
[7] M. Anne Britt and Cindy Aglinskas, "Improving Students’Ability to Identify and Use Source
Information," COGNITION AND INSTRUCTION, vol. 20, pp. 458-522.
[8] Marc Stadtler and Rainer Bromme, "Dealing with multiple documents on the WWW: The role of
metacognition in the formation of documents models," International Journal of Computer-Supported
Collaborative Learning, vol. 2, no. 2-3, pp. 191-210, 2007.
[9] Ivar Bråten and Helge I. Strømsø and Ladislao Ladislao, "Trust and mistrust when students read
multiple information sources," Learning and Instruction, vol. 21, no. 2, pp. 180-192, 2011.
[10] Jeremy Hall and Yasushi Kiyoki, "Creating a Personal-Context Oriented Real-Time Dynamic and
Collaborative Space," Information Modeling and Knowledge Bases Xxiv, p. 82, 2013.

18

Building a collective experience

More Related Content

What's hot (19)

Viewers also liked (7)

Similar to Building a collective experience (20)

Recently uploaded (20)

Building a collective experience