User-Centric Product Discovery for Personalized e-Commerce

Recommendations

Mustafa Keskin, Enis Teper and Sinan Kec¸eci

Hepsiburada, Turkey

Keywords:

Recommender Systems, Personalization, Graphsage, node2vec, Approximate Nearest Neighbor Search,

user2user Similarity.

Abstract:

Personalized recommendations in e-commerce platforms often rely on user-item interactions or product sim-

ilarity. In this work, we explore a user2user recommendation paradigm, where products are recommended

based on purchases made by similar users. We investigate three methods for modeling user similarity: binary

category vectors with sparse dot-product search, GraphSAGE embeddings trained on a user–product bipartite

graph, and behavioral user embeddings obtained by averaging Node2Vec-based product vectors. Recommen-

dations are drawn from complementary or previously browsed categories and ranked using recency-aware,

diversity-promoting strategies. Ofﬂine experiments using HitRate@K demonstrate that graph and embedding-

based methods signiﬁcantly outperform the category-based baseline, effectively capturing latent user prefer-

ences and surfacing relevant, novel items.

1 INTRODUCTION

Modern e-commerce platforms host millions of prod-

ucts and users, creating both an opportunity and a

challenge for delivering personalized shopping expe-

riences. Recommender systems have become essen-

tial tools for guiding users through large product cat-

alogs by suggesting items that align with their prefer-

ences. While item-based and content-based methods

are commonly used in production systems, they often

rely heavily on product metadata or the user’s indi-

vidual history, limiting their effectiveness in scenarios

with sparse data or fast-changing inventories.

In this study, we implement a user2user rec-

ommendation pipeline designed speciﬁcally for e-

commerce applications. The core idea is to identify

users who exhibit similar purchase or browsing be-

haviors and recommend products that these similar

users have bought, but which the user has not previ-

ously engaged with. Unlike traditional collaborative

ﬁltering approaches that focus on item-to-item rela-

tionships, our method emphasizes user-level similar-

ity, making it particularly effective for surfacing prod-

ucts from the long tail or handling cold-start users

with limited activity.

This approach is inspired by the intuition that

“users like you” often discover items you might also

be interested in. By leveraging behavioral signals

across the user base, such as co-purchases, session

data, or category preferences, the user2user method

captures latent interests that are difﬁcult to model with

item features alone. It also enables the system to gen-

erate novel recommendations that are not just simi-

lar to what the user has already seen, but reﬂective of

what like-minded users are exploring.

We evaluate the effectiveness of our method us-

ing both ofﬂine metrics (e.g., precision, recall, cover-

age) and online simulations. We further discuss im-

plementation challenges in large-scale environments,

such as maintaining real-time similarity graphs, han-

dling popularity bias, and ensuring recommendation

diversity.

2 RELATED WORKS

Personalized recommendation has been an important

study area in information retrieval and recommender

systems research, with a variety of approaches rang-

ing from collaborative ﬁltering to deep representa-

tion learning. In this section, we review relevant

literature across three key dimensions aligned with

our proposed framework: user-based collaborative

ﬁltering, graph-based recommendation models, and

embedding-based user modeling.

Keskin, M., Teper, E. and Keçeci, S.

User-Centric Product Discovery for Personalized e-Commerce Recommendations.

DOI: 10.5220/0014287000004848

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

In Proceedings of the 2nd International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences (ICEEECS 2025), pages 25-30

ISBN: 978-989-758-783-2

2.1 User-Based Collaborative Filtering

User-based collaborative ﬁltering (UBCF) is one of

the earliest and most intuitive approaches to rec-

ommendation, where users receive recommendations

based on the preferences of similar users [Sarwar et

al., 2001] (Sarwar et al., 2001). Traditional UBCF

relies on computing similarity scores between users,

typically via cosine similarity or Pearson correlation,

based on their interactions (e.g., ratings, purchases,

or clicks). Despite its simplicity, UBCF suffers from

well-known limitations such as sparsity, cold-start is-

sues, and scalability concerns in large datasets.

Several works have proposed enhancements to

mitigate these problems. For example, item clus-

tering [Linden et al., 2003], neighborhood prun-

ing [Desrosiers & Karypis, 2011] (Desrosiers and

Karypis, 2011), and hybrid models that combine user

and item-based signals have shown promise. How-

ever, most UBCF approaches are shallow in represen-

tation and fail to capture the rich structural or tem-

poral context of user behavior, especially in domains

with high item churn like e-commerce.

2.2 Graph-Based Recommendation

Models

Graphs have emerged as powerful structures for mod-

eling interactions between users and items. Bipartite

user–item graphs allow for the capture of both direct

and higher-order co-occurrence patterns. Early meth-

ods such as random walks [Tong et al., 2006] (Tong

et al., 2006) and label propagation have demonstrated

success in capturing implicit feedback paths. More

recently, the use of Graph Neural Networks (GNNs)

has brought signiﬁcant advances. Models such as

GraphSAGE [Hamilton et al., 2017] (Hamilton et al.,

2017), PinSage [Ying et al., 2018] (Ying et al., 2018),

and LightGCN [He et al., 2020] (He et al., 2020) have

been shown to learn high-quality user and item em-

beddings directly from the interaction graph.

In our work, we adopt GraphSAGE due to its

ability to generalize to unseen nodes and inductively

learn embeddings by aggregating information from

a node’s neighborhood. This approach is particu-

larly well-suited for sparse or dynamic e-commerce

graphs, where new users and products frequently en-

ter the system.

2.3 Embedding-Based User Modeling

Learning low-dimensional embeddings to represent

users and items has become a dominant strategy in

modern recommender systems. These embeddings

can be learned via matrix factorization [Koren et

al., 2009] (Koren et al., 2009), sequence models

(e.g., GRU4Rec [Hidasi et al., 2016] (Hidasi et al.,

2016), SASRec [Kang & McAuley, 2018] (Kang

and McAuley, 2018)), or graph-based methods like

Node2Vec [Grover & Leskovec, 2016] (Grover and

Leskovec, 2016). In e-commerce, item embeddings

are often trained on user interaction sequences or

product co-occurrence graphs, capturing substitute or

complementary relationships between products.

Our approach builds on this by learning prod-

uct embeddings via Node2Vec on a product–product

co-occurrence graph constructed from user browsing

and purchase sessions. These embeddings are then

averaged per user to construct behavioral user vec-

tors, allowing us to compute similarity in a latent in-

tent space. Similar techniques have been explored

in session-based recommendation [Quadrana et al.,

2017] (Quadrana et al., 2017) and ofﬂine customer

modeling [Grbovic & Cheng, 2018] (Grbovic and

Cheng, 2018).

2.4 Diversity and Complementarity in

Recommendations

Another key challenge in recommender systems is

balancing relevance with diversity and novelty. With-

out explicit constraints, many systems tend to over-

recommend popular or similar items, leading to user

fatigue. Approaches such as result re-ranking [Ziegler

et al., 2005] (Ziegler et al., 2005), determinantal point

processes (DPP) [Kulesza & Taskar, 2012] (Kulesza

and Taskar, 2012), and intent-aware diversiﬁcation

[Vargas & Castells, 2011] (Vargas and Castells, 2011)

have been proposed to address this issue. Our round-

robin sampling strategy across categories aims to in-

crease inter-category diversity while preserving topi-

cal relevance by incorporating recent browsing con-

text.

Finally, the concept of recommending from com-

plementary categories, rather than just similar ones,

has been underexplored, despite its relevance in do-

mains like e-commerce where users often buy re-

lated products over time (e.g., phone → case →

charger). Recent works in complementary item pre-

diction [McAuley et al., 2015] (McAuley et al., 2015)

and basket completion have laid the groundwork for

this direction, which our method expands upon.

3 METHODOLOGY

We designed and implemented a user2user recom-

mendation framework tailored for a large-scale e-

ICEEECS 2025 - International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences

commerce platform. The system is built upon be-

havioral data spanning the past 12 months, and it tar-

gets users with at least 10 distinct product purchases.

This threshold ensures that the similarity estimation

is grounded in meaningful behavioral patterns.

The goal is to retrieve the top-10 most similar

users for a given target user, and recommend products

that those similar users have purchased but the target

user has not. To capture different aspects of user sim-

ilarity, we developed and evaluated three distinct ap-

proaches: category-based sparse vectors, graph-based

embeddings, and behavior-driven product vector ag-

gregation.

3.1 Category-Based User Similarity

In this method, we represent each user as a high-

dimensional binary vector indicating their interac-

tions with product categories. Each vector dimension

corresponds to a product category, where a value of 1

indicates at least one purchase in that category. Given

the sparsity of the data, we utilize the sparse dot topn

eron, 2018) library to efﬁciently compute approxi-

mate top-N cosine similarities. This allows us to re-

trieve the 10 most similar users for each target user at

scale.

Figure 1: User-Category Binary Matrix.

The Fig 1 illustrates a sparse binary matrix in

Compressed Sparse Row (CSR) format, where rows

corresponding to users and columns to product cate-

gories. A value of 1 indicates at least one purchase

in that category, capturing user-category interactions

compactly.

3.2 Graph-Based User Embeddings

To model deeper structural relationships between

users and products, we construct a bipartite graph

where nodes represent users and products, and edges

denote purchase interactions (Kazemi et al., 2019).

We then train a GraphSAGE model using a link pre-

diction objective to learn node embeddings. Graph-

SAGE operates by iteratively aggregating and trans-

forming feature information from a node’s local

neighborhood. For each node v at layer k, its em-

bedding h

is computed based on its own embedding

from the previous layer h

k−1

and the aggregated in-

formation from its neighbors N (v). A common ap-

proach, using the mean aggregator, can be expressed

as:

N (v)

= AGGREGATE

({h

k−1

: u ∈ N (v)}) (1)

Here, AGGREGATE

typically refers to an aggre-

gation function (such as mean, sum, or max-pooling)

applied to the embeddings of node v’s neighbors from

the previous layer k −1. For the mean aggregator, this

would like:

N (v)

|N (v)|

∑

u∈N (v)

k−1

(2)

Following aggregation, the node’s embedding is

updated by combining its own previous embedding

with the aggregated neighbor information, typically

followed by a non-linear activation:

= σ



· CONCAT(h

k−1

, h

N (v)

) + b



(3)

In this update step, σ is a non-linear activation

function, W

is a learnable weight matrix, b

is a

learnable bias vector, and CONCAT represents the

concatenation of the node’s embedding and its aggre-

gated neighborhood embedding. This iterative pro-

cess allows each node’s embedding to capture in-

creasingly broader structural and feature information

from its multi-hop neighborhood.

This enables us to encode not only direct

user-product interactions but also higher-order co-

purchase patterns. After training, we extract user em-

beddings and perform approximate nearest neighbor

(ANN) search using the HNSW (Hierarchical Navi-

gable Small World) (Malkov and Yashunin, 2016) al-

gorithm to ﬁnd the top 10 most similar users.

3.3 Behavioral Product Embeddings

and Averaging from Orders

Our experiments also evaluate two production-grade

user modeling baselines that generate user embed-

dings by averaging behavioral product vectors. In

both methods, user embeddings are formed by aver-

aging the embeddings of products found in their his-

torical orders. The key difference lies in how the indi-

vidual product embeddings are constructed: the V2V

model (Attokurov et al., 2022) learns product embed-

dings from co-view patterns within the same session

(i.e., ”users who viewed this also viewed”), while

User-Centric Product Discovery for Personalized e-Commerce Recommendations

Figure 2: Overview of the methodology: Bipartite user-

product interaction graph and its corresponding embedding

space. On the left, users (User1, User2, etc.) are connected

to products (Product1, Product2, etc.) they have purchased

or interacted with. Through network embedding techniques

GraphSAGE, the graph structure is transformed into a la-

tent vector space (right), where proximities reﬂect behav-

ioral or structural similarities. This representation enables

user-to-user or item-to-item recommendation based on spa-

tial closeness.

the FBT model (Keskin et al., 2024) trains on co-

purchase graphs derived from same-day orders (i.e.,

”frequently bought together”). Both models utilize

the Node2Vec algorithm over their respective prod-

uct–product graphs. These embeddings serve as rep-

resentative behavioral priors in our user2user similar-

ity computations.

Figure 3: Getting average product embedding from user his-

tory

4 RECOMMENDATION

GENERATION

Once the top-10 most similar users are identiﬁed for a

given seed user using one of the methods described in

the previous section, we proceed to generate personal-

ized product recommendations by leveraging the be-

havioral footprints of those similar users. Our goal is

not merely to recommend frequently purchased items,

but to surface products that are novel, contextually

relevant and diverse.

4.1 Extracting Candidate Categories

For each target user, we begin by aggregating the

product categories purchased by the 10 nearest users.

This aggregated set reﬂects the broader interest space

of similar users. To ensure novelty, we ﬁlter out the

categories in which the seed user has already made

purchases, and instead focus on:

• Categories the user has previously browsed or in-

teracted with, but not purchased

• And/or categories that are considered comple-

mentary to the user’s previous purchases (e.g., if a

user bought a camera, accessories like tripods or

memory cards are deemed complementary).

This step ensures that recommendations explore

adjacent areas of interest while maintaining contex-

tual relevance.

4.2 Candidate Product Pool

Construction

We enhanced the recently purchased products of users

by incorporating substitute and complementary cate-

gories. This creates a large pool of candidate items.

To avoid temporal bias and ensure relevance, we fur-

ther rank these candidate products based on the re-

cency of the seed user’s browsing activity in each

corresponding category. That is, products from cat-

egories the user has recently interacted with are prior-

itized higher.

4.3 Ranking and Diversiﬁcation

Strategy

To construct the ﬁnal recommendation list, we adopt

a round-robin selection strategy across categories to

increase diversity and avoid item redundancy. Rather

than recommending multiple products from the same

category consecutively which often leads to diminish-

ing returns in user engagement we iteratively sample

one product from each relevant category in a rotating

manner. This ensures that the resulting recommenda-

tion list is category-balanced and prevents overcon-

centration in any single interest area.

The ﬁnal ranking thus reﬂects a careful balance

between:

• Behavioral similarity (from user2user modeling)

• Recency-aware personalization

• And inter-category diversity

This approach is particularly effective in e-

commerce platforms where users are often multi-

ICEEECS 2025 - International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences

Table 1: Ofﬂine performance comparison of ﬁve user similarity methods on HitRate and NDCG at different cutoffs.

Method HitRate NDCG

@10 @20 @50 @10 @20 @50

SparseDotTopn 0.001041 0.001387 0.001635 0.000586 0.000677 0.000737

V2V 0.002684 0.004016 0.004858 0.001380 0.001722 0.001908

FBT 0.001174 0.001390 0.001408 0.000711 0.000774 0.000778

GraphSAGE (raw) 0.000839 0.001131 0.001319 0.000456 0.000532 0.000575

GraphSAGE (avg) 0.000656 0.000854 0.000994 0.000362 0.000413 0.000444

intent, and engagement improves with exposure to a

variety of relevant options rather than a narrow focus.

5 EVALUATION

To evaluate the effectiveness of our user similarity

methods in a real-world e-commerce setting, we con-

ducted an ofﬂine evaluation using historical purchase

data. For each eligible user, we held out the products

purchased in the following 30 days as ground truth,

simulating a next-period recommendation task. Our

goal was to assess whether the users retrieved by our

similarity methods would help uncover relevant prod-

ucts that the target user is likely to purchase in the

near future.

We compared ﬁve methods for user similarity:

• SparseDotTopn: cosine similarity over binary

category vectors.

• V2V: user embeddings averaged from product

vectors trained on session-based co-view graphs.

• FBT: user embeddings averaged from product

vectors trained on same-day co-purchase graphs.

• GraphSAGE (raw): user embeddings directly

obtained from a trained GraphSAGE model on the

user–product bipartite graph.

• GraphSAGE (avg): user embeddings obtained

by averaging product embeddings from Graph-

SAGE.

Performance was measured using standard top-

k ranking metrics: HitRate@k and nDCG@k (Nor-

malized Discounted Cumulative Gain). HitRate@k

checks whether at least one of the held-out items

appears in the top-k recommended products, while

nDCG@k takes into account the rank position of cor-

rect predictions, rewarding higher placements.

As shown in Table 1, the v2v method achieved the

best overall performance across all metrics and cut-

offs. This suggests that session-based co-view em-

beddings provide a strong behavioral signal for re-

trieving similar users. The fbt approach also per-

formed well, outperforming GraphSAGE variants and

SparseDotTopn in most cases, reﬂecting the effective-

ness of co-purchase-based embeddings.

The GraphSAGE (avg) method, while concep-

tually simple, lagged behind v2v and fbt, poten-

tially due to over-smoothing effects during aggrega-

tion. GraphSAGE (raw) produced more expressive

representations but struggled with sparsity and noise

in the bipartite graph. Finally, the SparseDotTopn

method offered moderate performance, demonstrat-

ing that even lightweight sparse vector techniques can

be competitive with proper category engineering.

These results highlight the trade-offs between

simplicity, scalability, and expressiveness when de-

signing user representation models for e-commerce

recommendations.

6 CONCLUSION AND FUTURE

WORKS

In this work, we proposed a scalable user-based rec-

ommendation framework for large-scale e-commerce

platforms by implementing binary category inter-

action vectors and graph-based embedding tech-

niques. Our method effectively identiﬁes simi-

lar users through approximate top-N cosine similar-

ity computations, enabling personalized and inter-

pretable product recommendations. We demonstrated

that simple, sparse interaction-based modeling (v2v)

outperforms more complex graph neural approaches

in cold-start and sparse settings, especially when re-

lying solely on purchase data.

In future work, we plan to incorporate richer be-

havioral signals such as product clicks, add-to-cart

events, and dwell time to construct more ﬁne-grained

user representations. By integrating these signals, we

expect to capture earlier stages of user intent and ex-

tend the recommendation coverage to a broader user

base including those without recent purchase history.

Moreover, temporal modeling and causal inference

approaches could help further reﬁne recommendation

quality and robustness.

User-Centric Product Discovery for Personalized e-Commerce Recommendations

ACKNOWLEDGEMENTS

This project was made possible by the individual con-

tributions of each member of the recommendation

team within Hepsiburada technology group. Also,

this project would not have been possible if the tech-

nology group management of Hepsiburada had not

supported and encouraged the data science team in

innovation.

REFERENCES

Attokurov, U., Kaya, O., and Sezgin, M. S. (2022). Product

recommendation based on embeddings: People who

viewed this product also viewed these products. In

2022 IEEE International Conference on Big Data and

Smart Computing (BigComp), pages 296–299. IEEE.

Desrosiers, C. and Karypis, G. (2011). A comprehensive

survey of neighborhood-based recommendation meth-

ods. Recommender Systems Handbook, pages 107–

144.

Grbovic, M. and Cheng, H. (2018). Real-time personaliza-

tion using embeddings for search ranking at airbnb.

Proceedings of the 24th ACM SIGKDD International

Conference on Knowledge Discovery & Data Mining,

pages 311–320.

Grover, A. and Leskovec, J. (2016). node2vec: Scal-

able feature learning for networks. In Proceedings of

the 22nd ACM SIGKDD International Conference on

Knowledge Discovery and Data Mining, pages 855–

864.

eron, A. (2018). sparse-dot-topn: Efﬁcient sparse ma-

trix multiplication for top-n cosine similarity. https:

//github.com/ing-bank/sparse\ dot\ topn. Accessed:

2025-07-20.

Hamilton, W. L., Ying, R., and Leskovec, J. (2017). In-

ductive representation learning on large graphs. In

Advances in Neural Information Processing Systems

(NeurIPS), pages 1024–1034.

He, X., Deng, K., Wang, X., Li, Y., Zhang, Y., and Wang,

M. (2020). Lightgcn: Simplifying and powering graph

convolution network for recommendation. In Pro-

ceedings of the 43rd International ACM SIGIR Con-

ference on Research and Development in Information

Retrieval, pages 639–648.

Hidasi, B., Karatzoglou, A., Baltrunas, L., and Tikk, D.

(2016). Session-based recommendations with re-

current neural networks. In Proceedings of the In-

ternational Conference on Learning Representations

(ICLR).

Kang, W.-C. and McAuley, J. (2018). Self-attentive se-

quential recommendation. In Proceedings of the IEEE

International Conference on Data Mining (ICDM),

pages 197–206.

Kazemi, S. M., Goel, R., Jain, K., Kobyzev, I., Sethi, A.,

Forsyth, P., and Poupart, P. (2019). Relational repre-

sentation learning for dynamic (knowledge) graphs: A

survey. CoRR, abs/1905.11485.

Keskin, M., Teper, E., and Kurt, A. (2024). Comparative

evaluation of word2vec and node2vec for frequently

bought together recommendations in e-commerce. In

2024 9th International Conference on Computer Sci-

ence and Engineering (UBMK), pages 1–5. IEEE.

Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factor-

ization techniques for recommender systems. Com-

puter, 42(8):30–37.

Kulesza, A. and Taskar, B. (2012). Determinantal point

processes for machine learning. In Foundations and

Trends in Machine Learning, volume 5, pages 123–

286. Now Publishers Inc.

Malkov, Y. A. and Yashunin, D. A. (2016). Efﬁcient

and robust approximate nearest neighbor search us-

ing hierarchical navigable small world graphs. CoRR,

abs/1603.09320.

McAuley, J., Targett, C., Shi, Q., and van den Hengel,

A. (2015). Inferring networks of substitutable and

complementary products. In Proceedings of the 21th

ACM SIGKDD International Conference on Knowl-

edge Discovery and Data Mining, pages 785–794.

Quadrana, M., Karatzoglou, A., Hidasi, B., and Cremonesi,

P. (2017). Personalizing session-based recommen-

dations with hierarchical recurrent neural networks.

CoRR, abs/1706.04148.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001).

Item-based collaborative ﬁltering recommendation al-

gorithms. In Proceedings of the 10th international

conference on World Wide Web (WWW), pages 285–

295. ACM.

Tong, H., Faloutsos, C., and Pan, J.-Y. (2006). Fast random

walk with restart and its applications. Proceedings

of the Sixth IEEE International Conference on Data

Mining (ICDM), pages 613–622.

Vargas, S. and Castells, P. (2011). Rank and relevance in

novelty and diversity metrics for recommender sys-

tems. In Proceedings of the ﬁfth ACM conference on

Recommender systems, pages 109–116.

Ying, R., He, R., Chen, K., Eksombatchai, P., Hamilton,

W. L., and Leskovec, J. (2018). Graph convolutional

neural networks for web-scale recommender systems.

In Proceedings of the 24th ACM SIGKDD Interna-

tional Conference on Knowledge Discovery & Data

Mining, pages 974–983.

Ziegler, C.-N., McNee, S. M., Konstan, J. A., and Lausen,

G. (2005). Improving recommendation lists through

topic diversiﬁcation. In Proceedings of the 14th inter-

national conference on World Wide Web, pages 22–32.

ACM.

ICEEECS 2025 - International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences