Classification of Video Viewing Task Types and
Recommendation of Videos
Tatsuro Ide
1a
and Hiroshi Hosobe
2b
1
Graduate School of Computer and Information Sciences, Hosei University, Tokyo, Japan
2
Faculty of Computer and Information Sciences, Hosei University, Tokyo, Japan
Keywords: Adaptive Information Retrieval, Machine Learning, Video Recommendation.
Abstract: YouTube is one of the largest and most sophisticated recommendation systems and a useful source of
information for users. In video search on YouTube, even the same user may have different purposes in mind
depending on the user’s state. However, videos are recommended based on the relevance of videos and the
user's viewing history, regardless of the user's state. This paper proposes a classification of video viewing task
types based on the user’s behavioral characteristics. By classifying the user's purpose as a task type, it enables
higher-order recommendation that fits the task type. Behavioral characteristics are momentary characteristics
of the user that appear from actions such as screen scrolling. The system implicitly records the user’s actions,
classifies the task type based on these parameters, and recommends the related video list on a mobile
application that imitates YouTube. We conducted experiments to evaluate classification of task types and
recommendation of videos.
1 INTRODUCTION
YouTube is one of the most popular video sharing
platforms and a useful source of information for
users. In video search on YouTube, even the same
user may have different purposes in mind depending
on the user’s current state. However, at present, it is
presumed that videos are recommended based on the
degree of relevance of videos and the past viewing
history, regardless of the user's current state. For
example, even when searching for content for a
learning purpose, videos of subscribed channels or
content that ignores the current purpose may be
recommended. Also, even when exploring a wide
range of videos, the user might be unable to make new
discoveries by returning to the video group that the
user habitually watches. Displaying related videos in
this way makes it possible that the user may be
trapped in a closed search space.
Previous research on personally adaptive
information retrieval has been actively conducted on
document retrieval systems (Athukorala et al., 2016).
On the other hand, there is almost no research on
information retrieval on video sharing systems such
a
https://orcid.org/0000-0001-5787-0443
b
https://orcid.org/0000-0002-7975-052X
as YouTube. This may be due to the difficulty of
extracting features from videos, the large number of
video groups, or the difficulty of acquiring video and
user data. A study on YouTube video search by
Google (Covington et al., 2016) showed that they
analyzed such a huge amount of personal information
by various measures and applied it to
recommendation of videos.
In this paper, we propose a classification of video
viewing task types and a method for recommending
videos appropriate for task types that are analyzed
according to the user’s behavioral characteristics of
viewing videos. We base the classification and the
method on a previous study on adaptive information
retrieval for a document retrieval system (Athukorala
et al., 2016). Our aim is to make higher-order
recommendations of videos by classifying the user’s
dynamic purpose from the user’s behaviors. Our
method uses behavioral characteristics that are the
instantaneous features of the user appearing from
detailed actions such as scrolling the screen. We
constructed a classifier that determines task types
based on the parameters obtained from such
behavioral characteristics. The classifier is a decision
Ide, T. and Hosobe, H.
Classification of Video Viewing Task Types and Recommendation of Videos.
DOI: 10.5220/0010824700003116
In Proceedings of the 14th International Conference on Agents and Artificial Intelligence (ICAART 2022) - Volume 3, pages 373-380
ISBN: 978-989-758-547-0; ISSN: 2184-433X
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
373

tree that was obtained by imposing video viewing
tasks on users.
We implemented a classification and
recommendation system as a mobile application that
imitated YouTube. The system records the user's
actions, classifies task types, and recommends related
videos. It implements video search by using the
YouTube Data application programming interface
(API), and realizes recommendation by filtering the
search for related videos and adding the separately
searched videos to the related video list.
To evaluate the task type classification and the
video recommendation, we conducted an experiment
by using this system. To reproduce the participants’
usual use of YouTube in the proposed method, we
also conducted a pre-questionnaire about their
interests in subscribed channels and video categories.
The result of the experiment indicated that the
classification accuracy of the task types was 60%,
which was higher than 1/3. However, regarding the
video recommendation, we found limitations due to
incorrect task types and unsuitable recommended
videos.
2 RELATED WORK
2.1 Classification of Task Types
According to Behavioral
Characteristics
Search activities can be divided into two major
categories: lookup and exploration (Marchionini et
al., 2006). In lookup search, in order for the user to
reach the correct area of the information space, the
user first accurately expresses the information that the
user has, quickly refers to the related result, and
finally arrives at the most suitable item. On the other
hand, in exploratory search, user behavior is dynamic.
The user starts the search with an unclear search
purpose in mind and initially issues an inaccurate
search query. In addition, the user reads the search
results and repeatedly reformulates queries according
to the newly found keywords.
Athukorala et al. (2015) showed that these lookup
and exploratory searches could be categorized by
easily measurable behavioral characteristics. Query
length, scroll depth, reading time, task completion
time, and cumulative numbers of clicks were shown
as effective behavioral characteristics for
classification. Based on this research, they further
constructed a classifier that recorded implicit search
behaviors on a paper search engine such as Google
Scholar and classified exploration tasks and research
tasks according to their parameters (Athukorala et al.,
2016). This classifier was implemented in the article
search system, and the determined tasks were used for
recommendation. In this system, query length,
reading time, and cumulative numbers of clicks were
used as behavioral characteristics.
2.2 YouTube Video Recommendation
Google has adopted various measures for YouTube
video recommendation. Among them, in the study
on the application of deep learning to a
recommendation system (Covington et al., 2016)
and the research on efficient input of context
information (Beutel et al., 2018), implicit features
were introduced and used to construct
recommendation systems. Both studies used implicit
features such as viewing histories and user genders,
but they did not use detailed and instantaneous user
actions such as scroll depth.
The former study showed that the conventional
matrix factorization-based recommendation
algorithm was replaced with deep learning to
improve accuracy. They proposed that solving
difficult problems with large amounts of fresh
content could be roughly divided into two stages: (1)
narrow down candidates from millions of videos; (2)
rank the videos according to their scores. In stage
(1), viewing histories, search results, user genders,
viewing areas, training sample ages, etc. were
inputted as feature quantities. The age of the training
sample is the time elapsed since the video was
uploaded, and it was observed that fresh content
tended to be viewed more frequently regardless of
taste. In stage (2), the embedded vectors of videos
and the numbers of recommendations were used for
scoring. The numbers of recommendations were
used for learning to lower the scores of unselected
videos even if they were displayed multiple times.
The latter study attempted to solve the problem
that the model size became large when the
embedded vectors were connected to the user's
context information. The used information was the
elapsed time before and after viewing, the device to
be viewed, and page information. Page information
was a feature of each page such as the top page and
the video playback page, and there was a tendency
for new content to be viewed on the top page.
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
374

3 DECISION TREE
GENERATION
In this paper, C4.5 (Quinlan et al., 1993), which is a
classification learning algorithm among the machine
learning algorithms of Weka (Witten et al., 2006), is
used to generate a decision tree. C4.5 is an algorithm
based on the divide-and-conquer method. Weka takes
a dataset in a special format called ARFF format as
input and outputs the result by the selected
classification learning algorithm and evaluation
method. Each data in the dataset is a set of the input
variable that is the branching condition of the
decision tree and the possible output of the leaf node
that has no children. We use the data obtained by
imposing a task on users as an input dataset, and
describe the query length, reading time, scroll depth
as input variables, and the task type as possible
output. The process of decision tree generation is
shown below.
3.1 Quantification of Conditions based
on Entropy
Based on information theory, C4.5 uses entropy to
quantify the discriminating power of the leaves of the
decision tree. In the set 𝐶 of the dataset, the possible
outputs belong to the set 𝐷, and the probability at
which 𝑥∈𝐷 occurs is expressed as 𝑝


𝐶

. The
entropy 𝑀

𝐶

for the set 𝐶 of the dataset is as
follows:
𝑀

𝐶

=−𝑝


𝐶

log𝑝


𝐶

∈
When the number of classes that divide the base of
the logarithm (possible output 𝑥) is set, the maximum
of 𝑀

𝐶

becomes 1. When it is close to 1, the dataset
is in a messy state.
3.2 Selection of Conditions
The information gain obtained by dividing 𝐶 into 𝑘
pieces with the input variable as a condition is 𝐺

𝐶

:
𝐺

𝐶

=𝑀

𝐶

−
|
𝐶

|
|
𝐶
|
×𝑀

𝐶




The information gain can be interpreted as the degree
to which the disorder is reduced depending on the
conditions. The quality of division can be defined by
this information gain. The dataset is divided under
each condition, and the one with the large information
gain is set in the leaf node. This is done recursively in
each subtree of the child to generate the decision tree.
4 PROPOSED METHOD
In this paper, the user's dynamic purpose which can
be read from the user's behavioral characteristics is
classified as a task type and applied to
recommendation. The proposed method mainly
consists of three components: a user interface, a
classifier, and a recommender:
1. The user interface records the user's actions when
viewing videos, which is realized as a YouTube
client application. This is almost the same as that
of YouTube Mobile, but it implicitly records the
actions.
2. The classifier obtains the parameters that the user
interface extracted from the actions. Then it
determines the task type.
3. The recommender filters the video search by the
task type and also adds the separately searched
video to the list.
4.1 Definition of Task Types and
Behavioral Characteristics
The task type is defined based on the paper
(Athukorala et al., 2016). They defined two task
types, “lookup” and “exploration”, for article search.
In this paper, we define the following three task types
for video search by newly adding “repeat”.
• Lookup: A task where the video to be searched
for is decided in advance; the user searches for a
specific video as a target (White et al., 2006).
• Exploration: A task where the video to be
searched for is not decided; the user searches a
wide range of content based on their interests.
• Repeat: A task where the video to be search for
is habitually checked by the user.
Behavioral characteristics are behaviors that
represent instantaneous user characteristics. In this
paper, the following three behavioral characteristics
were recorded and used as parameters.
• Query length: The number of words entered in
the query in the first search session; count by
separating them with spaces (Jansen et al., 2001).
• Scroll depth: Depth of scrolling up and down the
view of the video list.
• Reading time: Time to start watching the first
video.
Classification of Video Viewing Task Types and Recommendation of Videos
375

4.2 Classifier Parameter Determination
In this paper, Weka is used to select the parameters of
behavioral characteristics. The data obtained by
imposing a task on the participants on this application
is used as a dataset. Table
1 shows the assigned tasks
and each recorded parameter. The tasks assigned to
the participants are 4 lookup tasks, 8 exploration
tasks, and 8 repeat tasks for a total of 20 tasks. In the
lookup task, participants searched for the video that
they watched two hours before. In the exploration
task, participants searched videos in the category of
interest that they answered in advance. In the repeat
task, we used a group of videos that participants
habitually checked.
Table 1: Participant task completion data used in the
dataset.
Participants Query
len
g
th
Reading
time
Scroll
de
p
th
Task
1 2 35 0 Looku
p
2 113 30 Ex
p
loration
2 102 71 Exploration
1 185 30 Repeat
1 58 2 Repeat
2 3 97 51 Looku
p
2 187 74 Ex
p
loration
3 97 38 Ex
p
loration
1 181 24 Repeat
2 46 51 Repeat
3 2 25 0 Looku
p
1 256 162 Ex
p
loration
1 116 71 Ex
p
loration
0 389 85 Repeat
0 427 238 Repeat
4 2 35 24 Lookup
1 112 133 Ex
p
loration
2 262 67 Ex
p
loration
0 96 48 Re
p
eat
0 22 37 Repeat
Using this dataset as input data, J48, which
generates a decision tree based on C4.5 (Quinlan et
al., 1993), was selected as the machine learning
algorithm. The reason for adopting this algorithm was
that it was easy to understand the cause of
classification failure from the excellent visibility of
the decision tree. Figure 1 shows the decision tree
generated using cross-validation for the training data.
Leaf nodes with children represent query length,
reading time, and scroll depth respectively. Leaf
nodes that have no children represent the three tasks,
i.e., lookup, expansion, and repeat. The branch
comparison operation branches according to the
parameters of the parent node of each data.
Figure 1: Task type decision tree based on user data.
The query length was selected as the first
condition in the generation of this decision tree
because its information gain was larger than those of
the scroll depth and the reading time. The information
gains of the datasets in Table 1 are calculated as
follows. Since the data is divided into three
categories, lookup, exploration, and repeat, and also
since the numbers of data in their classes were 5, 10,
and 10 respectively, the entropy is the following:
𝑀

𝐶

=
5
25
log

25
5

10
25
log

25
10

10
25
log

25
10
≅ 0.960
The information gain obtained by dividing the
query length on the condition that it is larger than 0 is
the following:
𝐺

𝐶

= 0.960 − 
10
20

3
10
log

10
3

7
10
log

10
7


10
20

4
10
log

10
4

5
10
log

10
5

1
10
log

10
1
 ≅ 0.253
4.3 Recommender
The search parameters are changed according to the
task type determined by the classifier. If the task type
is classified as exploration, videos of the category of
interest are added to the related videos. If the task type
is classified as lookup, new videos and live streaming
of subscribed channels are displayed in descending
order of the relevance of the videos. If the task type is
classified as repeat, the new video of the subscribed
channel of interest is added to the related video
display regardless of the relevance of the video.
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
376

5 IMPLEMENTATION
We implemented our system as a mobile application
by using Android Studio and Google Pixel 4a. We
obtained YouTube video data by using the YouTube
Data API and played them back by using the Android
Player API. This application records the user's
behavioral characteristics while the user is viewing
videos on YouTube, classifies task types, and applies
them to recommendations. The query length, reading
time, and scroll depth are recorded as behavioral
characteristics.
The user interface imitates YouTube Mobile.
Figure 2 shows screenshots of the top page and the
video playback page. The valid bottom tabs are
Home, Search, and Subscribed Channels, and a pre-
questionnaire searches the video list for each
participants. The user presses the magnifying glass
icon at the top of the screen to start the search action.
If the user taps a video from each tab or the video list
of the search results, the video will be played. On the
video playback page, a list of related videos is
displayed below the video player. The user plays the
first video by searching for a video or selecting a
video from the video list on each tab. After that, the
user can search for a video by selecting a video from
the related video display or swiping to return to the
previous screen. The user presses the account icon to
the right of the magnifying glass icon at the top of the
screen to go to the login page. On this page, it
provides OAuth authentication to use YouTube user
data associated with the Google account.
Figure 2: Implemented application.
6 EXPERIMENT
We conducted an experiment to evaluate the dynamic
classification of task types and recommendation of
videos according to behavioral characteristics. Our
system uses a classifier that records data for three user
interactions (query length, scroll depth, and reading
time). In addition, the classifier treats these input data
as parameters and predicts whether the user's task is
lookup, exploration, or repeat. In this section, the
system that enables this classifier is called the full
system; the system that disables this classifier is
called the baseline system and is used for comparison.
There were 5 participants, the average age was 22
years old, and they habitually used YouTube. A pre-
questionnaire was conducted to simulate the video list
on each tab of YouTube.
6.1 Design
Each participant was asked to complete a total of 5
tasks, each of which consisted of 1 lookup task, 2
exploration tasks, and 2 repeat tasks, for the full
system and the baseline system. Also, the task order
was balanced to avoid the order effect.
The lookup task asked participants to search for
the video that they watched 2 hours before the
experiment. By this method, we ensured that the
participants would remember information about the
video that they watched, but forget detailed
information such as the video title and channel name
(Kane et al., 2000). For the exploration task, we asked
them to answer the categories of videos in which they
were interested but about which they did not know
well (Schraefel et al., 2005), and then to select the
appropriate kind of videos and investigate them
freely. This attempted to reproduce learning, which is
a typical exploratory search (White et al., 2006).
Table 2 shows the correspondence between the
categories selected by the participants in the
exploration task. In the repeat task, the videos that the
participants habitually watched were answered in
advance and used as usual. After the two exploration
tasks and the two repeat tasks were completed, we
asked a question about the related video display. This
allowed the full system and the baseline system to
change the orders of related videos.
Table 2: Video categories selected by the participants in the
pre-questionnaire.
Participant Video category
1 Music
2 Science & Technology
3 News & Politics
4 Sports
5 Music
Classification of Video Viewing Task Types and Recommendation of Videos
377

6.2 Procedure
The procedure of the experiment was divided into
three stages: a pre-questionnaire, watching a video for
the lookup task, and the main experiment. In the pre-
questionnaire, we prepared questions to reproduce the
individual YouTube pages. We asked them to select
a frequently viewed category from the YouTube
categories for the content displayed on the home tab,
and to answer a list of frequently viewed subscribed
channels for the subscribed channel tab. In addition,
for the categories to be searched by the exploration
task, we asked them to answer multiple categories in
which they were interested but about which they did
not know well, and to select the one that could obtain
valid results from the API.
To reproduce the situation of the lookup task, we
asked the participants to watch a specific video
individually 2 hours before the experiment. To
prevent the content from being memorized in detail,
we did not mention the lookup task to be done later.
This video was about 5 minutes long, and when the
participants finished watching it, we asked them to
return to their respective tasks, and told them that they
would perform the experiment 2 hours later.
After 2 hours and before this experiment, we
explained the tasks and functions of the application.
Each participant was asked to perform a total of 5
tasks. The lookup task was limited to the maximum
of 15 minutes, and the participant could finish the task
when the video was found. The participants spent 20
minutes each on the exploration and the repeat task.
Each of these experiments took about 90 minutes.
6.3 Accuracy of the Classifier
Table 3 shows the task types judged by the full
system. Since each participant had 5 tasks and 5
people worked on it, the accuracy was calculated for
a total of 25 tasks. Among the 25 tasks, 15 were
classified correctly, and the overall accuracy was
60%. Among these, the lookup task were 5 tasks, and
3 tasks, i.e., 60% of the tasks, were classified
correctly. The accuracy of the exploration task was
50% because 5 out of 10 tasks were classified
correctly. The accuracy of the repeat task was 70%
because 7 of the 10 tasks were classified correctly.
6.4 Evaluation of the Recommender
As a result of the post-questionnaire, 20% of the
participants answered that the full system (i.e., with
recommendation) was suitable, 50% of the
participants answered that the baseline system (i.e.,
without recommendation) was suitable, and 30% of
the participants answered that they did not notice any
difference between the two systems. There were
positive evaluations such as the related video display
that caught the eye in the exploration task (participant
3). However, many of the participants answered that
they did not notice the difference in the related video
display throughout the task. Some people said that
they noticed that the related video display had a video
display that was completely different from the
intended one (participants 1, 2, and 5). This is because
the related video display included a video that had
nothing to do with the purpose because the judged
task was different from the original task. Other
participants answered that the same video was
displayed repeatedly (participants 1 and 3). This is
because the video was added to the related video
display as a recommendation.
Table 3: Participant behavioral characteristics and the task
types judged by the full system.
Participant Query
length
Reading
time
Scroll
depth
Judged task
type
Correctness
1 2 34 1 Lookup Correc
t
2 62 63 Exploration Correc
t
3 70 37 Exploration Correc
t
1 67 94 Exploration Incorrec
t
2 417 92 Exploration Incorrec
t
2 2 50 4 Lookup Correc
t
1 112 60 Exploration Correc
t
2 64 13 Exploration Correc
t
0 23 31 Repea
t
Correc
t
0 62 7 Repea
t
Correc
t
3 1 25 16 Repea
t
Incorrec
t
1 49 4 Repea
t
Incorrec
t
1 73 108 Exploration Correc
t
1 22 8 Repea
t
Correc
t
2 38 54 Lookup Incorrec
t
4 1 14 5 Repea
t
Incorrec
t
1 86 30 Repea
t
Incorrec
t
1 147 35 Repea
t
Incorrec
t
0 19 4 Repea
t
Correc
t
1 142 19 Repea
t
Correc
t
5 2 32 3 Lookup Correc
t
1 115 6 Repea
t
Incorrec
t
1 119 48 Repea
t
Incorrec
t
1 146 12 Repea
t
Correc
t
1 120 1 Repea
t
Correc
t
7 DISSCUSSION
Our experiments showed that the task types were
classified with a certain degree of accuracy. On the
other hand, we found two problems from the
participants' evaluations. The first problem consisted
of two cases: (1) a recommendation was obtained
from an incorrect task type; (2) an inappropriate
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
378

recommendation was obtained from a correct task
type. An example of case (1) is that a subscribed
channel video was added to the related video display
because it was classified as a repeat task during the
lookup task. In case (2), even if the task type was
correctly classified, the video added by
recommendation did not satisfy the user's intention.
This problem was related to the recommendation
evaluation method. Although qualitative evaluations
could be obtained through questionnaires and
interviews, quantitative evaluations of recommended
videos could not be performed; this is because we
could not implement a system that would lead to
quantitative evaluation by, e.g., analyzing videos in
the viewing history.
To evaluate the recommendation result, a
quantitative evaluation of recommended videos
should be performed. At present, we simply add
videos that have been categorically searched, and add
videos from subscribed channels. It is necessary to
consider what kind of recommendation is preferable
in consideration of the results obtained by the
quantitative evaluation.
Since the task type is determined in the first search
session and the recommendation is continued based
on the task type after that, inappropriate
recommendation is continuously made in the case of
an incorrect classification. It is necessary to consider
a system that redetermines the task after watching the
video several times. Then, even if the user’s purpose
in mind changes, it will be possible to continue to
adapt to it dynamically by periodically redetermining
the task type.
In recent years, research on neural networks for
recommendation systems has progressed. YouTube
also incorporates the context of user information such
as video viewing histories and search histories into a
neural network and uses it for recommendation
(Covington et al., 2016). In this paper, instead of such
a large amount of data, we focused on the user's
behavior, which may be based on the user's purpose.
At this time, an important problem is how the user's
purpose and the user's behavior are linked in the video
search. The high readability of the decision tree makes
it easier for us to understand this problem, which will
be difficult when a neural network is used instead.
Bhabad et al. (2017) realized the recommendation
of video related information by ASR and OCR. The
method recommended web links, image links, and
YouTube links based on the text data from images
and sounds cut out from the video. Based on the task
type of this paper, Bhabad’s method worked
effectively when the purpose was clearly defined such
as lookup tasks. On the other hand, it was not suitable
when the users wanted to search a wide range of
videos such as exploration tasks and repeat tasks.
Silva et al. (2017) showed that comments on videos
could devide into technical, or instructional videos,
and non-technical videos. Although this is similar to
our research background, the point of view is
defferent. Since Silva’s method focuses on the video
itself, there are problems with videos increasing every
day and videos with only few comments. By contrast,
since our method focuses on the user’s behavior, it is
possible to avoid problems caused by the video itself.
C4.5, which was used in the decision tree
generation algorithm, has the problem that the
decision tree cannot be updated sequentially.
Especially when the dataset and the user's behavior
are extremely different as in the experiment in this
study, an inappropriate decision is made. Supervised
learning such as C4.5 requires input and correct
answer data and given tasks, and such data cannot be
analyzed sequentially by a decision tree generation
algorithm. It will be necessary to consider measures
that can be updated sequentially, such as
implementing the decision tree generation algorithm
itself in the application.
We adopted query length, scroll depth, and
reading time as behavioral characteristics because
they were general behaviors in search systems. In
addition to these, YouTube has other characteristic
operations such as video preview and maximization
and minimization of the video player. It is necessary
to verify whether these actions and other actions that
are effective in document retrieval are also effective
in video search.
Although we attempted to reproduce the function
of YouTube, some part of it imposed difficulty. For
example, the list of soaring videos on the exploration
tab and query search could be implemented with the
current YouTube Data API. However, the list of
recommended videos for a user displayed on the
home tab could not be implemented because it was
excluded from the current API. Also, even if the list
of subscribed channels is obtained, the list of videos
in order of posting date and time cannot be obtained.
Therefore, we conducted a pre-questionnaire to
simulate YouTube without using these unavailable
functions.
8 CONCLUSIONS AND FUTURE
WORK
We proposed a classification of video viewing task
types and a method for recommending videos
Classification of Video Viewing Task Types and Recommendation of Videos
379

appropriate for the task types. To dynamically
classify the task type of a user, the parameters of the
behavioral characteristics were recorded and
analyzed by a decision tree. We implemented an
application that implicitly recorded query length,
scroll depth, and reading time, determined the task
type by using a decision tree, and reflected it in the
related video display. We conducted an experiment to
evaluate the classification accuracy and the
recommendation of videos. Improvement of the
evaluation method such as implicitly evaluating the
search result list with bookmarks as in the paper
(Athukorala et al., 2016) is a future task.
REFERENCES
K. Athukorala, A. Medlar, A. Oulasvirta, G. Jacucci and D.
Glowacka. (2016). Beyond Relevance: Adapting
Exploration/Exploitation in Information Retrieval. In
Proc. ACM IUI, pp. 359-369.
P. Covington, J. Adams and E. Sargin. (2016). Deep Neural
Networks for YouTube Recommendations. In Proc.
RecSys, pp. 191-198.
G. Marchionini. (2006). Exploratory search: From finding
to understanding. In Com. ACM, vol. 49, no. 4, pp. 41-
46.
K. Athukorala, D. Głowacka, A. Oulasvirta, J. Vreeken and
G. Jacucci. (2015). Is exploratory search different? A
comparison of information search behavior for
exploratory and lookup tasks. JASIST, vol. 67, no. 11,
pp. 2635-2651.
A. Beutel, P. Covington, S. Jain, C. Xu, J. Li, V. Gatto and
E. H. Chi. (2018). Latent Cross: Making Use of Context
in Recurrent Recommender Systems. In Proc. WSDM,
pp. 46-54.
J. Quinlan. (1993). C4.5: Programs for Machine Learning.
Morgan Kaufmann.
I. H. Witten, E. Frank. (2005). Data Mining: Practical
Machine Learning Tools and Techniques. Morgan
Kaufmann.
R. W. White, B. Kules and S. M. Drucker. (2006).
Supporting exploratory search. Com. ACM, vol. 49, no.
4, pp. 36-39.
B. J. Jansen and U. Pooch. (2001). A review of web
searching studies and a framework for future research.
JASIST, vol. 52, no. 3, pp. 235-246.
M. J. Kane and R. W. Engle. (2000). Working-memory
capacity, proactive interference, and divided attention:
Limits on long-term memory retrieval. Journal of
Experimental Psychology: Learning, Memory, and
Cognition, vol. 26, no. 2, pp. 336-358.
M. C. Schraefel, D. A. Smith, A. Owens, A. Russell, C.
Harris and M. L. Wilson. (2005). The evolving mSpace
platform: Leveraging the Semantic Web on the trail of
the Memex. In Proc. Hypertext, pp. 174-183.
D. Bhabad, S. Therese, M. Gedam. (2017). Multimedia
based Information Retrieval Approach based on ASR
and OCR and Video Recommendation System. In Proc.
CTCEEC, pp. 1168-1172.
H. Silva, I. Azevedo. (2017). Instructional Videos and
Others on YouTube. In Proc. CSEDU, pp. 418-425.
ICAART 2022 - 14th International Conference on Agents and Artificial Intelligence
380