Results intricate relationships among products, owing to its

increased breadth. However, when we endeavored to

Comparison with Baseline Models deepen our model by augmenting the number of
We evaluated the efficiency of our model against two transformer blocks, there was no discernible
stateof-the-art benchmarks - BERT4Rec and SASRec. We enhancement in the final instance. This can be possibly
included all the available data -purchase sessions, traced back to the average session length (11.24/14.57)
shopping cart sessions, and style embeddings- for present in our dataset. A deeper model excels at
comparing STYLE4Rec with the benchmarks. As you can recognizing complex relationships in prolonged sessions.
see in Table 2, across all recommendation list lengths, our As the average session lengths in our dataset were
model outshone both BERT4Rec and SASRec in terms of all relatively brief, assessing the model’s performance based
metrics. The HR@5 metric improved from 0.681 to 0.735, on a broader model demonstrated superior results,
NDCG@5 rose from 0.594 to 0.674, and MRR@5 ascended especially in instances when all the data was taken into
from 0.559 to 0.654. SASRec obtained the second-best account.
results in 6 evaluation metrics and BERT4Rec obtained the
second-best results in 3 evaluation metrics. Considering
Dynamic Recommendation
these results, we can conclude that utilizing style
information with the help of neural style transfer In this section, our aim was to demonstrate the
algorithms and utilizing shopping cart sessions in performance of our model under varying maximum
training/validation yields significant improvement in session lengths, for evaluating our model’s performance in
sequential product recommendation task. dynamic recommendation task. The dynamic
recommendation allows ecommerce websites to
Effect of Style Embeddings and Shopping Cart Data recommend products for their users, immediately after
those users enter the website. Waiting for users to finish
As shown in Table 3, we tested the model performance on
4 different configurations to evaluate the effect of each their sessions and then recommending products, might be
method separately. In Style4Rec1 shopping cart sessions an ineffective way for e-commerce websites. In prior
and style embeddings were removed. In Style4Rec2 experiments, we established a uniform maximum session
shopping cart sessions were removed. In Style4Rec3 style length of 20 across all models and configurations.
embeddings were removed. In Style4Rec, we utilized all As we increased the maximum session length, as in
the available data, purchase sessions, shopping cart Figure 5, we observed an increase in all metrics. Increasing
sessions, and style embeddings. the maximum session length allowed our model to utilize
the relationships of the products which are far away from
Comparing the Style4Rec1 and Style4Rec2 models in
Table 3, adding style embeddings to the purchase data the end product. In that way, our model learned more
improves the model performance on 6 out of 9 metrics, distant and complex relationships between products.
and for MRR@10 the results are the same. We can We also observed that increasing the maximum session
conclude that extracting style embeddings with the help
of the neural style transfer algorithm increases the
performance of the sequential product recommendation
task. Comparing Style4Rec1 and Style4Rec3 models, adding
shopping cart data to purchase data increases the model
performance on all of the metrics in all recommendation
list lengths. We can conclude that utilizing shopping cart
data on training and validation yields meaningful
contributions to the sequential product recommendation
task. Comparing the Style4Rec1 and Style4Rec models,
adding both shopping cart data and the style embeddings
at the same time increases the model performance on all
of the metrics in all recommendation list lengths.
However, in this case, we evaluated the model
performance on a wider model. We increased the number
of heads to 8 and the dimension of the learnable product
embeddings to 1024 to make the model wider and allow
more attention heads to make different predictions.
The substantial improvements we observed can be
attributed to the significant volume of information we
incorporated, exceeding all previous instances.
Consequently, our model had the capacity to discern more
 Table 3: Comparison of the 4 configuration settings of Style4Rec. The best results are in boldface, and the second-best
results are underlined. Improvement was calculated between the worst model and the best model, Style4Rec1 and
Style4Rec, respectively.
Configurations HR@5 HR@10 HR@20 NDCG@5 NDCG@10 NDCG@20 MRR@5 MRR@10 MMR@20
Style4Rec1 0.670 0.718 0.769 0.622 0.637 0.650 0.606 0.612 0.616
2
Style4Rec 0.681 0.735 0.788 0.623 0.641 0.654 0.604 0.612 0.615
Style4Rec3 0.718 0.765 0.825 0.667 0.682 0.697 0.650 0.656 0.661
Style4Rec 0.735 0.784 0.838 0.674 0.690 0.704 0.654 0.661 0.665
Improvement +0.065 +0.066 +0.069 +0.052 +0.053 +0.054 +0.048 +0.049 +0.049
Table 4: Effect of negative sampling
Configurations HR@5 HR@10 HR@20 NDCG@5 NDCG@10 NDCG@20 MRR@5 MRR@10 MMR@20
Style4Recneg 0.670 0.718 0.769 0.622 0.637 0.650 0.606 0.612 0.616
Style4Recw\o neg 0.505 0.549 0.585 0.434 0.448 0.457 0.410 0.416 0.419
Decrease -0.165 -0.169 -0.184 -0.188 -0.189 -0.193 -0.196 -0.196 -0.197
Figure 5: The effect of maximum session length. and shopping cart data were incorporated into the
transformer recommender network. Furthermore, our
model was shown to surpass existing state-of-the-art
length to more than 18 does not significantly improve benchmarks across multiple evaluation metrics. Notably,
performance, since the distant products might have less our model was designed with scalability in mind. By
effect on the final decision-making. storing learnable product vectors, we have enabled the
comparison of a new product’s vector with existing ones,
Effect of Negative Sampling allowing for an adaptable and continuously evolving
recommender system. However, while we have made
Negative sampling was employed as our primary strategy significant strides in improving recommendation quality, it
for assessing model performance. However, in real-world is essential to acknowledge that there is always room for
recommendation systems, the necessity arises to rank all further refinement. Our future work will be geared toward
available products to formulate the final product assessing the performance of our model in scalable tasks.
recommendations. As you can see in Table 4, we removed The challenge lies in maintaining and even improving
the negative sampling process in an attempt to simulate upon this level of performance as we scale to handle
our model’s performance under realistic conditions. This larger datasets and more complex recommendation
modification resulted in a decline across all metrics, with scenarios. However, we are confident that with further
decreases ranging between 0.165 and 0.197. With research and continuous development, our model will
negative sampling, the prediction scope was confined to continue to evolve and excel in the ever-changing
101 products; however, in its absence, the prediction landscape of product recommendation systems.
space expanded dramatically to encompass 2991
products. The decrease in prediction space using negative
References
sampling generally leads to improved results, as it reduces Ba, J. L.; Kiros, J. R.; and Hinton, G. E. 2016. Layer
the complexity of the task at hand. The disparity between Normalization. arXiv:1607.06450.
laboratory conditions and actual market complexities Chen, T.; Sun, Y.; Shi, Y.; and Hong, L. 2017. On Sampling
underlines the need for continuous innovation and model Strategies for Neural Network-based Collaborative
refinement to ensure these advanced algorithms can Filtering. arXiv:1706.07881.
reliably deliver optimal recommendations in practice. Gatys, L. A.; Ecker, A. S.; and Bethge, M. 2015. A Neural
Algorithm of Artistic Style. arXiv:1508.06576.
He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Residual
Conclusion Learning for Image Recognition. In 2016 IEEE Conference
We have introduced a multi-layer transformer encoder on Computer Vision and Pattern Recognition (CVPR), 770–
network that integrates the neural style transfer algorithm 778.
to incorporate style information. Alongside this, we He, R.; Kang, W.-C.; and McAuley, J. 2017.
devised a training methodology that emphasizes the Translationbased Recommendation. In Proceedings of the
distinctions between purchased and shopping cart Eleventh ACM Conference on Recommender Systems.
products. Our comprehensive experiments demonstrated ACM.
a significant performance enhancement in sequential He, X.; Liao, L.; Zhang, H.; Nie, L.; Hu, X.; and Chua, T.-S.
product recommendation tasks when style embeddings 2017. Neural Collaborative Filtering. arXiv:1708.05031.
Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2016.
Session-based Recommendations with Recurrent Neural
Networks. arXiv:1511.06939.
Hinton, G.; Vinyals, O.; and Dean, J. 2015. Distilling the
Knowledge in a Neural Network. arXiv:1503.02531.
Kang, W.-C.; and McAuley, J. 2018. Self-Attentive
Sequential Recommendation. arXiv:1808.09781.
Li, J.; Tu, Z.; Yang, B.; Lyu, M. R.; and Zhang, T. 2018. Multi-
Head Attention with Disagreement Regularization.
arXiv:1810.10183.
Lin, J.; Pan, W.; and Ming, Z. 2020. FISSA: Fusing
Item Similarity Models with Self-Attention Networks for
Sequential Recommendation. In Proceedings of the 14th
ACM Conference on Recommender Systems, RecSys ’20,
130–139. New York, NY, USA: Association for Computing
Machinery. ISBN 9781450375832.
Rendle, S.; Freudenthaler, C.; Gantner, Z.; and
SchmidtThieme, L. 2009. BPR: Bayesian Personalized
Ranking from Implicit Feedback. In Proceedings of the
Twenty-Fifth Conference on Uncertainty in Artificial
Intelligence, UAI ’09, 452–461. Arlington, Virginia, USA:
AUAI Press. ISBN 9780974903958.
Rendle, S.; Freudenthaler, C.; and Schmidt-Thieme, L.
2010. Factorizing Personalized Markov Chains for Next-
Basket Recommendation. WWW ’10, 811–820. New York,
NY, USA: Association for Computing Machinery. ISBN
9781605587998.
Simonyan, K.; and Zisserman, A. 2015. Very Deep
Convolutional Networks for Large-Scale Image
Recognition. arXiv:1409.1556.
Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and
Salakhutdinov, R. 2014. Dropout: A Simple Way to Prevent
Neural Networks from Overfitting. Journal of Machine
Learning Research, 15(56): 1929–1958.
Sun, F.; Liu, J.; Wu, J.; Pei, C.; Lin, X.; Ou, W.; and Jiang, P.
2019. BERT4Rec: Sequential Recommendation with
Bidirectional Encoder Representations from Transformer.
arXiv:1904.06690.
Tang, J.; and Wang, K. 2018. Personalized Top-N
Sequential Recommendation via Convolutional Sequence
Embedding. arXiv:1809.07426.
Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones,
L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017.
Attention Is All You Need. arXiv:1706.03762.
Wu, L.; Li, S.; Hsieh, C.-J.; and Sharpnack, J. 2020. SSEPT:
Sequential Recommendation Via Personalized
Transformer. In Proceedings of the 14th ACM Conference
on Recommender Systems, RecSys ’20, 328–337. New
York, NY, USA: Association for Computing Machinery. ISBN
9781450375832.
Zhang, S.; Tay, Y.; Yao, L.; and Sun, A. 2018. Next Item
Recommendation with Self-Attention. arXiv:1808.06414.