Page 174
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Link-Spam-Resistant Domain Ranking: A Comparative Evaluation
of PageRank, HITS, and TrustRank Under Simulated Link-Farm
Attacks Using Common Crawl Web Graphs
Romelyn J. Banaybanay
1
, Reagan B. Ricafort
2
1
Initao College, Initao, Misamis Oriental, Philippines
2
AMA University, Makati City, Philippines
DOI: https://doi.org/10.51583/IJLTEMAS.2026.150600016
Received: 17 June 2026; Accepted: 22 June 2026; Published: 03 July 2026
ABSTRACT
Link-based ranking is still an important part of web information retrieval, but it remains vulnerable to
coordinated link-farm attacks. This study examined the resistance of PageRank, HITS, and TrustRank to link
spam using simulated attacks on domain-level web graphs extracted from Common Crawl. A quantitative,
comparative, and controlled experimental design was used. Five independent base graphs were prepared, with
each graph containing 100,000 legitimate domain nodes. Four link-farm structures were tested: star-centered,
bounded-density clique, bipartite, and mixed. The attacks were applied at four intensity levels: 1%, 3%, 5%, and
10% of the legitimate graph size. Across all conditions, the experiment produced 480 attacked graph instances
and 1,440 verified algorithm-level runs. The algorithms were compared using spam target promotion, spam score
share, top-1000 spam infiltration, legitimate-domain displacement, ranking stability, execution time, peak
memory use, and iteration count. The results showed that PageRank was the most vulnerable algorithm. It had
the highest spam target promotion, highest spam score share, measurable top-1000 spam infiltration, greater
legitimate-domain displacement, and lowest ranking stability. HITS showed the strongest overall resistance and
computational efficiency. It recorded near-zero spam score share, zero top-1000 infiltration, zero legitimate-
domain displacement, highest stability, fastest runtime, and fewest iterations. TrustRank also blocked top-1000
infiltration and legitimate-domain displacement, although it allowed higher target promotion than HITS. The
bounded-density clique structure caused the most damage to PageRank, while higher attack intensity increased
spam impact in most cases. The findings show that link-spam resistance depends on the ranking algorithm, attack
structure, attack intensity, and evaluation metric used. A single ranking score or attack model is not enough to
evaluate resistance to link-farm manipulation.
Keywords: Web information retrieval, link spam, PageRank, HITS, TrustRank
INTRODUCTION
Web information retrieval needs ranking algorithms to decide which pages or domains appear first in search
results. Link-analysis methods use hyperlinks as signs of authority, popularity, or trust. PageRank, HITS, and
TrustRank offer three related but different ways to read link structure. PageRank estimates importance through
a random-walk model. HITS separates hub and authority scores. TrustRank limits rank propagation by starting
from trusted seed nodes. Recent work on sparse, dense, contextual, and late-interaction retrieval shows that
ranking remains an active research area [5, 6, 8, 15, 25]. Studies on exact lexical matching, dense passage
retrieval, sequence-to-sequence ranking, and reproducible IR tools also support the need for reliable ranking
methods [9, 14, 16, 21, 23]. Graph-based recommendation, knowledge graph, link-prediction, and graph neural
network studies show that graph structure remains useful for ranking and relational analysis [7, 11, 12, 24, 28].
Link-analysis methods are useful, but they are not safe from manipulation. Link spam happens when artificial
links are created to raise the rank of a target page or domain. A link farm is one example. It uses a group of spam
nodes to pass ranking influence to one or more targets. This behavior can promote low-quality or harmful
Page 175
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
domains and push legitimate sources downward. Recent work on spam and web-spam detection supports the
need to test ranking systems under adversarial conditions [1, 2, 13, 20, 26]. Fraud and graph-anomaly studies
also show that coordinated or camouflaged graph behavior can distort system outputs [3, 17, 18, 22].
This study responds to that problem by comparing PageRank, HITS, and TrustRank under controlled link-farm
attacks. It uses domain-level Common Crawl graphs and synthetic attack structures instead of surveys or private
search-engine logs. The aim is not to claim that the simulated attacks represent all real spam behavior. Rather,
the aim is to test how each algorithm behaves under clear, repeatable, and measurable attack settings.
The main research question was: How do PageRank, HITS, and TrustRank differ in their resistance to simulated
link-farm attacks across attack structures and intensities using domain-level Common Crawl web graphs? The
study measured differences in spam target promotion, spam score share, top-k spam infiltration, legitimate-
domain displacement, ranking stability, execution time, memory use, and convergence behavior.
The following null hypotheses were tested:
H01: There is no significant difference among PageRank, HITS, and TrustRank in spam target promotion.
H02: Link-farm structure does not significantly affect top-k spam infiltration.
H03: Attack intensity does not significantly affect legitimate-domain displacement.
H04: There is no significant algorithm-by-structure interaction for spam target promotion.
H05: There is no significant algorithm-by-structure-by-intensity interaction for ranking stability.
H06: There is no significant difference among PageRank, HITS, and TrustRank in execution time.
H07: There is no significant difference among PageRank, HITS, and TrustRank in peak memory consumption.
METHODOLOGY
Research Design
The study used a quantitative, comparative, and controlled experimental design. It compared PageRank, HITS,
and TrustRank under simulated link-farm attacks. The study did not involve human participants, surveys,
interviews, questionnaires, or confidential records. The independent variables were ranking algorithm, link-farm
structure, and attack intensity. The dependent variables were spam target promotion, spam score share, top-1000
spam infiltration, legitimate-domain displacement, Spearman ranking stability, rank-biased overlap, execution
time, peak memory consumption, and iteration count.
Data Source and Unit of Analysis
Domain-level web graph data from Common Crawl were used. In the graph, each node represented a pay-level
domain and each directed edge represented a hyperlink relationship between two domains. An edge from domain
u to domain v meant that at least one page from u linked to a page from v. The pay-level domain was the unit of
analysis.
Graph Sampling
Five independent base graphs were prepared. Each base graph contained 100,000 legitimate domain nodes.
Direction-aware random-walk-with-restart sampling preserved connected web-graph neighborhoods while
keeping the experiment manageable. After sampling, self-loops and duplicate directed edges were removed. The
cleaned graphs were used as the base testbeds.
Page 176
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
TrustRank Seed Selection
TrustRank needed trusted seed domains. Archived Tranco domain rankings were used as a reproducible proxy
for trusted seeds. For each base graph, the highest-ranked Tranco domains that also appeared in the sampled
graph were selected. Synthetic spam nodes were not allowed in the seed set.
Ranking Algorithms
PageRank was run using a random-surfer model with a damping factor of 0.85. HITS was run using authority
and hub updates over the directed adjacency matrix, and authority scores served as the main ranking output.
TrustRank was run as seed-personalized PageRank. All algorithms used sparse matrix representations in the
same computing environment. This workflow followed common scientific computing and reproducible IR tool
practices [10, 19, 27].
Link-Farm Attack Model
Each attack included one target spam domain and a set of supporting spam domains. Four link-farm structures
were tested: star-centered, bounded-density clique, bipartite, and mixed. Each attack used a fixed spam-
originated edge budget of 10m, where m was the number of synthetic spam nodes. External ingress links were
added from legitimate non-seed domains.
Attack Intensities and Replication
Four attack intensities were tested: 1%, 3%, 5%, and 10% of the 100,000 legitimate nodes in each base graph.
Each structure-intensity-base graph condition was repeated six times using independent attack realizations. The
final experiment produced 480 attacked graph instances and 1,440 algorithm-level runs.
Statistical Analysis
Descriptive statistics were calculated for each algorithm, structure, and intensity. Friedman tests were used for
paired nonparametric comparisons among algorithms. Pairwise Wilcoxon signed-rank tests with Holm
adjustment were used for post hoc analysis. Structure and intensity effects were examined by algorithm.
Interaction effects were tested using rank-based interaction ANOVA, following current procedures for
multifactor rank-based analysis [4]. The statistical workflow used Python-based scientific computing and
reproducible IR tools [10, 19, 27].
RESULTS
The experiment was completed successfully. It produced 480 attacked graph instances and 1,440 algorithm-level
metric records. All algorithm runs converged. The final master dataset had no missing values and no duplicate
experimental rows. It was balanced across algorithms, structures, intensities, and base graphs.
Across all attack conditions, PageRank had the highest mean target promotion at 0.930980. TrustRank followed
at 0.753470, while HITS had the lowest mean target promotion at 0.186051. PageRank also had the highest spam
score share, measurable top-1000 spam infiltration, greater legitimate-domain displacement, and the lowest
ranking stability.
Table 1. Main descriptive comparison of ranking algorithms.
Metric
HITS
PageRank
TrustRank
Mean Target Promotion
0.186051
0.930980
0.753470
Mean Spam Score Share
0.000003
0.032597
0.000058
Page 177
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Mean Infiltration@1000
0.000000
0.007198
0.000000
Mean P95 Legitimate Displacement
0.000000
9.173438
0.000000
Mean Spearman Stability
0.998953
0.906140
0.990426
Mean Rank-Biased Overlap
1.000000
0.972289
0.999985
Mean Execution Time (seconds)
5.005167
7.988084
7.727799
Mean Peak Memory (GiB)
0.729034
0.729571
0.731206
Mean Iterations
31.800000
98.964583
96.952083
Friedman tests showed significant algorithm effects for all tested metrics: target promotion, spam score share,
top-1000 spam infiltration, P95 legitimate displacement, Spearman stability, RBO, execution time, peak memory
consumption, and iterations. Pairwise Wilcoxon tests with Holm adjustment showed that HITS had lower target
promotion than PageRank and TrustRank. PageRank had higher target promotion than TrustRank.
Link-farm structure affected the algorithms in different ways. For PageRank, the bounded-density clique was
the most damaging structure. It had the highest spam score share, highest top-1000 infiltration, highest P95
displacement, lowest Spearman stability, and lowest RBO. For HITS, the star-centered structure produced the
highest target promotion, but HITS still had zero top-1000 infiltration and zero P95 displacement across all
structures. For TrustRank, the bipartite structure had the highest target promotion, while the clique structure had
the highest spam score share and lowest stability.
Higher attack intensity generally increased spam impact. For PageRank, higher intensity led to stronger spam
score share, higher infiltration, greater displacement, and lower ranking stability. HITS and TrustRank also
showed increases in target promotion and spam score share, but both kept zero top-1000 infiltration and zero
P95 legitimate displacement at all intensity levels.
The Algorithm x Structure interaction for target promotion was significant, F = 114.918537, p = 2.814419e-118,
partial eta squared = 0.327479. For ranking stability, the Algorithm x Structure x Intensity interaction was
significant when Spearman stability was used, F = 65.500370, p = 1.949972e-170, partial eta squared =
0.460189. The same three-way interaction was not significant when RBO was used, F = 0.873179, p = 0.612166,
partial eta squared = 0.011237.
Table 2. Final hypothesis-decision summary.
Hypothesis
Decision
Interpretation
H01
Reject H01
Algorithms differed significantly;
HITS had the lowest target
promotion, PageRank the highest,
Trust Rank intermediate.
H02
Partially reject H02
Structure affected spam infiltration
under PageRank; HITS and Trust
Rank produced tied zero infiltration.
H03
Partially reject H03
Intensity affected displacement
under PageRank; HITS and Trust
Rank produced tied zero
displacement.
H04
Reject H04
Algorithm x Structure interaction
was significant: F=114.918537,
Page 178
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
p=2.814419e-118, partial eta
squared=0.327479.
H05
Partially reject H05
Three-way interaction was
significant for Spearman Stability
but not significant for RBO.
H06
Reject H06
Execution time differed
significantly among algorithms;
HITS was fastest.
H07
Reject H07
Peak memory differed significantly,
although practical differences were
small.
DISCUSSION
PageRank, HITS, and TrustRank responded differently to the simulated link-farm attacks. PageRank was the
most vulnerable. HITS showed the strongest overall resistance. TrustRank showed strong but selective
resistance. These results reflect the different assumptions of the algorithms. PageRank spreads importance
through links. HITS separates hub and authority behavior. TrustRank limits influence through trusted seed
personalization. Related graph-based studies also show that recommendation, graph convolution, knowledge
representation, and graph learning methods depend on relational structure [7, 11, 12, 28, 30]. Link-prediction
research supports the same role of graph structure in relational inference and ranking behavior [24].
PageRank was vulnerable because it depends on link-based score propagation. When spam nodes are
coordinated, they can pass ranking influence to a target. In this study, the clique structure caused the most damage
because spam nodes reinforced one another before passing influence outward. This led to higher spam score
share, higher infiltration, greater displacement, and lower stability. The finding is consistent with fraud and
graph-anomaly studies showing that coordinated graph behavior can distort detection and ranking outputs [3,
17, 18, 22]. It also supports web-spam studies that recommend combining link-based signals with content-based
and anomaly-based defenses [1, 20, 26].
HITS performed best overall. It stopped top-1000 spam infiltration and legitimate-domain displacement across
all structures and intensities. It also showed the highest stability and the best computational efficiency. The spam
target still gained some rank position under some structures, especially star-centered attacks, but the effect did
not spread into wider ranking pollution.
TrustRank also showed strong resistance. It stopped top-1000 infiltration and legitimate-domain displacement
across all structures and intensities. This result supports the use of seed-personalized ranking to limit spam
diffusion. Still, TrustRank allowed higher target promotion than HITS. This means that trust personalization
reduced broad spam spread but did not fully stop the target spam domain from improving its rank. This point
agrees with graph-based research that shows the need to manage trust, links, and relational structure carefully
[12, 24, 28, 30].
The results also show that attack structure matters. PageRank was most affected by clique structures. HITS had
its highest target promotion under star-centered attacks. TrustRank had different weaknesses under bipartite and
clique structures. These differences show that one attack model is not enough to evaluate link-spam resistance.
Recent studies on graph anomaly detection, web-spam detection, and imbalanced graph learning also support
structure-aware testing of adversarial graph behavior [18, 20, 26, 29].
Overall, link-spam resistance is not defined by one metric alone. PageRank was weak against coordinated
reinforcement. HITS was strongest overall. TrustRank contained broad spam spread but was less effective than
Page 179
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
HITS in suppressing target promotion. These findings support multi-algorithm, multi-metric, and structure-
aware evaluation in web information retrieval.
CONCLUSION AND RECOMMENDATIONS
Conclusion
This study tested the link-spam resistance of PageRank, HITS, and TrustRank under simulated link-farm attacks
using domain-level Common Crawl web graphs. The experiment produced 480 attacked graph instances and
1,440 verified algorithm-level runs. The results showed clear differences among the three algorithms.
PageRank was the most vulnerable algorithm. It had the highest spam target promotion, highest spam score
share, measurable top-1000 spam infiltration, greater legitimate-domain displacement, and lowest ranking
stability. The bounded-density clique structure and higher attack intensities caused the most damage to
PageRank.
HITS showed the strongest overall resistance and computational efficiency. It had the lowest target promotion,
lowest spam score share, zero top-1000 infiltration, zero legitimate-domain displacement, highest ranking
stability, fastest execution time, and fewest iterations. TrustRank also performed well in preventing top-1000
infiltration and legitimate-domain displacement, but it allowed higher target promotion than HITS.
The study concludes that link-spam resistance depends on algorithm design, attack structure, attack intensity,
and the metric used to measure spam success. HITS was strongest overall. TrustRank limited broad spam spread.
PageRank was most vulnerable to link-farm manipulation.
Recommendations
Search systems should not rely on PageRank alone when link manipulation is possible. PageRank should be
combined with trust signals, spam-pattern detection, link-anomaly detection, content-quality signals, and
temporal link-growth monitoring. Spam and web-spam studies support the use of both content-based and link-
based anti-spam defenses [1, 13, 20, 26]. Fraud and anomaly detection studies also support graph-aware defenses
against coordinated behavior [3, 17, 18, 22].
Systems that need stronger resistance to link-farm attacks should include HITS-like authority and hub analysis
as part of a broader ranking model. HITS should still be combined with other anti-spam methods because target
promotion can still occur under some structures. Graph-based recommendation, graph convolution, knowledge
graph, and graph neural network studies support the use of relational structure in ranking and detection models
[7, 11, 12, 28, 30]. Link-prediction research also supports the value of modeling relationships among connected
entities [24].
TrustRank or seed-personalized ranking should be used when the goal is to limit spam influence from untrusted
graph regions. Careful seed selection, periodic seed validation, and added spam-detection features are still
needed to reduce target-level gains. This recommendation fits graph-based research showing that link structure,
knowledge representation, and graph neural methods depend on the quality of relational signals [12, 24, 28, 30].
Future evaluations should use several metrics, link-farm structures, and attack intensities. Future work should
extend the experiment to page-level graphs, labeled spam datasets, temporal spam patterns, and hybrid ranking
models. Future studies should compare link-based ranking with sparse, contextual, dense, and exact-matching
retrieval approaches under adversarial settings [5, 6, 8, 9, 14]. Late-interaction ranking should also be tested in
spam-resistant retrieval settings [15, 25]. Reproducible IR toolkits, sequence-to-sequence ranking, optimized
dense retrieval, and imbalanced graph learning should be included in future comparisons [16, 21, 23, 29].
Page 180
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
LIMITATIONS
This study used simulated link-farm attacks instead of naturally labeled spam networks. The findings therefore
measure resistance to the attack models used in the experiment. This matters because real spam behavior can
include email, social, content, review, and web-spam patterns that were not fully represented by the simulated
link farms [1, 2, 13, 20, 26]. Real adversarial behavior can also include fraud and graph-anomaly patterns that
were not directly modeled [3, 17, 18, 22].
The study used domain-level graphs instead of page-level graphs. Domain-level aggregation supports large-scale
analysis, but it removes page-level details such as anchor text, topical relevance, internal site structure, and link
placement. These details matter because lexical expansion, contextual matching, dense passage retrieval, and
late-interaction retrieval often work below the domain level [5, 6, 9, 14, 15]. Sequence-to-sequence ranking,
optimized dense retrieval, and lightweight late-interaction models also rely on document-level or text-level
signals [21, 23, 25].
The sampled graphs were controlled testbeds and should not be treated as complete representations of the web.
Random-walk sampling may give more weight to well-connected graph regions. Future replication should record
sampling settings, software versions, and computing tools to strengthen reproducibility [10, 16, 19, 27].
TrustRank used Tranco-based proxy seed domains. These seeds were reproducible, but popularity is not the
same as trustworthiness. Future work should compare different seed-selection methods and test whether trusted
seeds remain stable across time, domain categories, and graph regions [12, 18, 24, 28, 30].
Data and Reproducibility Statement
This study used public domain-level web graph data from the Common Crawl Host- and Domain-Level Web
Graphs October-November-December 2024 release. The domain-level graph was used to build five independent
legitimate-domain base graphs, each with 100,000 nodes. The original Common Crawl graph was not used as a
labeled spam dataset.
TrustRank seed domains were selected from archived Tranco ranking data as a reproducible proxy for trusted
domains. For each base graph, the highest-ranked Tranco domains found in the sampled graph were used as seed
nodes. Synthetic spam nodes were excluded from the seed set.
The experiment used controlled synthetic link-farm attacks so that testing could be repeated. Four attack
structures were simulated: star-centered, bounded-density clique, bipartite, and mixed. Four attack intensities
were tested: 1%, 3%, 5%, and 10% of the legitimate graph size. Each structure-intensity-base graph condition
was repeated using six independent attack realizations.
The reproducibility package contains the dataset source record, sampling notes, experimental pseudocode,
statistical output files, result screenshots, and appendix tables. These files document the data source, sampling
process, attack design, ranking workflow, statistical analysis, and reported results. Full reproduction requires
access to the cited public datasets and implementation of the documented experimental pipeline.
REFERENCES
1. Akinyelu, A. A. (2021). Advances in spam detection for email spam, web spam, social network spam,
and review spam: ML-based and nature-inspired-based techniques. Journal of Computer Security, 29(5),
473-529. https://doi.org/10.3233/JCS-210022
2. Alom, Z., Carminati, B., & Ferrari, E. (2020). A deep learning model for Twitter spam detection. Online
Social Networks and Media, 18, Article 100079. https://doi.org/10.1016/j.osnem.2020.100079
3. Dou, Y., Liu, Z., Sun, L., Deng, Y., Peng, H., & Yu, P. S. (2020). Enhancing graph neural network-based
fraud detectors against camouflaged fraudsters. Proceedings of the 29th ACM International Conference
on Information and Knowledge Management, 315-324. https://doi.org/10.1145/3340531.3411903
Page 181
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
4. Elkin, L. A., Kay, M., Higgins, J. J., & Wobbrock, J. O. (2021). An aligned rank transform procedure for
multifactor contrast tests. Proceedings of the 34th Annual ACM Symposium on User Interface Software
and Technology, 754-768. https://doi.org/10.1145/3472749.3474784
5. Formal, T., Lassance, C., Piwowarski, B., & Clinchant, S. (2022). From distillation to hard negative
sampling: Making sparse neural IR models more effective. Proceedings of the 45th International ACM
SIGIR Conference on Research and Development in Information Retrieval, 2353-2359.
https://doi.org/10.1145/3477495.3531857
6. Formal, T., Piwowarski, B., & Clinchant, S. (2021). SPLADE: Sparse lexical and expansion model for
first stage ranking. Proceedings of the 44th International ACM SIGIR Conference on Research and
Development in Information Retrieval, 2288-2292. https://doi.org/10.1145/3404835.3463098
7. Gao, C., Zheng, Y., Li, N., Li, Y., Qin, Y., Piao, J., Quan, Y., Chang, J., Jin, D., He, X., & Li, Y. (2023).
A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM
Transactions on Recommender Systems, 1(1), Article 3. https://doi.org/10.1145/3568022
8. Gao, L., & Callan, J. (2021). Condenser: A pre-training architecture for dense retrieval. Proceedings of
the 2021 Conference on Empirical Methods in Natural Language Processing, 981-993.
https://doi.org/10.18653/v1/2021.emnlp-main.75
9. Gao, L., Dai, Z., & Callan, J. (2021). COIL: Revisit exact lexical match in information retrieval with
contextualized inverted list. Proceedings of the 2021 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, 3030-3042.
https://doi.org/10.18653/v1/2021.naacl-main.241
10. Harris, C. R., Millman, K. J., van der Walt, S. J., Gommers, R., Virtanen, P., Cournapeau, D., Wieser,
E., Taylor, J., Berg, S., Smith, N. J., Kern, R., Picus, M., Hoyer, S., van Kerkwijk, M. H., Brett, M.,
Haldane, A., del Río, J. F., Wiebe, M., Peterson, P., Gérard-Marchant, P., Sheppard, K., Reddy, T.,
Weckesser, W., Abbasi, H., Gohlke, C., & Oliphant, T. E. (2020). Array programming with NumPy.
Nature, 585, 357-362. https://doi.org/10.1038/s41586-020-2649-2
11. He, X., Deng, K., Wang, X., Li, Y., Zhang, Y., & Wang, M. (2020). LightGCN: Simplifying and
powering graph convolution network for recommendation. Proceedings of the 43rd International ACM
SIGIR Conference on Research and Development in Information Retrieval, 639-648.
https://doi.org/10.1145/3397271.3401063
12. Ji, S., Pan, S., Cambria, E., Marttinen, P., & Yu, P. S. (2021). A survey on knowledge graphs:
Representation, acquisition, and applications. IEEE Transactions on Neural Networks and Learning
Systems, 33(2), 494-514. https://doi.org/10.1109/TNNLS.2021.3070843
13. Kaddoura, S., Chandrasekaran, G., Popescu, D. E., & Duraisamy, J. H. (2022). A systematic literature
review on spam content detection and classification. PeerJ Computer Science, 8, Article e830.
https://doi.org/10.7717/peerj-cs.830
14. Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W. (2020). Dense
passage retrieval for open-domain question answering. Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing, 6769-6781. https://doi.org/10.18653/v1/2020.emnlp-main.550
15. Khattab, O., & Zaharia, M. (2020). ColBERT: Efficient and effective passage search via contextualized
late interaction over BERT. Proceedings of the 43rd International ACM SIGIR Conference on Research
and Development in Information Retrieval, 39-48. https://doi.org/10.1145/3397271.3401075
16. Lin, J., Ma, X., Lin, S.-C., Yang, J.-H., Pradeep, R., & Nogueira, R. (2021). Pyserini: A Python toolkit
for reproducible information retrieval research with sparse and dense representations. Proceedings of the
44th International ACM SIGIR Conference on Research and Development in Information Retrieval,
2356-2362. https://doi.org/10.1145/3404835.3463238
17. Liu, Z., Dou, Y., Yu, P. S., Deng, Y., & Peng, H. (2020). Alleviating the inconsistency problem of
applying graph neural network to fraud detection. Proceedings of the 43rd International ACM SIGIR
Conference on Research and Development in Information Retrieval, 1569-1572.
https://doi.org/10.1145/3397271.3401253
18. Ma, X., Wu, J., Xue, S., Yang, J., Zhou, C., Sheng, Q. Z., Xiong, H., & Akoglu, L. (2023). A
comprehensive survey on graph anomaly detection with deep learning. IEEE Transactions on Knowledge
and Data Engineering, 35(12), 12012-12038. https://doi.org/10.1109/TKDE.2021.3118815
Page 182
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
19. Macdonald, C., & Tonellotto, N. (2020). Declarative experimentation in information retrieval using
PyTerrier. Proceedings of the 2020 ACM SIGIR International Conference on the Theory of Information
Retrieval, 161–168. https://doi.org/10.1145/3409256.3409829
20. Makkar, A., & Kumar, N. (2020). An efficient deep learning-based scheme for web spam detection in
IoT environment. Future Generation Computer Systems, 108, 467-487.
https://doi.org/10.1016/j.future.2020.03.004
21. Nogueira, R., Jiang, Z., & Lin, J. (2020). Document ranking with a pretrained sequence-to-sequence
model. Findings of the Association for Computational Linguistics: EMNLP 2020, 708-718.
https://doi.org/10.18653/v1/2020.findings-emnlp.63
22. Pang, G., Shen, C., Cao, L., & van den Hengel, A. (2021). Deep learning for anomaly detection: A review.
ACM Computing Surveys, 54(2), Article 38, 1-38. https://doi.org/10.1145/3439950
23. Qu, Y., Ding, Y., Liu, J., Liu, K., Ren, R., Zhao, W. X., Dong, D., Wu, H., & Wang, H. (2021).
RocketQA: An optimized training approach to dense passage retrieval for open-domain question
answering. Proceedings of the 2021 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, 5835-5847.
https://doi.org/10.18653/v1/2021.naacl-main.466
24. Rossi, A., Barbosa, D., Firmani, D., Matinata, A., & Merialdo, P. (2021). Knowledge graph embedding
for link prediction: A comparative analysis. ACM Transactions on Knowledge Discovery from Data,
15(2), Article 14. https://doi.org/10.1145/3424672
25. Santhanam, K., Khattab, O., Saad-Falcon, J., Potts, C., & Zaharia, M. (2022). ColBERTv2: Efficient and
effective retrieval via lightweight late interaction. Proceedings of the 2022 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies,
3715-3734. https://doi.org/10.18653/v1/2022.naacl-main.272
26. Shahzad, A., Nawi, N. M., Gillani, S. M. Z. R., & Khan, A. (2021). An improved framework for content-
and link-based web-spam detection: A combined approach. Complexity, 2021, Article 6625739.
https://doi.org/10.1155/2021/6625739
27. Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau, D., Burovski, E.,
Peterson, P., Weckesser, W., Bright, J., van der Walt, S. J., Brett, M., Wilson, J., Millman, K. J., Mayorov,
N., Nelson, A. R. J., Jones, E., Kern, R., Larson, E., Carey, C. J., et al. (2020). SciPy 1.0: Fundamental
algorithms for scientific computing in Python. Nature Methods, 17(3), 261-272.
https://doi.org/10.1038/s41592-019-0686-2
28. Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2021). A comprehensive survey on graph
neural networks. IEEE Transactions on Neural Networks and Learning Systems, 32(1), 4-24.
https://doi.org/10.1109/TNNLS.2020.2978386
29. Zhao, T., Zhang, X., & Wang, S. (2021). GraphSMOTE: Imbalanced node classification on graphs with
graph neural networks. Proceedings of the 14th ACM International Conference on Web Search and Data
Mining, 833-841. https://doi.org/10.1145/3437963.3441720
30. Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang, L., Li, C., & Sun, M. (2020). Graph neural
networks: A review of methods and applications. AI Open, 1, 57-81.
https://doi.org/10.1016/j.aiopen.2021.01.001
Page 183
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
APPENDICES
Appendix A. Experimental Design Summary
Appendix Table A1. Experimental design summary.
Component
Verified value
Base graphs
5 independent graphs
Legitimate nodes per base graph
100,000 domains
Algorithms
PageRank, HITS, TrustRank
Structures
Star, Clique, Bipartite, Mixed
Attack intensities
1%, 3%, 5%, 10%
Attack realizations
6 per structure-intensity-base graph condition
Attacked graph instances
480
Algorithm-level runs
1,440
Primary data source
Common Crawl domain-level web graph
Trust seed proxy
Archived Tranco ranking
Appendix B. Direction Table by Algorithm and Structure
Appendix Table B1. Mean direction by algorithm and link-farm structure.
Algorith
m
Structur
e
Target
Promotio
n
Spam
Score
Share
Infil@100
0
P95
Disp.
Spearma
n
RBO
Time
Memor
y
Iterations
HITS
Bipartite
0.138682
0.00000
3
0.000000
0.000000
0.999011
1.00000
0
4.93769
0
0.72936
7
31.800000
HITS
Clique
0.141947
0.00000
3
0.000000
0.000000
0.998585
1.00000
0
5.00012
3
0.72906
3
31.800000
HITS
Mixed
0.217248
0.00000
3
0.000000
0.000000
0.998532
1.00000
0
4.94923
0
0.72936
3
31.800000
HITS
Star
0.246326
0.00000
3
0.000000
0.000000
0.999684
1.00000
0
5.13362
6
0.72834
3
31.800000
PageRan
k
Bipartite
0.931403
0.02555
5
0.001000
1.000000
0.934824
0.95821
4
7.83539
6
0.72990
7
95.600000
PageRan
k
Clique
0.931355
0.07169
4
0.022917
22.71666
7
0.817822
0.94676
0
8.69789
1
0.72960
1
109.05833
3
PageRan
k
Mixed
0.930872
0.02086
5
0.003875
9.676250
0.886683
0.98927
0
7.72502
4
0.72990
8
95.600000
PageRan
k
Star
0.930290
0.01227
4
0.001000
3.300833
0.985231
0.99491
3
7.69402
7
0.72886
8
95.600000
TrustRan
k
Bipartite
0.820064
0.00004
3
0.000000
0.000000
0.995908
0.99999
2
7.77591
7
0.73146
7
96.800000
Page 184
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
TrustRan
k
Clique
0.802913
0.00013
0
0.000000
0.000000
0.976711
0.99999
2
7.71450
1
0.73122
4
97.408333
TrustRan
k
Mixed
0.715707
0.00003
7
0.000000
0.000000
0.990708
0.99997
8
7.67050
9
0.73145
3
96.800000
TrustRan
k
Star
0.675196
0.00002
1
0.000000
0.000000
0.998379
0.99997
6
7.75026
8
0.73068
1
96.800000
Appendix C. Direction Table by Algorithm and Attack Intensity
Appendix Table C1. Mean direction by algorithm and attack intensity.
Algorith
m
Intensit
y %
Target
Promotio
n
Spam
Score
Share
Infil@100
0
P95
Disp.
Spearma
n
RBO
Time
Memor
y
Iteration
s
HITS
1
0.162504
5.14E
-07
0.000000
0.000000
0.999841
1.00000
0
5.16220
7
0.72038
5
31.80000
0
HITS
3
0.169351
1.59E
-06
0.000000
0.000000
0.999459
1.00000
0
4.85836
5
0.72230
2
31.80000
0
HITS
5
0.185284
2.65E
-06
0.000000
0.000000
0.999002
1.00000
0
5.00836
9
0.72599
9
31.80000
0
HITS
10
0.227063
5.45E
-06
0.000000
0.000000
0.997511
1.00000
0
4.99172
8
0.74745
0
31.80000
0
PageRank
1
0.945636
7.39E
-03
0.002883
3.266667
0.976512
0.99553
9
8.08467
0
0.72080
5
98.05833
3
PageRank
3
0.938031
2.16E
-02
0.004333
5.925000
0.934219
0.98449
7
7.98358
7
0.72272
9
99.05000
0
PageRank
5
0.929841
3.52E
-02
0.006292
8.584167
0.896510
0.97108
1
7.97214
4
0.72657
0
99.30000
0
PageRank
10
0.910412
6.62E
-02
0.015283
18.91791
7
0.817318
0.93804
0
7.91193
6
0.74818
0
99.45000
0
TrustRan
k
1
0.677634
1.44E
-05
0.000000
0.000000
0.997806
0.99999
4
7.85463
4
0.72252
0
96.83333
3
TrustRan
k
3
0.744481
3.42E
-05
0.000000
0.000000
0.994031
0.99998
9
7.60142
4
0.72514
6
96.90833
3
TrustRan
k
5
0.782194
5.90E
-05
0.000000
0.000000
0.990004
0.99998
4
7.66860
1
0.72885
2
96.96666
7
TrustRan
k
10
0.809570
1.23E
-04
0.000000
0.000000
0.979865
0.99997
1
7.78653
7
0.74830
6
97.10000
0
Page 185
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Appendix D. Key Interaction Effects
Appendix Table D1. Key H04 and H05 rank-based interaction effects.
Test
Metric
Effect
F
p-value
Partial eta²
H04: Algorithm
x Structure
Target
Promotion
Algorithm x
Structure
114.918537
2.814419e-118
0.327479
H05: Algorithm
x Structure x
Intensity
Spearman
Stability
Algorithm x
Structure
1551.453659
0.000000e+00
0.870648
H05: Algorithm
x Structure x
Intensity
Spearman
Stability
Algorithm x
Structure x
Intensity
65.500370
1.949972e-170
0.460189
H05: Algorithm
x Structure x
Intensity
RBO
Algorithm x
Structure
67.681206
6.032906e-74
0.226980
H05: Algorithm
x Structure x
Intensity
RBO
Algorithm x
Structure x
Intensity
0.873179
0.612166
0.011237