=Paper=
{{Paper
|id=Vol-1388/DeCat2015-paper1
|storemode=property
|title=Automatic Task-Cluster Generation based on Document Switching and Revisitation
|pdfUrl=https://ceur-ws.org/Vol-1388/DeCat2015-paper1.pdf
|volume=Vol-1388
|dblpUrl=https://dblp.org/rec/conf/um/AbelaSH15
}}
==Automatic Task-Cluster Generation based on Document Switching and Revisitation==
https://ceur-ws.org/Vol-1388/DeCat2015-paper1.pdf
Automatic Task-Cluster Generation based on
Document Switching and Revisitation
Charlie Abela1 , Chris Staff1 , and Siegfried Handschuh2
1
Department of Intelligent Computer Systems,
University of Malta, Malta
{charlie.abela,chris.staff}@um.edu.mt
2
Department of Computer Science and Mathematics,
University of Passau, Bavaria, Germany
{siegfried.handschuh}@deri.org
Abstract. Personal Information Management (PIM) research is chal-
lenging primarily due to the inherent nature of PIM. Studies have shown
that people often adopt their own schemes when organising their per-
sonal collections, possibly because PIM tool-support is still lacking. In
this paper we investigate the problem of automatic organisation of per-
sonal information into task-clusters by transparently exploiting the user’s
behaviour while performing some tasks. We conduct a controlled experi-
ment, with 22 participants, using three different task-execution strategies
to gather clean data for our evaluation. We use our PiMx (PIM analytix)
framework to analyse this data and understand better the issues asso-
ciated with this problem. Based on this analysis, we then present the
incremental density-based clustering algorithm, iDeTaCt, that is able
to transparently generate task-clusters by exploiting document switch-
ing and revisitation. We evaluate the algorithm’s performance using the
collected datasets. The results obtained are very encouraging and merit
further investigation.
Key words: Density-Based Clustering, Personal Information Manage-
ment, Task Clusters
1 Introduction
When we are performing some task on our desktop, we tend to spend a consider-
able amount of time looking back, establishing past references and remembering
[10]. Whether performing the task requires us to search for information on the
Web, reply to some email we’ve received, or resume writing some other document
which we’ve worked on the day before, we tend to rely on our organisational skills
and the support of search, bookmarking and history tools [9].
The common feature in these tools is their ability to help us find or re-find
information by exploiting revisitation [10]. However, most of these tools tend to
consider the user’s information-seeking activities as unrelated events, unlike the
way we actually organise things, which is usually in terms of directories (on our
desktop) and tasks (conceptually) [11, 13]. Furthermore, humans are by nature
or by force multi-taskers, and tool-support should cater for situations whereby
users switch between one task and another or are interrupted [4, 11].
In this paper we align our research with efforts such as those of [12, 1] and
investigate how to transparently cluster documents such as Web-browsed doc-
uments, office-related documents and emails, which are viewed by the user and
belong to the same task.
As documents are presented to users in windows and tabs, and users switch
between them, we collect and exploit evidence of the users “window switching
behaviour” to identify and generate task-clusters3 . These clusters can be used
to re-find specific task-related documents and to resume a task, if or when, the
user is interrupted. Currently, we do not label the identified task-clusters and
we are not considering the content of the accessed documents as was the case in
[12].
We adopt an unsupervised method that treats the accesses to documents
as an undirected activity graph, Ga (V, E), based on which we create the task-
cluster’s graph Gc (V, E), which is both weighted and undirected. The edge weight
w(u, v) in Gc (V, E) reflects the strength of the association between two nodes u
and v. In the text we tend to use interchangeably the words “document” and
“node”, depending on the perspective we consider.
We performed a controlled experiment, conducted with 22 participants, using
3 different task-execution strategies to gather user’s window-switching behaviour
data for our evaluations. We separate the collected data into 3 groups depend-
ing on the task-execution strategy adopted: in succession with no interleaving
(used as baseline), interleaved and interleaved over different sessions. We also
developed the PIM analytix PiMx framework through which it was possible to
simulate, off-line, the task-execution over the collected data and at the same
time exploit network-analytics to analyse and visualise Ga (V, E) and Gc (V, E).
Through PiMx we were able to understand better which algorithmic approach
was more suitable.
In our approach we have factored-in an important feature, which to our
knowledge has not been addressed yet. We refer to the incremental nature of
the task and task-clusters, with nodes and edges being added over time as the
user visits and re-visits documents. We extend the work of [6], and propose an
incremental density-based clustering algorithm, which we call iDeTaCt that
identifies the dense regions from the less denser ones in the accessed documents’
space, identifying the task-clusters in the process.
We evaluated our approach to verify how well our algorithm is able to: (i)
identify those nodes that belong together in a task-cluster and (ii) identify when
a switch between two nodes is effectively a task-switch. The results are very
encouraging and iDeTaCt managed to cleanly separate all the tasks, using the
data of all the participants from the baseline group which performed the tasks
in succession. When we used the data with interleaved tasks, iDeTaCt was
3
A task-cluster is a group of documents that pertain to a task.
found to be sufficiently reliable in more than 50% of the cases, at the expense of
capturing less documents from the referenced tasks.
The rest of the paper is structured as follows. In Sec. 2 we present research
which is closely related to our own. We follow up with a detailed description of
the controlled experiment that we conducted to collect reliable data. In Sec. 4 we
give an overview of the PiMx framework which we use to analyse the collected
data. We introduce our iDeTaCt algorithm in Sec. 5, which is followed by the
evaluation and future work sections.
2 Related Work
In our work we draw parallels with research related to task-switching and identi-
fication, and others that have adopted the graph data structure as the underlying
representation for the user’s switching and revisitation behaviour.
In his thesis [10], Mayer presented an integrative history, visualization tool
entitled SessionGraphs. The tool allows a user to view her browsing activity as an
animated, interactive graph and to organise the visualisations according to her
tasks. We have adopted a similar graphical approach for our PiMx framework,
however the scope behind PiMx was that of allowing a researcher to better
understand the issues related to the problem of automatic task-cluster generation
rather than to provide browsing support.
The search bar presented in [11] persistently maintains a hierarchical Web
history organised around search topics and queries. It assists users in organising
complex searches and re-acquiring the context of a suspended search, based
mainly on topic and query-driven groupings. The multitasking bar presented
by [13] copes with both multiple tasks as well as multiple session tasks. They
considered a task as having different states and it was up to the user to maintain
and organise a task. Our approach aims to transparently automate the task-
cluster generation without requiring any user intervention.
A semi-automated approach to task-identification was adopted by [3] which
used activity-log analysis to group the accessed resources based on cues generated
by an individual while performing some task. We adopt this same approach to
collect the user’s activity information through dedicated application plug-ins.
We maintain a global desktop history with information about all the created,
accessed and edited documents which also includes Web-related documents and
queries.
To identify documents pertaining to a task [12] computed document con-
tent similarity and applied a maximal clique-finding algorithm over the user’s
switching activities. Unlike this approach, we currently do not intend to exploit
the content of the accessed documents, however we will consider the incremental
characteristics underlying an information-finding process.
In [1] a PageRank-like association heuristic was used to compute the asso-
ciation between windows opened on the desktop and presented a visualisation
through which windows that were frequently clicked in sequence, were displayed
closer together. However, no task-clusters were explicitly generated.
In [7, 8] an incremental density-based graph clustering approach is used to
cluster documents and find interesting subgraphs, respectively. Density-based
clustering is quite interesting since it is capable of coping effectively with noise.
In our case this will be a major challenge since users tend to constantly switch
between tasks, with the result that it would be more difficult to deal with those
documents that are accessed in between tasks.
3 Controlled Data Collection Experiment
Evaluating PIM related research is inherently difficult, in particular due to the
lack of readily available datasets. We therefore conducted a data-collection exper-
iment in a controlled environment, to collect clean data related to the window-
switching behaviour of the participants while performing some predefined tasks.
We set up a cluster of machines in one of our laboratories, each running Win-
dows OS and having two activity-monitoring applications installed on them. One
of the applications monitored browsing activity on Firefox4 while the other mon-
itored file browsing activity (e.g. of word processing documents) on the desktop.
The participants were advised about this monitoring and were assured that the
data would be anonymised and used only for the specified research purpose be-
fore the start of the experiment. This consisted of all the participants performing
the same three, predefined information-seeking tasks, by answering a number of
questions related to specific topics. Apart from seeking out information, partici-
pants had to compile a document with the relevant answers for each task, which
they had to email to us at specified intervals. Our methodology was in line with
that adopted by [11, 10].
The tasks required participants to provide specific information about the
planning of a vacation in a specific country; answering questions related to the
research area of human computation; and providing information about any two
upcoming music events. The tasks were conducted either over single or multiple
sessions. At pre-established intervals, unknown to the participants, we sent out
emails that either informing them what a task entails or else requesting that
they switch to another task.
There were in total 22 participants, 25% of whom were female. The partic-
ipants were students and members of staff (lecturing and administration) from
the Faculty of ICT within the University of Malta. The students were compen-
sated e 10 for their participation in the experiment.
The participants were split into three groups. Each group performed the
experiment separately from the others. The groups were split as follows:
i. Group 1 : the 7 participants in this group completed each of the three tasks
in sequence, starting with task 1, followed by tasks 2 and 3, without inter-
ruptions. We use the data from this group as the baseline for our algorithm,
since the tasks are clearly separated from each other;
4
https://www.mozilla.org/en-US/firefox/desktop/
ii. Group 2 : there were 10 participants in this group. They performed the three
tasks in a single session. They started working on task 1 but were interrupted
with an email from us requesting that they start task 2. After some more
time we sent another email requesting the participants to stop working on
task 2 and start working on task 3. We later interrupted them with yet
another email requesting that they switch back to task 2, finish it, and then
switch to, and complete tasks 1 and 3, in this order. In this way the tasks
were interleaved and thus identifying which documents pertained to which
task, becomes even more challenging;
iii. Group 3 : the 5 participants in this group performed the tasks in the same
order as Group 2 and with similar interruptions, however they were stopped
30 minutes into the session. Later on we asked them to continue the ex-
periment in another session, which took place some days later. During the
second session they had to resume the tasks and complete them in a sequen-
tial order with no further interruptions. In this way we wanted to introduce
some more challenges to the participants, since they had to remember what
they had been working on and recall the state of the task/s before they were
stopped.
Although we tried to have an equal number of participants in each group, due
to availability issues of our participants we had to somewhat relax this aspect.
Furthermore, the data collected from one participant from Group 1 and another
from Group 2 was unusable due to issues with the data-logging applications,
which were unfortunately, not noticed in time.
The logged data included information about the type of event (e.g. naviga-
tional and tabbed events), the application that generated the event, the times-
tamp, the URL of the document accessed as a result of the event, an excerpt of
text from the window caption. Other information, specific to particular events
was also captured. This included, the URL of the page that was in focus be-
fore the event was triggered, as is the case of the navigational events. We also
captured the file name and whether a document was edited or not, in the case
of the desktop’s file-related events. The data logged from each participant was
anonymised and cleaned for further processing.
4 PiMx: tool for analysing the data
We implemented a tool, called PiMx (Personal information Management ana-
lytix) to analyse the collected data and understand better how we can algorith-
mically exploit document switching and revisitation to generate the task-clusters.
PiMx allows a researcher to load a user’s activity-log and to simulate the
execution of the task-trail for that user. This process can be paused and resumed
at any time, allowing the researcher to analyse and compare the evolving task-
trail through different views, see Fig. 1. The PiMx-History is similar to the tool
developed by [5] and allowed us to view details related to all accessed documents,
including the URI and amount of revisitations. It is also possible to filter the
PiMx
Viz
PiMx
History
PiMx
Stats
PiMx
Clusters
Fig. 1: PiMx Interface
data by different time windows (e.g. last hour, last 4 hours, today, yesterday
etc.), as well as by application and file-type.
The PiMx-Viz is an interactive component inspired by the SessionGraph
described in [10]. This provides visualisations of the unadulterated activity-graph
as it evolves over time and the task-clusters as they are generated by the applied
algorithm. The size of the nodes is relative to the number of accesses and it
is also possible to click on each node separately and to visualise the induced
subgraph generated by the nodes’ neighbourhood.
The PiMx-Stats component was inspired by graphical tools such as Visone5
and Gephi6 . It presents a number of graph related statistics, such as the number
of vertices and edges, the clustering coefficient, the average distance and diameter
of the graph. There is also information about the number of search and removed
nodes. Search nodes represent the pages associated with a search engine query.
The relevant query and the number of times that this search node was accessed
are displayed. Information about the type and number of occurrences of the
events that were triggered is also provided.
Through the PiMx-Clustering view it is possible to view the details of the
documents pertaining to each task-cluster. Each cluster is assigned a unique
ID and each document in a cluster has associated with it a ranking value and
information about the status generated by the algorithm. More details about
this algorithm are found in Sect. 5
5 Incremental Graph Clustering Approach
In this section we give an overview of the incremental density-based task clus-
tering approach that we’ve adopted. The scope behind our algorithm iDeTaCt
5
http://visone.info/
6
http://gephi.github.io/
is two fold: (i) identify those nodes that belong together in a task-cluster and
(ii) identify when a switch between two nodes is effectively a task-switch.
Clustering entities into dense parts allows for the discovery of interesting
groups in different networks. Furthermore, clustering on time-evolving networks
is still an open research problem that has been addressed through different ap-
proaches including incremental clustering [2] which tracks the granular dynam-
ics of a network, such as edge and node addition and deletion, rather than a
time-window. This approach is quite applicable to the dynamics of information-
seeking behaviours, whereby new documents are added over time, which in turn
need to be assigned to an existing or new task-cluster.
5.1 iDeTaCt: incremental Density-based Task Clustering
The density-based clustering algorithm DBSCAN proposed by [6] produces par-
titional clustering, whereby a cluster is considered to be a continuous area of
arbitrary shape that is denser than its surroundings. DBSCAN relies on the idea
that the neighbourhood of a node up till some given radius defines the “im-
portance” of that node. Nodes that have a minimum number, η, of other nodes
at a distance less then are termed as core nodes. On the other hand, a node
that has no such neighbourhood is given the status of noise node, unless it is
contained within the neighbourhood of a core node, in which case it is assigned
the status of a border node. Thus and η ensure that node neighbourhoods are
dense areas.
In our case, we consider that a switch between two windows initiates an
association between them, and this increases as more switches are effected. This
incremental nature of the data can be represented by a graph, Ga (V, E) that
evolves with the introduction of new nodes (documents) and edges (switches).
The edge weights represent the association between the nodes.
The clustering of those documents that pertain to the same task can also be
represented by a graph, Gc (V, E) that will also need to be updated incrementally,
since a switch to a new document will trigger a decision process do deal with
the change. The changes to Gc (V, E) that our clustering algorithm has to deal
with include:
i. the creation of a new cluster: when the association between two nodes ex-
ceeds the threshold;
ii. merging of two clusters: when either a core or border node in one cluster
gets strongly associated with another core or border node in another cluster;
iii. absorption (a growing cluster): when the association between a core or a
border node and a new node exceeds the threshold .
Consider a typical situation whereby the initial edge weight between two
nodes u and v is > . With increased switches, the association strength increases
since the edge weight will decrease and possibly become ≤ . At this point,
either, or both, of u and v can become core nodes (depending on η) with the
consequence that nodes in their neighbourhood can either change status as well,
form a cluster or merge with an existing one. It might also be the case that either
u or v, or both, become border nodes, and thus form a potential cluster.
In iDeTaCt we use an association edge-weighting function W : IR → IR that
maps the number of window-switches between two documents to an edge weight
w(u, v) in Ga (V, E). We do not consider the direction of the edge, that is, an
edge from node A to node B is considered the same as an edge from B to A. The
resulting edge weight is inversely proportional to the number of window-switches.
Thus a high number of switches will result in a lower edge weight. This is similar
to the proximity and influence functions used in [7, 8] respectively, whereby two
nodes are considered to be closer together if the edge weight between them is
less.
We compute the number n of edges between two nodes as a fraction of a
defined maximum number of edges, h. This maximum number is empirically set
to 10 which is considered to be sufficiently indicative of a strong association
between any two documents. Thus if the number of edges is 1, the value passed
1
on to the W would be equal to 10 .
The edge-weighting function W ( nh ) is based on the Epanechnikov kernel [8]
and is defined as:
3 2
W (x) = 4 (1 − x ) |x| ≤ 1
(1)
0 else
iDeTaCt takes as parameters the newly generated edge e and the old clus-
tered graph Gc (V, E) and works as follows:
– The association edge-weighting function computeAssociation takes as pa-
rameter the number of switches between u and v and returns the updated
weight w(u, v) of e.
– This weight is used to increase the ranking of nodes u and v within a cluster.
If w(u, v) is less than or equal to the node’s ranking is increased by a factor
of 0.85, otherwise it is increased by a factor of 0.15. This is in line with the
way that Firefox’s frecency algorithm7 assigns a bonus to recently viewed
pages.
– If edge e does not exist in Gc then e is added and Gc (V, E) is updated. This
results in endpoints u and v of e becoming connected in Gc (V, E).
– Then for both u and v, if they are not core, we consider all their incident
edges to check whether the changes have effected their status.
– If there are η or more such edges incident on node u then its status is set to
core.
– If u is not core but is adjacent to a core node then its status is defined as
border.
– Nodes that are neither core nor border are considered as noise and are placed
on a stack for later consideration.
7
https://developer.mozilla.org/en-US/docs/Mozilla/Tech/Places/Frecency_
algorithm
– Then iDeTaCt calls updateClusters which performs a Breadth-First-
Search over Gc to find the updated induced subgraphs. Each subgraph rep-
resents a cluster.
– In the process, updateClusters tries to include particular nodes from the
stack that are still labelled as noise using the procedure findWeakNodes.
– In findWeakNodes, if a node i is found to be a neighbour to u and v in
Gc (V, E) which have a status of core than the status of i is changed to weak
and it is added to Gc (V, E). Although such nodes have a weak relation with
the surrounding nodes, in that they fall short of the threshold, they are
connected to nodes which in turn are strongly connected.
– iDeTaCt returns a list of clusters that can be visualised through the PiMx-
Viz and the PiMx-Clusters components.
6 Evaluation
In our evaluation we wanted to verify whether iDeTaCt was able to: (i) identify
those nodes that belong together in a task-cluster and (ii) identify when a switch
between two nodes is effectively a task-switch.
For the evaluation we made use of the PiMx framework to simulate the task
execution trails of the users from groups 1 (considered as the baseline group)
and 2 (interleaved tasks). Details about the number of pages visited and the
number of switches made by participants in these two groups can be seen in Fig.
2 and Fig. 3. We did not use the data from Group 3 since we wanted to initially
evaluate our approach on data that was collected during a single session.
Fig. 2: Pages/Switches for Grp 1 Fig. 3: Pages/Switches for Grp 2
For each trail we used iDeTaCt with core parameter η values of 1, (D 1 )
and 2, (D 2 ). With η = 1, two nodes will become core when they are connected
by a single edge whose weight w(u, v) ≤ . Similarly with η = 2, a node will
need to be connected to two other nodes through two similar edges. was set
to the maximum value of 0.72 which is equivalent to a minimum of two window
switches (using equation Eq. 1).
We used standard information retrieval metrics to evaluate the clusters for
each of the three tasks. These metrics involve (i) precision, defined as the per-
centage of documents correctly assigned to a task-cluster over the total number
of documents in the task-cluster, (ii) recall, defined as the percentage of docu-
ments correctly assigned to a task-cluster over the total number of documents
that should have been assigned to that task-cluster, and (iii) F1-measure, de-
fined as the combined measure that assesses a trade off between precision and
recall. Whenever the algorithms generated two or more clusters for documents
from the same task, we considered the cluster which was more representative of
the task, that is, it contained the highest number of documents. In the case of
the interleaved tasks in Group 2, we expect the algorithm to be able to cluster
the interleaved tasks as if they were actually none interleaved.
Precision was 100% when we tried both D 1 and D 2 on all the task-trails
from Group 1. However when we used D 1 on the dataset from Group 2 it was
less then 100% in all cases except one. When we changed η to 2 on the data
from Group 2 we got 100% precision in 66% of the cases, in all the 3 tasks. In
the rest, the precision was less than 100% for only one of the tasks.
We focus on the more interesting recall and F1-measure. The averaged results
are shown in Fig. 4 and Fig. 5 respectively. We again compute the recall and
F1-measure for all the task-trails from Groups 1 and 2, and we do this for every
task separately.
Recall for the task-clusters generated on the data from Group 1 was highest
when we used D 1, with the averaged recall being highest for task 2, at 70.7%.
Task 3 had the lowest averaged recall at 36.5% due to the algorithm generat-
ing multiple, 2 or 3-node clusters. The possible reason for this could be due to
the familiarity of the participants with this topic. The fact that a very impor-
tant music event was forthcoming when the experiment was conducted, might
have effected the participants’ information seeking behaviour with many of them
knowing where to search and thus the number of re-visits was low.
The F1-measure for the task-clusters from Group 1 was consistent with the
recall and was again highest when we used D 1. As expected, the number of
captured nodes with this value of η was higher than with D 2 and the resultant
task-clusters where denser, yet still separate. Task 3 had once again the least
averaged F1-measure irrespective of η.
We now consider the recall and F1-measure based on the data from Group 2.
The values obtained for the task-clusters associated with task 2 were the highest,
and ranged between 50% and 60%. This trend is in line with the results we got
for the task-clusters from Group 1, however both values for task 3 are slightly
higher then those for task 1 except for the F1-measure based on D 1.
The recall and F1-measure based on the data of 2 of the participants from
this group resulted in exceptionally low values for the task-cluster related to
task 3. This was independent of η. For another user from the same group, we
observed the same low values for the task-cluster related to task 1. The common
feature observed across the data of these 3 participants was that the relevant
task was fragmented in multiple 2 or 3-node clusters.
From the generated graphs it was possible to observe that all the participants
tend to open up a number of documents which they only visit once. This was
more accentuated in almost 50% of the cases from Group 2. For some of these,
we manually inspected their log file and found that in fact these participants
typically opened up a number of tabs in succession, as in the case of a result page
associated with some query. They however only visited some of those opened tabs
once. Different individuals however did revisit some of the documents multiple
times and these acted like hubs/authorities to the other accessed documents thus
allowing for the generated clusters to be more consistent.
The fact that iDeTaCt managed to cleanly separate all the tasks, using the
data of all the participants from the baseline group is already encouraging. It
is even more encouraging when we used the interleaved tasks and found the
algorithm to be sufficiently reliable in more than 50% of the cases, even though
this was at the expense of capturing less documents from the referenced tasks.
Fig. 4: Average Recall Fig. 5: Averaged F1-measure
7 Future Work and Conclusion
We intend to take into consideration the type of edge, which requires us to extend
iDeTaCt to handle navigational events, especially those relating a query with its
result pages. Navigational events accounted on average for 30% of all switches
performed by each participant. We also plan to consider the window captions
and apply a similarity function, such as cosine similarity, over this content, based
on the clusters generated by iDeTaCt. Furthermore, we want to make use of
iDeTaCt for task-resumption and in-line task/document recommendations. We
intend to evaluate these extensions and compare our results.
In this paper we described the experiment we conducted to collect data for the
evaluation of our iDeTaCt incremental graph clustering algorithm. We executed
the algorithm over the task collections and used the PiMx framework to analyse
its performance. The results showed a high precision and recall over the data
from the baseline group and a recall of more than 50% over the data for the
interleaved tasks. This is considered to be very encouraging and motivates us to
further investigate how to improve our algorithm.
References
1. Bernstein, M., Shrager, J., Winograd, T.: Taskpose: Exploring Fluid Boundaries in
an Associative Window Visualization. In: 21st ACM Symposium on User Interface
Software and Technology, pp. 231-234. ACM Press New York, NY, USA (2008)
2. Charikar, M., Chekuri, C., Feder, T. and Motwani, R.: Incremental clustering and
dynamic information retrieval. In Proceedings of the twenty-ninth annual ACM
symposium on Theory of computing (STOC ’97), pp. 626-635. ACM, New York,
NY, USA (1997)
3. Costache, S., Gaugaz, J., Ioannou, E., Niederee, C., Nejdl, W.: Detecting contexts
on the desktop using bayesian networks. In: Desktop Search Workshop co-located
with SIGIR (2010)
4. Dabbish L., Mark, G. and Gonzlez, V.M.: Why do I keep interrupting myself?:
environment, habit and self-interruption. In Proceedings of the SIGCHI Conference
on Human Factors in Computing Systems (CHI ’11), pp.3127-3130. ACM, New
York, NY, USA (2011)
5. Dumais, S., Cutrell, E., Cadiz, J.J., Jancke, G., Sarin, R., and Robbins, D.C.: Stuff
I’ve seen: a system for personal information retrieval and re-use. In Proceedings of
the 26th annual international ACM SIGIR conference on Research and development
in informaion retrieval (SIGIR ’03), pp.72-79. ACM, New York, NY, USA (2003)
6. Ester, M., Kriegel, H.-P., Sander, J., Wimmer, M., and Xu, X.: Incremental cluster-
ing for mining in a data warehouse environment. In Proc. of 24th VLDB Conference
(1998).
7. Falkowski, T., Barth, A. and Spiliopoulou, M.: DENGRAPH: A Density-based Com-
munity Detection Algorithm. In Proceedings of the IEEE/WIC/ACM International
Conference on Web Intelligence (WI ’07). IEEE Computer Society, Washington,
DC, USA, 112-115 (2007)
8. Günnemann, S. and Seidl, T.: Subgraph Mining on Directed and Weighted Graphs.
In Proc. of the 14th Pacific-Asia Conference on Knowledge Discovery and Data Min-
ing (PAKDD 2010), pp. 133-146. Hyderabad, India. Springer - Heidelberg, Germany.
(2010)
9. Jones, W.P., Teevan, J.: Personal Information Management. ISBN 9780295987378,
University of Washington Press (2007)
10. Mayer, M.: Visualizing web sessions: improving web browser history by a better
understanding of web page revisitation and a new session- and task-based, visual
web history approach. PhD thesis, University of Hamburg (2008)
11. Morris, D., Ringel Morris, M., Venolia, G.: Searchbar: a search-centric web history
for task resumption and information re-finding. In: 26th annual SIGCHI conference
on Human factors in computing systems, CHI ’08, pp. 1207-1216. ACM Press, New
York, NY, USA (2008)
12. Oliver, N., Smith, G., Surendran, A.C.: SWISH: Semantic Analysis of Window
Titles and Switching History. In: 10th International Conference on Intelligent User
Interfaces, pp. 194-201. ACM Press, New York, NY, USA (2006)
13. Wang, Q. and Chang, H.:Multitasking bar: prototype and evaluation of introducing
the task concept into a browser. In Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems (CHI ’10). ACM, New York, NY, USA, 103-112.
(2010)