Knowledge Graph Embedding in Intent-Based

Networking
Kashif Mehmood, Katina Kralevska, and David Palma
Department of Information Security and Communication Technology (IIK)
NTNU—Norwegian University of Science and Technology, Trondheim, Norway
Email: {kashif.mehmood, katinak, david.palma}@ntnu.no

Abstract—This paper presents a novel approach to collection of operational goals and outcomes that a net-
network management by integrating intent-based net- work endeavors to achieve, articulated declaratively. This
arXiv:2405.07850v1 [cs.NI] 13 May 2024

working (IBN) with knowledge graphs (KGs); thus, cre- definition, however, stops short of specifying the means or
ating a more intuitive and efficient pipeline for service
orchestration. By mapping high-level business intents methods for actualizing these goals and outcomes. IBN’s
onto network configurations using KGs, the system core philosophy enables stakeholders, such as network
dynamically adapts to network changes and service administrators, application developers, network operators,
demands, ensuring optimal performance and resource and subscribers, to define their desired network behavior
allocation. This integration facilitates a deeper under- using a high-level language, bypassing the complexities
standing of network states and dependencies, enabling
predictive adjustments and real-time troubleshooting. of low-level network configurations. IRTF specifies the
We utilize knowledge graph embedding (KGE) to ac- following phases as the core part of an intent’s life cycle: i)
quire context information from the network and service user profiling and expression, ii) modeling and translation,
providers. The trained KGE model maps intents to iii) planning and conflict resolution, iv) activation and
services via service prediction and intent validation deployment, and v) assurance [5]. This work focuses on
processes in the proposed intent processing pipeline.
We evaluate the trained model for its efficiency in intent modeling and translation. It utilizes the available
the service mapping and intent validation tasks using context using KGs in the network domain to understand
simulated environments and extensive experiments. the intent expression and pave the path for its activation
The service prediction and intent verification accuracy in the underlying network infrastructure.
≥ 80% is achieved for the trained KGE model on a The integration of resource description framework
custom service orchestration intent knowledge graph
(IKG) based on TMForum’s intent common model. (RDF) and web ontology language (OWL) into the model-
Index Terms—service model, intent-based network, ing of IBN represents a pivotal advancement in the pursuit
knowledge graph learning, intent translation. of intelligent, adaptable, and semantically rich network
management systems. By leveraging these technologies,
I. Introduction the development of detailed knowledge graphs that fa-
Network and service providers face the daunting task cilitate a nuanced understanding of network dynamics,
of supporting applications and user demands at mas- policies, and, eventually, intents is made possible. By
sive scales, all while ensuring compliance with essential modeling networks as knowledge graphs, operators can
business and operational requirements such as security, understand the network’s current state and how it relates
availability, and latency [1]. Traditional network manage- to the desired operational state dictated by an intent.
ment systems are often time-consuming and error-prone This semantic richness enables more precise and dynamic
due to manual configurations and maintenance proce- mapping of intents to network configurations, significantly
dures [2]. Moreover, managing networks in multi-domain improving the network’s adaptability and responsiveness
environments, characterized by heterogeneous resources to changing network conditions.
with varying abstractions, introduces complexities and Randles et al. [6] propose an ontology for mapping the
potential incompatibilities. To address these challenges, intent flow in closed loop (CL) automation framework.
changes in network architecture are underway, leveraging Inspired by the MAPE-K (Monitor-Analyze-Plan-Execute
automation techniques such as software-defined network- over a shared Knowledge) framework, an intent ontology
ing and network function virtualization [3], [4]. describes different components such as metrics, action,
IBN has been met with significant interest as a pi- goals. Dzeparoska et al. [7] utilize a knowledge base of
oneering approach to network automation and service known intents and their actions to query and train the
orchestration aiming to alleviate these challenges. The intent translation model for network policy generation.
internet research task force (IRTF) network management Wang et al. [8] explore different sources of information
research group (NMRG) has made strides toward estab- in a vertical industry that can be incorporated in a KG
lishing a standardized understanding of “intent” within and propose a KG construction scheme using these sources
the IBN framework [5]. They describe an intent as a of information. Daroui et al. [9] focus on developing a
distributed knowledge base concept concerning different
deployable intents in the network and propose a solu-
tion for state management of these distributed knowledge
bases. However, there is still a lack of literature regarding
integrating KGs in the intent life cycle for aiding intent
translation and deployment phases.
The contributions of this paper are as follows:
• We design a knowledge graph storage solution by
embedding the available RDF triples as projections
in Gaussian space. This operation provides a trained
KGE model for the completion and verification of
intents during the intent processing and translation.
• We propose an intent processing pipeline with active
reporting and planning modules to adapt to the
changing network state during the life cycle of a
deployed intent. The pipeline uses the proposed IKG
and a trained KGE model to map intents to respective
services for deployment.
• For validation of the proposed pipeline, we design and Figure 1: An Intent Knowledge Graph (IKG) for IBN [13].
analyze extensive experimental campaigns covering
several boundary conditions and challenges encoun- offerings, resource types, and associated KPIs for the
tered in the simulated environment. defined services. For example, a triple (GBR, rdfs:subclass,
The rest of the paper is organized as follows: Section NonMcpttGBRService) represents a subclass of guaran-
II describes the proposed pipeline’s components and their teed bit rate (GBR) resource type, i.e., non-mission critical
integration with knowledge. Section III presents the eval- GBR service. Furthermore, the triple (PropertyParameter,
uation results. Section IV concludes the paper. rdfs:subclass, latency) represents a KPI parameter known
as ‘latency’ for the defined services.
II. Proposed Intent Processing Pipeline with
1) A Simple Knowledge Graph Model: A KG G is de-
Knowledge Graph Embedding
scribed as the following formulation,
This section first explains the key components of the
proposed pipeline — intent processing and KGE learning. G = (E, R, F ) (1)
Next, it presents the proposed pipeline for integrating
context as a knowledge graph in the intent life cycle. The where E = {e1 , e2 , e3 , ..., e|E| } is the set of entities (sub-
relationships between intents, services, and associated key ject or object), R = {r1 , r2 , r3 , ..., r|R| } is the set of rela-
performance indicators (KPIs) follow the service orches- tions, and F ⊆ E × R × E is the set of fact triples and
tration concept in 5G and B5G networks. each triple is denoted as (h, r, t). Here, h and t are head
and tail entities, and r denotes their relation. An entity is
A. Integration of Knowledge Graph in Intent Processing an object that can be classified as a class, type, or literal.
KGs consist of triples describing facts arranged in a Class objects utilize the resource description framework
systematic language using a well-known structure such schema (RDFS) to differentiate types of entities as distinct
as OWL [10] and RDF [11]. Intents also have a specific classes in a KG. Type object is used to identify class
structure with information fields describing a particular relationships as rdf:type objects; for example, class1 and
characteristic of the application use case. For example, class2 have similar properties, and they are of the same
“IntentA has an expectation Exp1” is represented as the type, that is, class A. A literal object is based on rdf:literal
triple (IntentA, hasExpectation, Exp1). In this way, any objects containing specific values as strings or integers. An
data object can be represented using a KG in the form of example of a KG is depicted in Figure 2.
triples. Hence, we can model an intent as a KG with dif- 2) Modeling Intents from IKG: In the IKG, one
ferent entities and relationships between them as depicted of the intent triples is (icm:PropertyExpectation,
in the intent common model (ICM) by Telemanagement icm:hasParameter, icm:PropertyParameter) representing
Forum (TMForum) [12]. In this paper, we utilize the IKG, the relationship between an expectation and its associated
given in Figure 1, that is based on an extension of the parameter (i.e., “PropertyExpectation has a Parameter
TMForum’s ICM [13]. that is PropertyParameter1 ”). Here, the intent triple is
The service, resource, and KPI extension models in the complete and utilized to process an intent.
IKG are essential components for defining service require- In contrast, the triple (icm:Expectation, icm:hasTarget,
ments in a service template. These components in the icm:Target) shows the relationship (i.e., “Expectation has
IKG represent information relevant to the available service a Target which is Target1 ”) between the intent expectation
 Figure 2: An example of a KG.

Figure 3: An incomplete service intent template.

and the required target to be reached. For example, the
intent translation process updates the information regard-
ing the expected target and required service, and KPI h ∼ N (µh , Ch )
parameters are appended from the IKG to complete the r ∼ N (µr , Cr ) (2)
designed intent model. This is accomplished via a series of t ∼ N (µt , Ct )
trained models using knowledge embedding [14] to predict
unknown entities in intent triples. where {µh , µr , µt } ∈ Rd and {Ch , Cr , Ct } ∈ Rd×d are
In other words, a service intent template is defined using mean vectors and covariance matrices, respectively.
the IKG, and the incomplete information, represented as 4) Training of KG Embedding Model: The KG2E model
‘???’ in Figure 3, is completed by augmenting information is trained on the IKG to minimize the scoring function,
related to service, resource, and expected KPI from the aiming to reduce the distance between the distributions of
IKG in Figure 1 using KGE learning. related entities and relations. This makes the embeddings
of correct triples more similar than those of incorrect
3) Knowledge Graph Embedding: KG analytics and rea-
triples. The scoring function fr (h, t) is defined to mea-
soning are enhanced by mapping knowledge into a vector
sure the correctness of the fact (h, r, t) in the embedding
space. This involves three steps: translating entities and
space. The scoring methods in KG2E [14] are modeled
relations into the vector space, establishing a scoring
as asymmetric Kullback-Leibler (KL)-divergence and sym-
function to evaluate their relationships, and undergoing
metric expected likelihood, with the scoring function being
a training process to refine the representations of entities
categorized as translational distance-based. The scoring
and relations.
function is defined using the probability inner product as:
Different techniques map knowledge and context from R
fr (h, t) = Nx (µt − µh , Ct + Ch ) . Nx (µr , Cr )dx
natural language to another latent space [15]. However, a (3)
∝ −µT C −1 µ − ln(det(C))
Gaussian embedding (GE) model effectively combines the
where µ = µh − µr − µt and C = Ch + Cr + Ct . Here,
mapping of various models such as DistMult, ComplexIE,
µ is the difference between mean vectors and captures
and TransE. Moreover, the robustness, scalability, and
the relationship between the head h and tail t entities
ability to support domain-specific KGs allow a GE-based
using relation r. It represents the relational embedding
model to suit appropriately with the proposed IKG and its
for (h, r, t) and µ encodes the direction and magnitude
extended version [15]. Therefore, we adopt a translational
of change from h to t via relation r.
distance model based on GE [14], which treats entities
The training uses the KG2E [14], where each triple
and relations as random vectors drawn from Gaussian
is evaluated to represent entities and relations in vector
distributions. This allows the model to capture entities’
space accurately. Open world assumption (OWA) is used
and relations’ central tendency (mean) and uncertainties
to prepare the dataset for the training algorithm. The KG
(covariance).
dataset for the learning algorithm is defined as the positive
Given a triple (h,r,t), knowledge graph with Gaussian triples F + = (h, r, t) and negative triples F − = (h′ , r,′ t′ ).
embedding (KG2E) models each entity and relation as a Here, the negative triples are constructed by randomly
Gaussian distribution with parameters (µ, C), where µ is choosing head and tail entities from the positive triples.
the mean vector and C is the covariance matrix. These
vectors are distinct in their semantic meanings and vary B. The Proposed Intent Processing Pipeline
in uncertainty compared to other entities with analogous We next present the pipeline for utilizing KGE learning
semantics. For any given entity or relation, the average in the intent processing life cycle with active control of the
of its vector embedding marks the core of its semantic deployed services (Figure 4).
interpretation, while the covariance matrix indicates the The different stages in the translation and mapping
degree of uncertainty. The distributions for entities and of user intents to orchestrated service deployment are
relations are denoted as: described as follows:
 Figure 4: Proposed intent processing pipeline with KGE.

Table I: Trained KGE model performance.

• Intent expression and recognition (A);
Link Prediction Metrics Classification Metrics
• Creation of service intent template (B);
• Identification of incomplete intent triples (C);
# of triples 1575 accuracy 71.66
Mean Rank 4.5 f1 score 0.9830
• Service predictions from trained KGE model (D); hits@1 0.58823 TPR 0.75722
• Completion of service intent template (E); hits@3 0.76470 TNR 0.98301
• Verification of intent as network intent (F).
hits@5 0.82352 FPR 0.24271
hits@10 0.91176 FNR 0.01698
Step A involves identifying and extracting named enti-
ties from the user’s intent, which delineates the search pa- of NER is implemented using spacy 2 library with a custom
rameters and scope for subsequent phases of intent trans- word corpus consisting of key terms relevant to the context
lation. Subsequently, Step B creates an intent template and expression style of the intent stakeholders. A public
structured according to a hierarchical model, leveraging version of the intent processing pipeline is available via
the intent ontology stipulated by TMForum [16]. However, github3 .
missing details concerning available resources, mapped
A. KGE Model Training (KG2E)
service offerings, and KPIs remain absent from the ser-
vice intent template. Step C is devoted to identifying The training for the KGE model is performed with 50
incomplete triples within the service intent, requiring sup- epochs and an adaptive learning rate. In addition, this
plementation with service-specific data. Step D employs model uses a root mean square optimizer and calculates
link prediction techniques facilitated by a trained KGE the distance measure based on KL-divergence. A total
model to infer resource, parameter, and KPI values. The of 1575 triples (both negative and positive) are utilized
link prediction task provides a set of p predictions with for training with a (0.8, 0.1, 0.1) split for train, test,
the highest scores. The contextual information from the and validation sets. The performance of the training loss
named entity recognition (NER) and the list of predictions minimization is depicted in Figure 5 with convergence
are used to populate the missing information in the intent after 12 optimization epochs.
triples. In Step E, the predicted entities from the KG
embedding are incorporated and expanded within the
service intent, generating an unverified rendition of the
network intent. Although comprehensive, this state of the
intent necessitates verification, accomplished through a
triple classification procedure using the pre-trained KGE
model. Finally, Step F entails intent verification, whereby
modified triples are evaluated based on the scoring func-
tion defined in Eq. (3). The intent verification is a triple
classification task within the KGE modeling. It classifies
the complete intent triples from the service intent as valid
or invalid based on the pre-defined scoring function. At
this point, the validated network intent encapsulates all
requisite information for deploying the requested service,
as ascertained by the intent processing pipeline.
III. Performance Evaluation
Figure 5: KG2E Model Training Convergence
The intent processing pipeline is implemented using
python and the pykeen 1 package. In addition, the process
2 https://spacy.io/api/entityrecognizer
1 https://github.com/pykeen/pykeen 3 https://github.com/kashifme224/kg-embedding-pykeen
 Figure 6: Service intent completion and validation with link prediction and triple classification.

B. Trained KGE Model Performance 2) Intent Verification: We use KGE model’s triple clas-
sification protocol to verify the validity of the intents using
a classification-based evaluator. We first define basic terms
The performance of the trained KGE model for service
for the performance metrics in our current scenario as:
predictions for intent-to-service mappings during the in-
• True positive (TP): A correctly predicted intent as
tent translation process is evaluated using typical KGE
valid is also valid in the KG.
metrics [17] as shown in Table I.
• True negative (TN): A correctly predicted intent as
1) Service Predictions: We use KGE link predictions to invalid is also invalid in the KG.
perform service predictions and measure their performance • False positive (FP): An incorrectly predicted intent
with a rank-based evaluator. The rank-based evaluation as valid is invalid in the KG.
protocol performs two link prediction tasks: 1) right-side: • False negative (FN): An incorrectly predicted intent
a pair of the head entity and relation is used to predict the as invalid is valid in the KG.
tail entity (h,r,?), 2) left-side: a pair of the tail entity This type of evaluation performs the classical sensitiv-
and relation is used to predict the head entity (?,r,t). The ity and specificity analysis for prediction tasks using a
scores are calculated using the scoring function in Eq. standard set of metrics such as 1) Accuracy: ratio of
(3) for both prediction tasks, and the scores are sorted the number of correct classifications to the total number,
in the order of decreasing score to determine the rank of 2) f1-score: harmonic mean of precision calculated as
the true choice (from the KG). The rank is the index in T P /(T P + F P ) and recall calculated as T P /(T P + F N ), 3) True

the sorted list of scores. The evaluation metrics for rank- positive rate (TPR): the probability of correctly predicting
based analysis are as follows: 1) Mean Rank: the average a valid intent, 4) True negative rate (TNR): the probability
rank of the correct entities and relations, 2) Hits@p: the of correctly predicting an invalid intent, 5) False positive
proportion of valid entities or relations ranked in top p rate (FPR): the probability of incorrectly predicting a
predictions. Each service prediction task provides a list of valid intent, 6) False negative rate (FNR): the probability
candidate p predictions to complete the incomplete triples. of incorrectly predicting an invalid intent.
These predictions are analyzed according to their score
and given context for an intent to select a valid completion C. An Example Intent Processing and Translation Flow
combination for the incomplete triples. A good embedding An example flow of the intent processing pipeline is
model should be able to have a low mean rank and a high shown in Figure 6. The process starts with identifying
proportion of hits for a given value of p. service keywords (‘reliable’ and ‘video’) from the natural
language user intent expression (Step A). A blueprint intents, service dependencies, and resource relationships.
service intent template is created and populated with The incorporation of KGE models facilitates the extrac-
relevant complete and incomplete triples from the IKG tion of meaningful insights from these knowledge graphs,
(Step B). Afterward, the recognized service keywords are enabling accurate prediction and completion of intent
used for getting and selecting relevant service predictions triples within the IBN pipeline.
from the trained KGE model for service intent completion
References
(Steps C, D). The service intent completion (Step E)
produces a service intent with all triples in a completed [1] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and
M. Ayyash, “Internet of things: A survey on enabling tech-
state. Intent verification classifies the service intent as nologies, protocols, and applications,” IEEE Communications
valid or invalid based on the IKG (Step F). Surveys & Tutorials, vol. 17, no. 4, pp. 2347–2376, 2015.
1) Service Intent Completion: Incomplete triples are [2] D. Kreutz, F. M. V. Ramos, P. E. Verı́ssimo, C. E. Rothenberg,
S. Azodolmolky, and S. Uhlig, “Software-defined networking: A
supplemented with missing tail entries by leveraging the comprehensive survey,” Proceedings of the IEEE, vol. 103, no. 1,
available head and relation entities. For instance, the triple pp. 14–76, 2015.
(icm:PropertyExpectation, icm:hasTarget, ???) identifies [3] K. Mehmood, K. Kralevska, and D. Palma, “Intent-driven au-
tonomous network and service management in future cellular
its pertinent prediction of nonmcptt:ConvVideo having the networks: A structured literature review,” Computer Networks,
highest prediction score (-180.9916) from the pool of vol. 220, p. 109477, 2023.
predictions. Each incomplete intent triple determines the [4] A. Leivadeas and M. Falkner, “A survey on intent-based net-
working,” IEEE Communications Surveys & Tutorials, vol. 25,
service, resource, and parameter necessary for completing no. 1, pp. 625–655, 2023.
the service intent template. The hierarchical predictions [5] A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura,
for each component are determined based on the scoring “Intent-Based Networking - Concepts and Definitions.” RFC
9315, Oct. 2022.
function delineated in Eq. (3). Furthermore, the intent [6] A. Randles, D. O’Sullivan, J. Keeney, and L. Fallon, “Ontology
verification process (Step F) within the intent processing driven closed control loop automation,” in 2023 IEEE 9th
pipeline enables the filtering of suitable predictions from International Conference on Network Softwarization (NetSoft),
pp. 528–533, 2023.
the prediction pool. [7] K. Dzeparoska, A. Tizghadam, and A. Leon-Garcia, “Emer-
2) Verified Network Intent: The integration of the gence: An intent fulfillment system,” IEEE Communications
intent processing and translation pipeline is enhanced Magazine, pp. 1–6, 2024.
[8] D. Wang, S. Zhang, R. Su, H. Li, and X. Xia, “An intent-based
by incorporating an intent verification step, facili- network empowered by knowledge graph: Enhancement of intent
tated by the concept of triple classification [18] uti- translation and management function for vertical industry,” in
lizing a trained KGE model. To illustrate, consider 2023 IEEE/CIC International Conference on Communications
in China (ICCC Workshops), pp. 1–6, 2023.
the scenario depicted in Figure 6, where the incom- [9] D. Daroui, M. D’Angelo, L. Mokrushin, M. Orlić, and A. C.
plete triple (nonmcptt:ConvVideo, icm:targetResource, Baktir, “State management of knowledge base in intent manage-
???) is supplemented by the top-ranked prediction (ser- ment functions in 6g networks,” in Int. Balkan Conference on
Communications and Networking (BalkanCom), pp. 1–6, 2023.
vice:NonMcpttGBRService) with the highest prediction [10] T. W. W. W. C. (W3C), “Owl 2 web ontology language;
score (-134.3381). However, it is worth noting that the structural specification and functional-style syntax,” technical
prediction (service:McpttGBRService) also warrants vali- report, 2012.
[11] T. W. W. W. C. (W3C), “Rdf 1.1 primer,” technical report,
dation, as it possesses the second-highest prediction score 2014.
(-140.0011) among the available recommendations, and it [12] TMForum, “Intent common model v2.1.0,” Technical Report
is also a valid completion for the incomplete intent triple. TR292, TMForum, 2022.
[13] K. Mehmood, K. Kralevska, and D. Palma, “Knowledge-based
Conversely, the remaining predictions fail to adhere to the intent modeling for next generation cellular networks,” in 2023
fundamental ontology of the knowledge graph, rendering IEEE International Mediterranean Conference on Communica-
them ineligible for validation. Consequently, an alternate tions and Networking (MeditCom), pp. 181–186, 2023.
[14] S. He, K. Liu, G. Ji, and J. Zhao, “Learning to represent
prediction is sought to ensure the selection of a valid knowledge graphs with gaussian embedding,” in Proceedings of
intent triple completion. Finally, the incomplete triple the 24th ACM International on Conference on Information and
(kpi:latency, icm:valueBy, ???) is completed with the top- Knowledge Management, CIKM ’15, (New York, NY, USA),
p. 623–632, Association for Computing Machinery, 2015.
ranked prediction (150msˆˆxsd:string) with the highest [15] Y. Lin, X. Han, R. Xie, Z. Liu, and M. Sun, “Knowledge
score of (-92.0514). The resultant completed and verified Representation Learning: A Quantitative Review,” Dec. 2018.
service intent is referred to as the network intent, per the arXiv:1812.10901 [cs].
[16] TMForum, “Intent management ontology v1.1.0,” Technical
terminology established by the TMForum [12], encompass- Report TR290, TMForum, 2022.
ing all requisite service parameters for deployment within [17] C. T. Hoyt, M. Berrendorf, M. Galkin, V. Tresp, and B. M.
the underlying network infrastructure. Gyori, “A unified framework for rank-based evaluation metrics
for link prediction in knowledge graphs,” 2022.
[18] S. Ji, S. Pan, E. Cambria, P. Marttinen, and P. S. Yu, “A survey
IV. Conclusion on knowledge graphs: Representation, acquisition, and applica-
In this paper, we have explored the utilization of KGE tions,” IEEE Transactions on Neural Networks and Learning
Systems, vol. 33, no. 2, pp. 494–514, 2022.
and its potential application for intent processing and
translation for IBN. We have highlighted the significance
of KGs in capturing and representing complex network