Convolution Neural Network-Based Expert Recommendation for
Alert Processing
Haidong Huang
1
, Liming Wang
1
, Rao Fu
2*
, Jing Yu
1
, Ding Yuan
2
and Danqi Li
2
1
State Grid Jiangsu Electric Power CO., LTD. Nanjing, China
2
Xuzhou Supply Company, State Grid Jiangsu Electric Power CO., LTD. Xuzhou, China
Keywords: Trouble Ticket, Expert Recommendation, Expert Profile, Convolution Neural Network.
Abstract: Alerts and associated trouble tickets provide extremely useful information for IT system maintenance.
However, the continuously occurrence of thousands of tickets also leads to a big challenge for accurately and
effectively dispatching them to skilled experts for quick problem-solving. To cope with such a challenge, this
paper develops a convolution neural network-based expert recommendation approach for in-time alert
processing. First, expert profile is built by extracting domain words from historical tickets, and is encoded
into a sentence like problem description. Second, an attention-based convolution neural network (CNN) is
developed not only to learn a unified representation for both problem description and expert profile but also
measure the semantic similarity between them. Finally, an ordered expert list is outputted. We evaluate our
approach on a real-world data set. Experimental results show that compared to the best baseline approach, our
approach can not only improve 2.8% in terms of p@1 but also shorten 11.7% in terms of the mean number of
steps to resolve (MSTR).
1 INTRODUCTION
Trouble tickets play a very important role in the
complex IT system maintenance. When an event
happens, or when special situations, errors, even
faults occur during the IT service consumption, a
trouble ticket, also called issue ticket, is generated,
which records the detailed problem symptom. And
then, the ticket management employed by the IT
system automatically dispatches the ticket to domain
experts for problem-solving. Once the problem is
fixed, the ticket is closed. System maintenance staffs
always attempt to quickly bring an abnormal service
back to normal by assigning skilled and well-matched
experts using the deployed expert recommendation
module. However, several real-world situations, such
as diverse troubles, vague problem descriptions
written in a natural language way, huge size of tickets
and low efficiency of manually assigning experts,
pose great challenges on the expert recommendation
module to avoid violating the signed Service Level
Agreement (SLA) with users. Therefore, rapid
problem-solving strongly depends on efficiently and
accurately expert recommendation, which motivates
us to focus on expert recommendation for trouble
tickets.
Typically, a trouble ticket contains at least five
fields, including ‘problem description’, ‘problem
type’, ‘expert name/id’, ‘resolution’ and ‘status’, as
shown in Figure 1. The problem description presents
the detailed symptom occurring at system runtime.
Each ticket is assigned to a specific problem type
belonging to the problem category. The expert name
is used to identify a specific expert who attempts to
solve the problem. The resolution records the detailed
approaches to fix the problem. The status is setting to
‘closed’ if the problem has been fixed, otherwise it is
setting to ‘open’. The ticket resolving process is
regarded as a ticket delivery sequence starting from
an initial expert to the final resolver. Initially, an
incoming ticket was assigned to an expert. If the
problem is fixed, the ticket is closed. Otherwise, the
ticket dispatching system delivers the ticket to
another expert. Such a delivery process is repeated
until the ticket is closed. The last expert who resolved
it is called a resolver.
Although a few studies (Shao Q, Shao Q, Agarwal
S, Botezatu M. M., Xu J, Zhou W) have been reported
to deal with ticket dispatching or expert
recommendation, there are still many limitations in
existing methods that deserve further investigation.
The traditional machine learning technology, such as
logistic regression (LR)( R. Qamili), decision tree
Huang, H., Wang, L., Fu, R., Yu, J., Yuan, D. and Li, D.
Convolution Neural Network-Based Expert Recommendation for Alert Processing.
DOI: 10.5220/0012284900003807
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 2nd International Seminar on Artificial Intelligence, Networking and Information Technology (ANIT 2023), pages 401-408
ISBN: 978-989-758-677-4
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
401
(DT) (R. Qamili) and support vector machine (SVM)
(Agarwal S), has been applied to conduct ticket
dispatching. An obvious advantage is that these
approaches can directly work on trouble tickets with
a little effort. However, the characteristics of ticket
problem descriptions, such as unformatted, large
vocabulary size and short texts, make expert
recommendation suffer from low accuracy. The main
reason lies in that the representation models, such as
the n-gram (R. Kallis), TF-IDF (R. Qamili) and LDA
(Zhou W), generally used by these approaches cannot
characterize trouble tickets well. To solve this issue,
several approaches based on the deep learning
technology has been proposed and shown its potential
in meeting the need of trouble ticket expert
recommendation and improving recommendation
performance. In this paper, to leverage the
characteristics of tickets well to further improve
recommendation accuracy, we propose a deep neural
ranking model-based expert recommendation
approach for trouble tickets by combining expert
profiling, vector-based ticket representation, the
attention-based convolution neural network. Further,
we evaluate the effectiveness of our model on a real
trouble ticket dataset.
Figure 1. An instance of the trouble tickets.
In summary, our contributions are the followings:
1) An expert profiling component is designed by
making full use of ticket problem descriptions and
resolutions to characterize expert’s professional
knowledge with domain words, which is helpful to
improve the efficiency and accuracy of ticket
assignment.
2) Two attention-based sentence models
characterizing problem description and expert profile,
respectively, are integrated into our recommendation
approach and mapped into the same vector, which
enables the semantic similarity measure between the
problem description and the expert profile and benefit
to improve the recommendation accuracy.
The rest of the paper is organized as follows. In
Section 2, a deep learning-based expert
recommendation approach is proposed. The
experiment settings are explained and the
performance evaluation results are discussed in
Section 3. At last, we conclude our work in Section 4.
2 THE PROPOSED APPROACH
2.1 Problem Formulation
Assume that
is a set of tickets
from a given IT system, where each ticket can be
represented by a quintuple, denoted as
>, . The first component denotes
the unique identifier of a ticket. The components
and are the problem type, the problem description
and the problem resolution, respectively. We assume
that the number of problem types is M. Resolutions
from multiple experts involved in the ticket routing
sequence are merged into the final problem resolution.
The last component
denotes a ticket routing
sequence containing k ordered experts. We operate on
historical tickets to output a set of instances, denoted
as =
, where
is the problem
description for the ith ticket,
is the ith expert
involving in its ticket routing sequence, and
is a
competency score of expert
. If expert
is a
resolver for the ith ticket,
, otherwise
.
Therefore, given the set , our purpose is to construct
a ranking model that calculates an optimal score
for each pair
, s.t. an expert with a strong
competency has a high score. Formally, the expert
recommendation task is to learn a ranking function
from historical pairs, as shown in Equation (1),
(1)
where function maps a pair <
to a
similarity vector, where each component reflects a
certain type of similarity, e.g., lexical, syntactic, or
semantic. The weight vector is a parameter of the
ranking model and is learned during the training.
2.2 Framework
The overview of our expert recommendation
approach is illustrated in Figure 2. Our approach
consists of four components, including data
preprocessing, expert profiling, vector-based
representation, and convolution neural network based
ranking model. Since problem descriptions record
textual information, the data preprocessing
component is essential to do some text preprocessing
operations by applying the natural language
techniques. And then, each problem description is
represented as a sentence vector using the learned
word vector from historical tickets. On the other hand,
domain words are extracted from problem
descriptions and resolutions to characterize expert
profile. Each expert is also represented as a sentence
GUI is failing with “Unable to Logon: RT11844:
Security exception: [CLI Driver] SQL30081N. A
communication error has been detected.
Communication protocol being used: “TCP/IP”.
Communication API being used: “SOCKETS”.
Location where the error was detected”.
Mary/
OSG021
stopped transition on g2 and g4 and recycled WAS on e8/
e9/ec/ed, then restarted transition. But still does not work.
Jose/
SMGNA054
There seems to be some authorization issue. DDF on
DB2B stopped with a mode force then restarted
Andrew/
SSAPHWOA023
the 2 threads that I canceled are stuck in DB2 and not
rolling back so they will continue to hold the locks they
have. DB2B was recycled under 5 minutes. The problem
has been solved, ticket is closed.
…
Problem Description
Resolutions
…
Experts
Problem Type
Status
DB2Security
ANIT 2023 - The International Seminar on Artificial Intelligence, Networking and Information Technology
402
vector. Further, each ticket is transformed into at least
one triplet containing a problem description vector,
an expert profile vector and a competency score.
Finally, an attention-based convolution neural
network ranking model is trained on these triplets.
When an incoming ticket arrives, the resulting
ranking model outputs an ordered expert list for the
ticket based on the matching score. Experts with the
top-N competency score are recommended to resolve
the incoming ticket one by one until it is resolved.
Figure 2. Overview of the proposed Approach.
2.3 Vector-Based Ticket
Representation
A proper ticket representation approach is crucial to
the task of expert recommendation. Recently, the
word vector technology (Zhou W), also known as
word embedding, has been successfully applied in the
domain of text representation and has been
demonstrated that it has a good effect in charactering
text semantic information. We argue that deriving an
effective representation for problem descriptions
plays an importance role in automating IT service
management. Thus, in this paper, the word vector
technology is also applied to ticket representation.
Assume that is the word vocabulary of tickets.
Let
be a d-dimensional word vector to
represent any word from ticket problem descriptions
or resolutions, and belongs to a word vector matrix
. Further, a ticket is represented as a
sentence combining its problem description and
resolution. Using the word vector matrix, each ticket
is mapped to their vector representation. Assume that
ticket contains words. The sentence for ticket is
represented as a matrix
.
Although there are some publicly pre-trained
word vector models, such as word2vec (Aggarwal V),
FastText (Athiwaratkun), and Bert. However, due to
the wide existence of non-dictionary words in tickets,
these pre-trained word vector models cannot be
applied to our work. According to the common
experience that a minimal size of the corpora required
for learning word vectors should be at least in the
order of hundreds of thousands, we use sufficient
historical tickets in our database as text corpora to
learn and extend the existing word vectors applicable
to IT service management.
2.4 Expert Profile and Representation
Expert profiling is a critical factor impacting the
efficiency and accuracy of ticket assignment. Here,
we use the professional knowledge or skills owned by
experts to characterize expert profile, where the
professional knowledge indicates his/her expertise to
resolve the ticket. For a given ticket, it is difficult to
determine whether an expert’s profile matches with
the ticket problem description. Thus, we represent
each expert profile as a group of domain words
extracted from problem descriptions and resolutions
to indicate general interests and activities of an
expert, defined as Definition 1.
Definition 1 (Expert Profile). A expert profile of
expert
is represented as a set of words,
, where
denotes
the set of K domain words describing general interests
and activities of expert
.
Different with the traditional way, we attempt to
automatically extract domain words from historical
tickets without human intervention. Using historical
tickets resolved by an expert, we think that those
words frequently occurring in problem descriptions
and resolutions of those tickets have a higher
probability to be domain words of the expert. For any
expert, we extract the same number of words to
characterize the expert.
We find tickets solved by a given expert from
historical tickets, and then candidate domain words
are fetched from these tickets. Domain words are
usually nouns or verbs provided by system
administrators or obtained from related documents
like the catalog taxonomy of system management.
After fetching candidate domain words, we measure
the importance of each word using the idea similar to
TF-IDF. For a candidate word w, we first measure its
frequency of appearing in tickets solved by expert
,
denoted as
. The higher
, the higher the
possibility that the word is a terminology is. Second,
we measure its expert frequency, denoted as
,
which represents the number of experts that have ever
solved the tickets containing the word. Third, we
measure its inverse expert frequency, denoted as
. The higher the value of
, the higher
the differentiation of the word among experts is. Last,
Expert
profile
Expert Profiling
Domain Word
Extraction
Expert Profile
Problem
Descriptions
Problem
Descriptions
and
Resolutions
Historical Ticket Data
Data Preprocessing
Vector-based
Representation
Triplets <Summary, Expert,
Competency Score / ?>
An Incoming
Ticket t
Convolution Neural Network
based Ranking Model
Matching score
Ranked
Expert List
Convolution Neural Network-Based Expert Recommendation for Alert Processing
403
we take these two factors into consideration to get the
importance of the word, and sort all candidate domain
words by the importance in a descending order and
then select the most important top-K words as the
domain words.
To accurately measure the semantic similarity
between a ticket and an expert, we make the expert
representation consistent with ticket representation.
An expert is represented as a sentence combining its
domain words. Using the same word vector matrix
used in the ticket presentation, each expert is mapped
to their vector representation. Assume that expert
profile contains words. The sentence for expert is
represented as a sentence matrix
.
2.5 CNN-Based Ranking Model
In this section, we build an attention-based CNN
ranking model to recommend experts, as shown in
Figure 3. Our model consists of two parts. The first
part consists of two attention-based sentence models
for mapping the trouble ticket problem and the expert
profile to their vector representation, respectively.
The second part is an expert ranking model that ranks
experts by learning the semantic similarity score
between the problem description and the expert
profile. We will describe these two parts in the
following texts. Generally, an incoming ticket can be
resolved by multiple experts with different
competency values. Thus, our model will recommend
all the candidates in the order that an expert with a
high competency ranks first.
In the embedding layer, the input is sentences
and , treated as sequences of words that represent a
ticket problem description and an expert capacity
respectively, where each word is drawn from a word
vocabulary . Words are represented by d-
dimensional vectors
. Input sentences and
are represented by matrices
=(
) and
=(
), respectively, where
and
are the number of words. To make our model focus
on domain words, we introduce the attention
mechanism. Note that although a sentence
representing the expertise profile is generated by
combining those domain words, expert’s expertise
differs from expert to expert. Thus, it is reasonable to
pay close attention to domain words related to query
topics. Specifically, an attention coefficient matrix
is generated by calculating the
similarity between two matrices and . For any
,
is a coefficient
matrix, where
and
denote the ith and jth word
vector in S and Q respectively, and
is a transformation function used to calculate the
attention coefficient. There are several alternative
approaches for the transformation function. Here, the
coefficient matrix calculation is defined as (2), where
denotes Euclidean distance between two words.
(2)
Figure 3. Overview of the CNN-based Ranking Model.
Further, the attention coefficient matrix is
transformed into an attention feature matrix with the
same dimension as sentence matrices, and then is
used together with the original sentence matrices as
inputs to the convolution operation. The attention
feature matrix enables the convolution operation to
learn domain words. The attention feature matrices
and
can be calculated by Equation (3) and
Equation (4), respectively, where the weight matrices
and
are parameters
learned during the training process. We randomly
initialize them.
(3)
(4)
The convolutional layer aims to obtain interesting
patterns of word sequences as ticket features and
expert capability features. We apply a one-
dimensional convolution operation on the sentence
vector
and the convolution kernel
in a wide convolution manner. The one-dimensional
convolution operation is taken in each h-size window
of sentence to obtain another sequence :
=
(5)
Sentence
matrix
Ws
Wq
Attention coefficient
matrix A
Ticket description
Expert capacity
k-max
Pooling
Folding
Wide
convolution
Fully
connected
clean full free space present
Xs
high space use full disk home
Softmax
Hidden layer
Join layer
Metric
Learning
Xq
M
ANIT 2023 - The International Seminar on Artificial Intelligence, Networking and Information Technology
404
where each row vector
in results
from a convolutional operation between jth row
vector in S and jth row vector in .
In practice, a set of filters, packed as
,
that work in parallel are applied in a deep learning
model, producing multiple feature maps
. A nonlinear feature of the text is
extracted using a Rectified Linear Unit (ReLU) as an
activation function after each convolutional layer.
The folding layer aims to capture the association
information of features between adjacent rows and
reduce the dimension of features, which sums up
every two rows in the feature map component-wise.
The pooling layer aims to extract the most
representative feature from sentences to reduce the
representation. We use the k-max pooling strategy to
select the top k features from all features according to
the sentence input length, and to keep the order
information of features, which can effectively
preserve the strength of recurring word features,
especially for words that may be repeated in problem
descriptions and solutions.
Using the fully connected layer, we fetch the
resulting representation vectors
and
of the
same dimension for processing tickets and expert
profiles.
The second part starts from using these two
vectors to be feed into the expert ranking model in
order to recommend experts. First, the similarity
between a ticket and an expert is calculated using
Equation (6) as the expert matching score.
(6)
where
is a similarity matrix, it acts as a
model of noisy channel approach for machine
learning, which has been commonly adopted as a
scoring model in information retrieval and question
answer. The similarity matrix M is a parameter of the
network and is optimized during the training. And
then, the joint vector is passed through a three-layer,
full connection, feed-forward neural network, which
allows rich interactions between a sentence pair from
one of the three components. The joint layer is
responsible for connecting these two eigenvector and
the similarity matrix using Equation (7).
(7)
The output of the hidden layer is calculated using
, where
is the weight vector of the
hidden layer, and is a nonlinear activation
function. Here, the ReLU function is used as the
activation function. Finally, the neurons output by the
hidden layer are passed through the softmax layer to
get the resulting probability as the expert
recommendation score, as shown in Equation (8),
where
represents the weight vector of the kth
problem area.
(8)
Our model is trained to minimize the binary cross-
function, as shown in Equation (9).
(9)
where
and
are the ground truth and the
prediction result for the ith pair of ticket and expert.
The parameters in the neural network are trained by
the mini-batch gradient descent approach, and the
sample size of each epoch is optimized during the
experiment. In order to mitigate the over-fitting issue,
we augment the cost function with L2-norm
regularization to constrain the parameters, and
employ the dropout strategy in the full-connection
layer to prevent feature co-adaption by dropping out
a portion of hidden units during the forward phrase.
2.6 Ranking-Based Expert
Recommendation
For an incoming ticket t and a set of experts D, ticket-
expert pairs
} are built and feed
into the ranking model. The expert recommendation
scores for these pairs are outputted using the trained
ranking model, and experts are sorted based on their
competency scores in a descending order. A
straightforward recommendation policy is to
recommend an expert who has not been
recommended from the ordered expert list each time
until the ticket is resolved.
3 EXPERIMENTS
3.1 Experiment Settings
The ticket dataset used in our experiments was
collected from an account of a large IT service
provider, which contains over 479079 tickets
belonging to 95 problem types and 582 system
maintenance experts. Statistically, the average
number of tickets resolved by an expert is close to
823. 10% of tickets in the dataset are randomly
selected to generate the testing dataset, while the rest
is used as the training dataset. After that, the natural
language processing technique is used to remove stop
words and build part-of-speech tags for ticket
problem descriptions and resolutions. The nouns,
adjectives and verbs are kept as the signature term
candidates and are concatenated into a sentence by
keeping the word order in its original texts. Further,
Convolution Neural Network-Based Expert Recommendation for Alert Processing
405
for each ticket in the training data, it is divided into
several instances with the form
, where
denotes a ticket represented by a sentence,
is an
expert involved in the ticket routing sequence and
represented by a sentence of domain words, and
denotes whether expert
solves ticket
or not,
if expert
is a resolver, otherwise
.
For each ticket in the testing data, it is transformed
into one instance with the form
by only
considering the resolver of this ticket.
Two common metrics in expert recommendation
are used to evaluate our approach, and they are
precision@ and Mean Steps to Resolve (MSTR)
(Shao Q, Xu J, Xu J). Precision@ (short for p@)
(Miao G, Aggarwal V) relates to precision, where
is a position parameter. For example, p@1 denotes
the probability that the first recommended expert is a
resolver, which is a natural way to indicate the
recommendation quality of the retrieved top-N
experts. Besides performance, we also use MSTR to
evaluate efficiency by measuring the mean steps of
resolving a ticket. Obviously, we prefer to a lower
MSTR for an efficient recommendation. Assume that
is a set of tickets,
, expert
is the resolver
of ticket
and
denotes a ordered expert
list recommended by any expert recommendation
algorithm. MSTR can be calculated as:
(10)
where
if
contains expert
and
denotes the position of expert
in the list, otherwise
.
Further, the four state-of-the-art approaches are
considered for comparison, including logistic
regression (LR) (R. Qamili), support vector machine
(SVM) (Agarwal S), STAR (Zhou W) and ABCNN-
1 (Yin W).
All recommendation models are implemented
using Python. For convenience, we name our model
as CNN-ATT. The testing machine is Windows 10
equipped with Intel Xeon E5-2699 V4 2.3GHz CPU
and 256GB RAM. We evaluate all models on the
testing data set using the above mentioned metrics.
3.2 Results
First, we evaluate the impact of different
representation models on recommendation
performance and efficiency. We compare four
common representation models, including TF-IDF,
Word2Vec, FastText and Bert, and present the results
in Table 1. We can see that the distributed
representation achieves a good performance.
Specifically, the algorithm using Bert performs better
than the algorithms using any other text
representation model in terms of P@1 and MSTR.
Our expert recommendation algorithm with a
combination of Bert and the deep learning technology
has a great improvement on P@1 over 22.1% and
6.95% compared to TF-IDF and Word2Vec,
respectively. The results using FastText are similar to
Word2Vec. On the other hand, our algorithm also has
an obvious efficiency improvement in terms of
MSTR, over 15.97% and 9.0% compared to TF-IDF
and Word2Vec, respectively. The main reason lies in
that the Bert model trained from historical tickets can
characterize the distribution of features automatically
and represent the tickets well, while the traditional
approaches highly depend on manual feature
engineering and may miss important features.
Second, we evaluate the impact of sentence length
on expert recommendation performance and
efficiency. We measure the distribution of the number
of words appearing in problem description and
characterizing expert profile from the historical
tickets, respectively. As for the length of problem
description, we can find that about 77% tickets
contain 25 to 100 words, and the maximum length is
197. As for the length of sentence representing expert
profile, we can find that about 68% tickets contain 10
to 120 domain words, and the maximum length is
237. Hence, we evaluate our algorithm by varying the
sentence length, and the results are shown in Figure
4, where the length of sentence model for problem
description varies from 20 to 100 while the length of
sentence model for expert profile varies from 10 to
120. We can observe that the recommendation
performance is optimal when the length of sentence
models for problem description and expert profile is
50 and 100, respectively, which means that the longer
length of sentence models do not always result in a
good recommendation performance, but spends more
computing time.
Table 1. Results from the model using different
presentation approaches.
Representation
P@1
MSTR
TF-IDF
0.7160
3.13
Word2Vec
0.8175
2.89
FastText
0.8104
2.95
Bert
0.8743
2.63
ANIT 2023 - The International Seminar on Artificial Intelligence, Networking and Information Technology
406
Figure 4. The impact of the sentence length on
recommendation performance.
Table 2. Overall performance comparison.
Approaches
P@1
MSTR
LR
0.6672
4.53
SVM
0.6895
4.12
STAR
0.8459
3.01
ABCNN-1
0.8464
2.98
CNN-w/o-ATT(ours)
0.8497
2.97
CNN-ATT(ours)
0.8743
2.63
Automating expert recommendation can be tackled
by applying different approaches. Here, we compare
several alternative algorithms from aspects of the
traditional machine learning and deep learning
technique in terms of p@1 and MSTR. Overall
performance results are shown in Table 2. We have
some interesting observations. First, compared to
traditional machine learning approaches, the
approaches based on the deep learning technology
perform better, which means that our CNN ranking
model-based solution is effective by making full use
of the semantic information from ticket problem
texts. Second, the approach using the attention
mechanism performs better than the others without it,
which means that the attention mechanism can help
our model identify key words in problem descriptions
and expert profiles and improve the recommendation
performance. Of course, we also see that the optimal
precision of the initial expert recommendation is
87.43%, which means that our approach dispatches a
trouble ticket to the following expert in the ordered
recommendation list until the ticket is resolved.
4 CONCLUSION
To quickly and accurately dispatch trouble tickets
from complex IT systems to skilled expert for
problem-solving, a deep learning-based expert
recommendation approach is proposed in this paper.
Our approach takes both problem description and
resolution into consideration to characterize expert
profile and build recommendation model. The unified
distributed representation is first applied to
characterize both problem description and expert
profile. And then, an attention-based convolution
neural network is built to learn a ranked expert
recommendation model by taking two sentence
models as inputs. For an incoming ticket, an expert
list is recommended to dispatch the trouble ticket. The
experimental results on real-world tickets show that
our approach performs better than both the traditional
machine learning-based approaches and the
convolution neural network-based approaches in
terms of p@1 and MSTR.
ACKNOWLEDGMENTS
The work was supported in part by State Grid Jiangsu
Electric Power CO., LTD (J2022014).
REFERENCES
Shao Q, Chen Y, Tao S, Yan X, Anerousis N. EasyTicket:
a ticket routing recommendation engine for enterprise
problem resolution(C). Proceedings of the VLDB
Endowment. 2008, 1(2):1436-1439.
https://doi.org/10.14778/1454159.1454193
Shao Q, Chen Y, Tao S, Yan X, Anerousis N. Efficient
ticket routing by resolution sequence mining(C).
Proceedings of the 14th ACM SIGKDD international
conference on Knowledge discovery and data mining.
2008, pp. 605-613.
https://doi.org/10.1145/1401890.1401964
Yin W, Schütze, Hinrich, Xiang B, et al. ABCNN:
Attention-Based Convolutional Neural Network for
Modeling Sentence Pairs (J).Transactions of the
Association for Computational Linguistics, 2016, 4:259-
272. https://doi.org/10.1162/tacl_a_00097.
Agarwal S, Sindhgatta R, Sengupta B. SmartDispatch:
enabling efficient ticket dispatch in an IT service
environment(C). Proceedings of the 18th ACM
SIGKDD International Conference on Knowledge
Discovery and Data Mining. 2012, pp. 1393-1401.
https://doi.org/10.1145/2339 530.2339744
Botezatu, M. M., Bogojeska, J., Giurgiu, I., Voelzer, H., &
Wiesmann, D. Multi-View Incident Ticket Clustering
for Optimal Ticket Dispatching(C). Proceedings of the
21th ACM SIGKDD international Conference on
Knowledge Discovery and Data Mining. 2015, 1711-
1720. https://doi.org/10.1145/2783258.2788607
R. Qamili, S. Shabani and J. Schneider, An Intelligent
Framework for Issue Ticketing System Based on
Machine Learning(C), Proceedings of 2018 IEEE 22nd
International Enterprise Distributed Object Computing
Convolution Neural Network-Based Expert Recommendation for Alert Processing
407
Workshop (EDOCW), Stockholm, 2018, pp. 79-86.
https://doi.org/10.1109/EDOCW.2018.00022
R. Kallis, A. Di Sorbo, G. Canfora and S. Panichella, Ticket
Tagger: Machine Learning Driven Issue
Classification(C), Proceedings of IEEE International
Conference on Software Maintenance and Evolution
(ICSME), Cleveland, OH, USA, 2019, pp. 406-409.
https://doi.org/10.1109/ICSME.2019.00070
Zhou W, Tang L, Li T, Shwartz L, Grabarnik GY.
Resolution recommendation for event tickets in service
management(C). Proceedings of 2015 IFIP/IEEE
International Symposium on Integrated Network
Management (IM) 2015 May 11, pp. 287-295.
https://doi.org/10.1109/INM.2015.7140303
Xu J, Zhang H, Zhou W, et al. Signature based trouble ticket
classification (J). Future Generation Computer Systems,
2018, 78:41-58.
https://doi.org/10.1016/j.future.2017.07.054
Zhou W, Li T, Shwartz L, Grabarnik GY. Recommending
ticket resolution using feature adaptation(C).
Proceedings of 11th International Conference on
Network and Service Management (CNSM), 2015 Nov
9 (pp. 15-21). IEEE.
https://doi.org/10.1109/CNSM.2015.7367333
Zhou W, Xue W, Baral R, et al. STAR: A System for Ticket
Analysis and Resolution(C)// Proceedings of the 23th
ACM SIGKDD international conference on Knowledge
Discovery and Data Mining. ACM, 2017, pp 2181-
2190. https://doi.org/10.1145/3097983.3098190
Athiwaratkun, Ben, A. G. Wilson, and A. Anandkumar.
Probabilistic FastText for Multi-Sense Word
Embeddings(C). In Proceedings of the 56th Annual
Meeting of the Association for Computational
Linguistics. 2018, pp.1-11.
https://doi.org/10.18653/v1/P18-1001
Xu J, He R. Expert recommendation for trouble ticket
routing (J). Data & Knowledge Engineering, 2018,
116(JUL.):205-218.
https://doi.org/10.1016/j.datak.2018.06.004
Xu J, He R, Zhou W, et al. Trouble Ticket Routing Models
and Their Applications (J). IEEE Transactions on
Network & Service Management, 2018, 15(2):530-543.
https://doi.org/10.1109/TNSM.2018.2790956
Miao G, Moser LE, Yan X, Tao S, Chen Y, Anerousis N.
Generative models for ticket resolution in expert
networks(C). In Proceedings of the 16th ACM
international conference on Knowledge discovery and
data mining (SIGKDD). 2010, pp. 733-742.
https://doi.org/10.1145/1835804.1835897
Aggarwal V, Agarwal S, Dasgupta GB, et al. ReAct: A
System for Recommending Actions for Rapid
Resolution of IT Service Incidents(C). In 2016 IEEE
International Conference on Services Computing
(SCC), 2016, pp. 1-8.
https://doi.org/10.1109/COMPSAC.2016.170
Quoc Le, Tomas Mikolov. Distributed representations of
sentences and documents(C). Proceedings of the 31st
International Conference on International Conference
on Machine Learning. 2014, 1188-1196.
https://dl.acm.org/doi/10.5555/3044805.3045025
ANIT 2023 - The International Seminar on Artificial Intelligence, Networking and Information Technology
408
