Third-Party Library Recommendations Through Robust Similarity

Measures

Abhinav Jamwal

and Sandeep Kumar

Department of Computer Science and Engineering, Indian Institute of Technology Roorkee, Roorkee, India

{abhinav j, sandeep.garg}@cs.iitr.ac.in

Keywords:

Third-Party Library Recommendation, Similarity Measures, Jaccard Similarity, Collaborative Filtering,

App-Library Interactions.

Abstract:

This research systematically investigates the impact of different similarity measurements on third-party li-

brary (TPL) recommendation systems. By assessing the metrics of average precision (MP), average recall

(MR), average F1 score (MF), average reciprocal rank (MRR) and average precision (MAP) at different levels

of sparsity, the research demonstrates the signiﬁcant impact of similarity measurements on recommendation

performance. Jaccard similarity consistently outperformed the measurements tested and performed better in

low-order and high-order app-library interactions. Its ability to reduce the number of sparse data sets and

achieve a balance between precision and recall makes it the optimal measurement for the TPL recommen-

dation. Other measurements, such as Manhattan, Minkowski, Cosine, and Dice, exhibited limitations to a

certain extent, most importantly under sparse conditions. This research provides a practical understanding

of the strengths and weaknesses of similarity measurements, which provides a basis for optimizing the TPL

recommendation system in practice.

1 INTRODUCTION

Rapid growth in mobile app development has fur-

ther intensiﬁed competition, requiring developers to

quickly publish and update apps to meet changing

user demands

. Third-party libraries (TPLs) have be-

come part of this process, increasing software qual-

ity and reducing development efforts (Li et al., 2021).

Studies show that the average Android app on Google

Play uses about 11.81 TPLs(He et al., 2020). How-

ever, the wide availability of TPLs makes it difﬁcult

to choose the most suitable libraries and combinations

of libraries (Lamothe and Shang, 2020; Salza et al.,

2020). Recommendation systems mitigate this limi-

tation by allowing developers to effectively ﬁnd suit-

able TPLs (Nguyen et al., 2020).

Collaborative ﬁltering (CF) has achieved ad-

mirable performance in a variety of recommenda-

tion tasks (Kangas, 2002). Methods such as Li-

bRec (Thung et al., 2013a) and CrossRec (Nguyen

et al., 2020) leverage CF-based similarity approaches

to suggest libraries. However, these approaches only

https://orcid.org/0000-0002-0213-3590

https://orcid.org/0000-0002-3250-4866

https://www.statista.com/statistics/289418/

rely on low-order interactions and overlook high-

order relations between apps and libraries (He et al.,

2020). Furthermore, the performance of these sys-

tems is highly dependent on the choice of similarity

measures, which have a direct inﬂuence on the qual-

ity of the recommendation.

The role of similarity measures is well studied

in machine learning and classiﬁcation, particularly in

K-Nearest Neighbors (KNN), where distance metrics

signiﬁcantly impact performance (Abu Alfeilat et al.,

2019). Although optimization techniques exist for

classiﬁcation, their application in TPL recommenda-

tion remains underexplored. This paper systemati-

cally examines the impact of similarity measures on

the TPL recommendation. Building on existing work

in classiﬁcation (Zhang, 2021) and TPL recommen-

dation (Li et al., 2024), we analyze the effects of co-

sine, Jaccard, and Dice similarity on recommendation

performance. Our study aims to:

• Identify the strengths and limitations of different

similarity measures in the TPL recommendation.

• Evaluate your impact on low-order and high-order

app-library interactions.

• Provide actionable insights to improve the accu-

racy and efﬁciency of TPL recommendation sys-

Jamwal, A. and Kumar, S.

Third-Party Library Recommendations Through Robust Similarity Measures.

DOI: 10.5220/0013366600003928

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 20th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2025), pages 635-642

ISBN: 978-989-758-742-9; ISSN: 2184-4895

635

tems.

Our contributions include a comprehensive anal-

ysis of similarity measures and their impact on rec-

ommendation quality, which elevates the state of TPL

recommender systems.

2 RELATED WORK

Third-party library (TPL) recommendation systems

have gained attention for accelerating software devel-

opment by assisting developers in selecting appropri-

ate libraries. Early research focused on recommend-

ing speciﬁc APIs or program snippets within TPLs.

Zheng et al. (Zheng et al., 2011) proposed an ap-

proach for API replacement during software develop-

ment, while Thung et al. (Thung et al., 2013b) lever-

aged textual feature requests to suggest relevant API

methods.

Beyond API recommendations, researchers have

explored app-library usage patterns to recommend en-

tire TPLs. Saied et al. (Saied and Sahraoui, 2016)

developed COUPMiner, which combined client- and

library-based mining to uncover trends in the use of

app-library. Ouni et al. (Ouni et al., 2017) introduced

LibFinder, which used semantic similarity between

source codes and TPLs for library detection. More

recently, Saied et al. (Saied et al., 2018) proposed

LibCUP, a multilayer clustering approach for catego-

rizing TPLs based on usage history.

Collaborative ﬁltering (CF) has played a signif-

icant role in the recommendation of TPL. LibRec

(Thung et al., 2013a) was among the ﬁrst to apply

CF to Java projects, integrating association rule min-

ing. Nguyen et al. (Nguyen et al., 2020) introduced

CrossRec, a CF-based approach for open-source soft-

ware projects, while He et al. (He et al., 2020) de-

veloped LibSeek, which incorporated matrix factor-

ization and adaptive weighting for diversiﬁed recom-

mendations. However, these approaches primarily

relied on low-order interactions, often overlooking

high-order dependencies. To address this, GRec (Li

et al., 2021) modeled the relationships between the

application and the library as graphs, using graph neu-

ral networks (GNNs) to capture low- and high-order

interactions, signiﬁcantly enhancing the precision of

the recommendation. HGNRec (Li et al., 2024) fur-

ther reﬁned this approach by decomposing the app-

library graph into two homogeneous graphs for efﬁ-

cient aggregation of interaction patterns.

Despite advances in TPL recommendation, the

impact of similarity measures on TPL recommen-

dations remains underexplored. This work bridges

this gap by systematically evaluating their effects

on TPL prediction and recommendation, drawing in-

sights from similarity-based classiﬁcation techniques.

3 MOTIVATING EXAMPLE

Figure 1 illustrates the role of similarity measures in

interactions between the app and the library. The left-

most graph represents an interaction graph, where

apps (A1–A6) are connected to libraries (L1–L5) based

on usage, with edges indicating app-library utiliza-

tion. The middle graph highlights Similar Apps (e.g.,

A1, A2, and A5), identiﬁed using measures like Jaccard

or Cosine, reﬂecting shared library usage. The right-

most graph identiﬁes Similar Libraries (e.g., L1, L2,

and L4), representing frequently used libraries across

different apps. These similarity patterns improve the

recommendation process by identifying libraries rele-

vant to a given application.

This example demonstrates the importance of sim-

ilarity measures in capturing both low-order (direct)

and high-order (indirect) interactions within app-

library graphs:

• Low-order interactions: Libraries directly con-

nected to an app (e.g., L1 and L2 for A1) are pri-

mary recommendation candidates.

• High-order interactions: Libraries connected via

multiple hops (e.g., L4 and L5) capture broader

patterns of co-usage, providing additional recom-

mendations.

Despite their importance in the TPL recommenda-

tion, the impact of similarity measures remains under-

explored. This study investigates their role in shaping

app-library relationships and improving recommen-

dation effectiveness.

4 METHODOLOGY

The effectiveness of third-party library (TPL) recom-

mendation relies on accurately identifying libraries

most likely to be adopted by an application based on

app-library interaction patterns. This study integrates

K-Nearest Neighbor (KNN) similarity measures with

Graph Neural Networks (GNN) to analyze the im-

pact of different distance measures on recommenda-

tion performance, addressing the gap in understand-

ing their role in TPL recommendations.

At the core of this methodology is the app-library

interaction matrix, a binary representation that indi-

cates whether an application has utilized a particu-

lar library. From this matrix, App Similarity Matri-

ces and Library Similarity Matrices are constructed

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

636

Similar Apps shown

by shaded regions

Similar Libraries shown

by shaded regions

Similar Apps: 𝐴1, 𝐴2, 𝐴5

Similar Apps: 𝐴3, 𝐴4, 𝐴6

Similar Libraries: L1, L2, L4

Similar Libraries: L3, L5

Interaction Graph

App Node (A1, A2, etc)

Library Node (L1, L2, etc.)

App-Library Interaction

Figure 1: Visualization of App-Library Interactions. The interaction graph (left) shows connections between apps and li-

braries. Similar apps (middle) and similar libraries (right) are identiﬁed using KNN similarity, highlighted by shaded regions.

using Jaccard, Cosine, Dice, Minkowski, and Man-

hattan similarity measures. As shown in Figure 2,

these matrices facilitate the identiﬁcation of neigh-

bouring applications and libraries based on shared in-

teraction patterns. Recognizing such neighbors is es-

sential to generate accurate and relevant recommen-

dations, which form the foundation of the similarity-

driven recommendation process.

4.1 Impact of Similarity Measures

To quantify proximity between apps and libraries,

we employ similarity measures such as Jaccard, Co-

sine, Dice, Minkowski, and Manhattan. These mea-

sures construct neighbour matrices that capture inter-

actions between the app and the library, which form

the basis for structural analysis, as shown in Figure 2.

Jaccard similarity evaluates shared elements between

two sets, emphasizing common library usage, while

cosine similarity measures the alignment of interac-

tion vectors. The similarity of the slices, a vari-

ant of Jaccard, assigns greater weight to the shared

interactions. Manhattan distance computes absolute

differences, and the Minkowski distance generalizes

the Manhattan and Euclidean distances by adjusting

a sensitivity parameter. These measures identify app

and library neighbors, providing structural informa-

tion on app-library relationships.

The proposed framework integrates these simi-

larity measures into a GNN-based recommendation

pipeline, as shown in Figure 3. Interaction data

are preprocessed and split into training and test-

ing sets, with similarity-based neighbour identiﬁ-

cation enhancing the app-library interaction graph.

The GNN propagates information through this aug-

mented graph, leveraging both direct interactions and

similarity-based neighbourhoods to learn node em-

beddings. These embeddings encode the relation-

ship between the application and the library, en-

abling ranked library recommendations. The frame-

work is evaluated using precision, recall, F1 score,

mean average precision (MAP), and mean reciprocal

rank (MRR) to systematically assess recommenda-

tion quality, ranking effectiveness, and retrieval per-

formance. This approach highlights the role of simi-

larity measures in improving the accuracy of the rec-

ommendation and capturing structural dependencies.

5 EXPERIMENTAL SETUP AND

RESULTS

This section presents the experimental setup used to

evaluate the framework, followed by the results ob-

tained. The evaluation is carried out on the publicly

available MALib-Dataset

, speciﬁcally designed for

third-party library (TPL) recommendation tasks. The

data set consists of 31,432 Android apps and 752

distinct TPLs, represented as nodes in a graph, with

537,011 app-library usage records forming the edges.

Constructed by collecting 61,722 Android ap-

plications from Google Play via AndroidZoo

, the

dataset captures app-library interactions, allowing a

comprehensive analysis of the use of third-party li-

braries (TPL). To ensure accuracy, TPL classiﬁcations

were cross-veriﬁed with libraries available in Maven

and GitHub, expanding the data set to 827 distinct

TPLs and 725,502 app library usage records. With its

extensive node and edge scale, the dataset supports

https://github.com/malibdata/MALib-Dataset

https://androzoo.uni.lu

Third-Party Library Recommendations Through Robust Similarity Measures

637

L1 L2 L3 L4 L5

111

1 1

0 0 0

L1 L2 L3 L4 L5

A1 A2

A3 A4 A5

Similarity Measure

L1 L2 L3 L4 L5

A1 A2

A3 A4 A5

1.0 0.25 0.0 0.0 0.5 0.0

0.0

0.00.330.51.00.25

0.0

0.5

0.0

0.33

0.5 1.0

1.0

0.0

0.33 0.33

0.0

0.33

0.0

0.33

0.0

App Similarity Matrix

Library Similarity Matrix

1.0

0.25 0.50 0.33 0.0

0.25

0.50

0.33

0.0

0.33

0.66 0.25

0.0

0.33

0.0

0.25 0.0

0.66

0.0

1 1

App neighbour Matrix

Library neighbour Matrix

SimApp ≥ 0.5

SimLib ≥ 0.5

App-Library Links

1 = connected, 0 = not connected

Figure 2: Illustration of identifying neighbouring nodes using similarity measures.

low- and high-order relationship modeling, making

it well suited for graph-based recommendation meth-

ods. The publicly accessible data set

ensures repro-

ducibility and validation by the research community.

5.1 Network Settings

Figure 3 illustrates the architecture of the GNN-based

framework, integrating various KNN similarity mea-

sures (e.g., Jaccard, Cosine, Dice, Minkowski, Man-

hattan) to assess their impact on the TPL recommen-

dation. The framework begins with the app-library in-

teraction matrix, where KNN similarity is applied to

construct app- and library-neighbor matrices, enhanc-

ing the structural information within the graph. For

evaluation, a cross-validation approach is employed.

Given an app u in the interaction matrix M, the rm li-

braries (rm ∈ {1, 3, 5}) are randomly removed to cre-

ate an incomplete representation of the app, and the li-

braries removed serve as the ground truth. The model

is trained on the remaining interactions and the num-

ber of selected neighbors k is varied to analyze its ef-

fect on performance. The evaluation is carried out us-

ing the metrics calculated at K = [5, 10, 20], ensuring

a comprehensive assessment of the similarity mea-

sures. This experimental setup enables a systematic

evaluation of KNN similarity measures to improve the

accuracy of the TPL recommendation.

5.2 Result Analysis

This study investigates the impact of different sim-

ilarity measures, namely Cosine, Dice, Minkowski,

Jaccard, and Manhattan, on TPL recommendation.

The evaluation was carried out on datasets with vary-

ing levels of sparsity (rm = 1, 3, 5), where rm rep-

resents the number of libraries removed from the

record of each app during the training phase. Met-

rics such as mean precision (MP), mean recall (MR),

mean F1 score (MF), mean reciprocal rank (MRR)

and mean average precision (MAP) were calculated

at K = 5, 10, 20. The results are summarized in Ta-

ble 1, and the trends in these metrics are visualized in

Figure 4.

The results reveal that the Jaccard similarity con-

sistently outperformed the other measures in multiple

evaluation scenarios. For rm = 1, Jaccard achieved

an MF of 0.263, an MRR of 0.646, and a MAP

of 0.646 in K = 5, outperforming the Cosine, Dice,

Minkowski, and Manhattan similarity measures. At

K = 10 and K = 20, Jaccard maintained its superi-

ority, demonstrating its robustness in accurately cap-

turing shared interactions between the app and the li-

brary even at higher values of K. In contrast, cosine

and Dice similarity measures struggled to achieve

competitive performance, with lower MF values of

0.223 and 0.226 at K = 5, respectively. These results

reﬂect their limitations in handling sparse data scenar-

ios.

For rm = 3, the advantage of the Jaccard similar-

ity became even more pronounced. It achieved an MF

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

638

GNN

algorithm

Recommendations

and Evaluation

Results

Distance Measure

Selecting the

Distance

measure

Dataset

Preparation

Similarity

Calculation

Train Dataset

Test Dataset

Collect app-library interaction data in matrix form,

capturing the relationships between apps and libraries.

Similarity between

apps or libraries

Split the dataset

Define the similarity metric to evaluate

relationships (e.g., Jaccard, Cosine, Manhattan)

Choose the optimal

similarity measure

Use a Graph Neural Network (GNN)

to learn embeddings and predict

library recommendations

Rank libraries for each app using GNN-generated

embeddings and recommend top-N libraries

Present the final output as the

top-N recommended libraries

for each app

Figure 3: The framework of our experiments designed to evaluate the impact of various distance measures on the performance

of the GNN-based recommendation system.

of 0.570 and an MAP of 0.841 at K = 5, which differ-

entiates it as the most effective similarity measure in

moderate sparsity. Minkowski and Manhattan mea-

sures showed competitive performance, with compa-

rable MAP scores of K = 10 and K = 20. However,

Jaccard consistently ranked higher in MF and MRR,

highlighting its ability to balance precision and recall

while maintaining consistency in ranking. The cosine

and Dice measures continued to lag, particularly in re-

call and MAP, showing their limitations in effectively

capturing app-library interactions.

At the highest level of sparsity, rm = 5, the Jaccard

similarity once again emerged as the leading mea-

sure. It achieved an MF of 0.672 and a MAP of

0.879 at K = 5, underscoring its robustness even in

highly sparse data sets. Manhattan and Minkowski’s

measures followed closely, delivering competitive re-

sults but falling slightly short in recall and ranking-

based metrics. The dice and cosine measures strug-

gled to adapt to the sparse setting, with their best MF

values reaching only 0.599 and 0.598, respectively.

The performance gap between Jaccard and other mea-

sures was more pronounced at K = 20, where Jaccard

achieved superior MAP and MRR scores, demon-

strating its effectiveness in retrieving relevant libraries

even under challenging conditions.

The trends observed in Figure 4 further validate

these ﬁndings. The Mean Precision (MP) curves in-

dicate that Jaccard consistently achieved the high-

est MP, reﬂecting its ability to effectively prioritize

relevant libraries. Similarly, the Mean Recall (MR)

curves highlight the superior recall values achieved

by the Jaccard and Manhattan measures, particularly

at K = 20. This demonstrates their ability to capture

a higher proportion of relevant libraries. The Mean

F1 Score (MF) curves reinforce Jaccard’s dominance,

as its harmonic balance of precision and recall con-

sistently exceeded that of other measures. The Mean

Reciprocal Rank (MRR) and Mean Average Precision

(MAP) metrics further emphasize Jaccard’s ability to

rank relevant libraries higher in the recommendation

list, ensuring that developers receive accurate and ac-

tionable recommendations.

The loss reduction curve shown in Figure 4(f) pro-

vides additional insights into the training dynamics of

each similarity measure. Jaccard and Manhattan mea-

sures demonstrated the fastest and most stable loss re-

duction, reﬂecting their ability to effectively minimize

errors and converge to an optimal solution. In con-

trast, the cosine and Dice measures exhibited slower

convergence and less stable loss reduction, indicating

challenges in learning high-quality embeddings under

sparse conditions. The superior loss reduction per-

formance of Jaccard and Manhattan highlights their

efﬁciency in leveraging the structural and interaction

information encoded in the app-library dataset.

The analysis demonstrates the critical role of sim-

ilarity measures in inﬂuencing the quality of TPL

Third-Party Library Recommendations Through Robust Similarity Measures

639

Table 1: Performance Comparison of Different Similarity Measures.

Dataset Similarity K = 5 K = 10 K = 20

MP MR MF MRR MAP MP MR MF MRR MAP MP MR MF MRR MAP

rm=1

Cosine 0.134 0.668 0.223 0.510 0.510 0.076 0.761 0.138 0.523 0.523 0.042 0.843 0.080 0.529 0.529

Dice 0.136 0.679 0.226 0.524 0.524 0.077 0.773 0.141 0.536 0.536 0.043 0.853 0.081 0.542 0.542

Manhattan 0.154 0.768 0.256 0.622 0.622 0.084 0.837 0.152 0.632 0.632 0.044 0.889 0.085 0.635 0.635

Minkowski 0.154 0.769 0.256 0.625 0.625 0.084 0.840 0.153 0.634 0.634 0.044 0.888 0.085 0.638 0.638

Jaccard 0.158 0.789 0.263 0.646 0.646 0.085 0.854 0.155 0.655 0.655 0.045 0.907 0.086 0.658 0.658

rm=3

Cosine 0.399 0.672 0.500 0.824 0.780 0.231 0.779 0.356 0.828 0.740 0.127 0.854 0.221 0.829 0.704

Dice 0.400 0.675 0.502 0.826 0.781 0.233 0.784 0.358 0.830 0.739 0.128 0.860 0.222 0.832 0.703

Manhattan 0.404 0.681 0.506 0.829 0.785 0.233 0.784 0.359 0.833 0.745 0.127 0.855 0.221 0.834 0.711

Minkowski 0.406 0.684 0.509 0.827 0.782 0.233 0.785 0.359 0.830 0.744 0.127 0.856 0.221 0.832 0.711

Jaccard 0.454 0.766 0.570 0.882 0.841 0.250 0.843 0.386 0.885 0.811 0.133 0.894 0.231 0.886 0.785

rm=5

Cosine 0.590 0.597 0.593 0.899 0.850 0.364 0.736 0.487 0.900 0.795 0.205 0.827 0.328 0.901 0.746

Dice 0.595 0.603 0.599 0.894 0.848 0.365 0.739 0.489 0.896 0.795 0.205 0.830 0.329 0.896 0.746

Manhattan 0.567 0.575 0.571 0.874 0.825 0.357 0.722 0.478 0.877 0.768 0.202 0.818 0.324 0.877 0.719

Minkowski 0.566 0.573 0.569 0.872 0.824 0.356 0.720 0.476 0.875 0.767 0.202 0.817 0.324 0.876 0.718

Jaccard 0.668 0.676 0.672 0.916 0.879 0.393 0.795 0.526 0.918 0.834 0.214 0.865 0.343 0.918 0.796

recommendations. Among the evaluated measures,

the Jaccard similarity consistently emerged as the

most effective, excelling in all metrics and K-values.

Its robustness and reliability in capturing both low-

order and high-order interactions make it the pre-

ferred choice for TPL recommendation tasks. Al-

though Manhattan and Minkowski’s measures offered

competitive alternatives, their performance lagged be-

hind Jaccard’s. The ﬁndings emphasize the impor-

tance of selecting an appropriate similarity measure to

ensure accurate and efﬁcient recommendations, par-

ticularly in sparse data scenarios.

6 DISCUSSION

This study underscores the critical role of similarity

measures in third-party library (TPL) recommenda-

tion systems. Among the measures evaluated, the Jac-

card similarity consistently outperformed others in all

metrics and values of K, as shown in Table 1 and Fig-

ure 4. Its ability to capture low-order (direct) and

high-order (multi-hop) interactions makes it particu-

larly effective in sparse datasets. The high mean F1

score (MF), the mean reciprocal rank (MRR), and the

mean average precision (MAP) further highlight its

balanced precision-recall trade-off and ranking effec-

tiveness.

Manhattan and Minkowski measures performed

well in moderately sparse scenarios (rm = 3) but

showed reduced effectiveness in highly sparse set-

tings (rm = 5). Cosine and Dice similarity, relying

on vector alignment and weighted intersections, ex-

hibited consistent limitations across all sparsity lev-

els, particularly in capturing complex app-library in-

teractions. Jaccard’s efﬁciency in learning optimal

embeddings and its faster convergence during training

reinforce its suitability for the TPL recommendation.

These ﬁndings emphasize the importance of selecting

appropriate similarity measures to improve the preci-

sion and efﬁciency of recommendations.

7 CONCLUSION

This study evaluated the impact of similarity mea-

sures on TPL recommendation, focusing on mean

precision (MP), mean recall (MR), mean F1 score

(MF), mean reciprocal rank (MRR) and mean aver-

age precision (MAP) at varying sparsity levels. The

results show that Jaccard similarity is the most ef-

fective, capturing both low-order (direct) and high-

order (multihop) app-library interactions. Its empha-

sis on shared interactions and balanced union enables

superior performance in all metrics. Although Man-

hattan and Minkowski performed well in moderately

sparse scenarios, they were outperformed by Jaccard

in highly sparse datasets. The cosine and Dice mea-

sures struggled with sparse interaction matrices, re-

sulting in lower performance. These insights high-

light the strengths and limitations of different simi-

larity measures, which guide the optimization of TPL

recommendation systems. Future work may explore

additional similarity measures and hybrid approaches

to enhance adaptability across diverse datasets.

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

640

(a) Mean Precision (MP) for Different Similarity Mea-

sures.

(b) Mean Recall (MR) for Different Similarity Measures.

sures.

(d) Mean Reciprocal Rank (MRR) for Different Similar-

ity Measures.

(e) Mean Average Precision (MAP) for Different Simi-

larity Measures.

(f) Overall Loss Reduction Across Similarity Measures.

Figure 4: Comparison of trends across different similarity measures, highlighting their performance and loss.

REFERENCES

Abu Alfeilat, H. A., Hassanat, A. B., Lasassmeh, O.,

Tarawneh, A. S., Alhasanat, M. B., Eyal Salman,

H. S., and Prasath, V. S. (2019). Effects of distance

measure choice on k-nearest neighbor classiﬁer per-

formance: a review. Big data, 7(4):221–248.

He, Q., Li, B., Chen, F., Grundy, J., Xia, X., and Yang,

Y. (2020). Diversiﬁed third-party library prediction

for mobile app development. IEEE Transactions on

Software Engineering, 48(1):150–165.

Kangas, S. (2002). Collaborative ﬁltering and recommen-

dation systems. VTT information technology, pages

18–20.

Lamothe, M. and Shang, W. (2020). When apis are in-

tentionally bypassed: An exploratory study of api

workarounds. In Proceedings of the ACM/IEEE 42nd

International Conference on Software Engineering,

pages 912–924.

Li, B., He, Q., Chen, F., Xia, X., Li, L., Grundy, J., and

Third-Party Library Recommendations Through Robust Similarity Measures

641

Yang, Y. (2021). Embedding app-library graph for

neural third party library recommendation. In Pro-

ceedings of the 29th ACM Joint Meeting on European

Software Engineering Conference and Symposium on

the Foundations of Software Engineering, pages 466–

477.

Li, D., Gao, Y., Wang, Z., Qiu, H., Liu, P., Xiong, Z., and

Zhang, Z. (2024). Homogeneous graph neural net-

works for third-party library recommendation. Infor-

mation Processing & Management, 61(6):103831.

Nguyen, P. T., Di Rocco, J., Di Ruscio, D., and Di Penta,

M. (2020). Crossrec: Supporting software develop-

ers by recommending third-party libraries. Journal of

Systems and Software, 161:110460.

Ouni, A., Kula, R. G., Kessentini, M., Ishio, T., German,

D. M., and Inoue, K. (2017). Search-based soft-

ware library recommendation using multi-objective

optimization. Information and Software Technology,

83:55–75.

Saied, M. A., Ouni, A., Sahraoui, H., Kula, R. G., Inoue,

K., and Lo, D. (2018). Improving reusability of soft-

ware libraries through usage pattern mining. Journal

of Systems and Software, 145:164–179.

Saied, M. A. and Sahraoui, H. (2016). A cooperative ap-

proach for combining client-based and library-based

api usage pattern mining. In 2016 IEEE 24th In-

ternational Conference on Program Comprehension

(ICPC), pages 1–10. IEEE.

Salza, P., Palomba, F., Di Nucci, D., De Lucia, A., and Fer-

rucci, F. (2020). Third-party libraries in mobile apps:

When, how, and why developers update them. Empir-

ical Software Engineering, 25:2341–2377.

Thung, F., Lo, D., and Lawall, J. (2013a). Automated li-

brary recommendation. In 2013 20th Working confer-

ence on reverse engineering (WCRE), pages 182–191.

IEEE.

Thung, F., Wang, S., Lo, D., and Lawall, J. (2013b). Au-

tomatic recommendation of api methods from feature

requests. In 2013 28th IEEE/ACM International Con-

ference on Automated Software Engineering (ASE),

pages 290–300. IEEE.

Zhang, S. (2021). Challenges in knn classiﬁcation. IEEE

Transactions on Knowledge and Data Engineering,

34(10):4663–4675.

Zheng, W., Zhang, Q., and Lyu, M. (2011). Cross-library

api recommendation using web search engines. In

Proceedings of the 19th ACM SIGSOFT symposium

and the 13th European conference on Foundations of

software engineering, pages 480–483.

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

642