Available via license: CC BY 4.0
Content may be subject to copyright.
SAFERec: Self-Attention and Frequency Enriched
Model for Next Basket Recommendation
Oleg Lashinin1, Denis Krasilnikov2, Aleksandr Milogradskii2, and Marina
Ananyeva3,2
1Moscow Institute of Physics and Technology, Russia
2T-Bank, Russia
3National Research University Higher School of Economics, Russia
fotol764@gmail.com
di.krasilnikov@gmail.com
alex.milogradsky@gmail.com
ananyeva.me@gmail.com
Abstract. Transformer-based approaches such as BERT4Rec and SAS-
Rec demonstrate strong performance in Next Item Recommendation
(NIR) tasks. However, applying these architectures to Next-Basket Rec-
ommendation (NBR) tasks, which often involve highly repetitive inter-
actions, is challenging due to the vast number of possible item combi-
nations in a basket. Moreover, frequency-based methods such as TIFU-
KNN and UP-CF still demonstrate strong performance in NBR tasks,
frequently outperforming deep-learning approaches. This paper intro-
duces SAFERec, a novel algorithm for NBR that enhances transformer-
based architectures from NIR by incorporating item frequency informa-
tion, consequently improving their applicability to NBR tasks. Extensive
experiments on multiple datasets show that SAFERec outperforms all
other baselines, specifically achieving an 8% improvement in Recall@10.
Keywords: Next Basket Recommendations ·Recommender Systems ·
Transformer ·E-Commerce Recommendations
1 Introduction
The field of next-basket recommendations (NBR) has attracted significant inter-
est from both academic [16, 23,25] and industry [12,13], primarily caused by the
NBR’s critical role in enhancing the user experience on e-commerce platforms.
Predicting users’ subsequent purchases is essential in today’s digital shopping
landscape, supporting features like item reminders [14], product discovery [9],
and efficient order creation [27,30]. Additionally, the increasing complexity and
size of e-commerce platforms demonstrate a strong demand for innovations and
further improvements in NBR methodologies.
Some of the best performing methods in NBR are frequency-based algo-
rithms, such as TIFU-KNN [8] and UP-CF [6], which outperform even deep learn-
ing methods on these tasks [16]. However, despite their superior performance,
arXiv:2412.14302v1 [cs.IR] 18 Dec 2024
2 O. Lashinin et al.
these methods have some limitations. For instance, they struggle to adapt to di-
verse data types, such as images or textual descriptions, and rely on user-based
K-nearest Neighbors methods, leading to scalability issues in large datasets.
Furthermore, in recent years, transformer-based models [29] such as SAS-
Rec [11], gSASRec [21], and BERT4Rec [28] have demonstrated remarkable per-
formance in next-item recommendations (NIR) owing to the ability to model the
sequential nature of user’s behavior and complex optimization objectives [19,20,
31]. However, applying these methods to NBR remains challenging because of the
vast number of potential baskets. Nevertheless, the application of transformer-
based algorithms remains a promising topic in NBR tasks, as advancements in
these models could significantly improve the effectiveness of NBR methodologies.
In addition, recent studies [1,3,22] have highlighted the importance of inte-
grating transformer-based and frequency-based methods. In particular, frequency-
based methods can enhance user-specific recommendations by effectively captur-
ing repeated items, which are crucial for NBR tasks. Our study shows that the
item frequencies component is a necessary part that significantly boosts the per-
formance of the standard SASRec model, at times even surpassing state-of-the-
art methods on specific datasets. In this work, we propose a novel architecture
for NBR called SAFERec that combines transformer-based and frequency-based
methods. To the best of our knowledge, there is a lack of work that utilizes
frequency-aware components to adapt the NIR models for the NBR tasks.
The contributions of this work are twofold. First, we propose a novel NBR
method, SAFERec, which integrates a transformer layer and a frequency-aware
module for NBR. To support reproducibility, we have publicly released the code
of SAFERec4.Second, we conduct offline experiments on three public datasets.
The results show that SAFERec outperforms other state-of-the-art methods,
specifically by up to 8% in Recall@10, and recommends a more novel set of items.
2 SAFERec
Problem Formulation. In a typical Next Basket Recommendation (NBR)
scenario, we have a set of users Uand items V. For each user in the dataset
there is an ordered set of purchases with baskets Bu=bu
1, . . . , bu
|Bu|, where
bu
k= (vu,j
1, . . . , vu,j
|bu
k|)represents an unordered set of items. The objective of NBR
is to predict the next basket bu
|Bu|+1 for a user u. In offline experiments, models
usually predict bu
|Bu|based on the purchase history, known as the leave-one-out
evaluation protocol. However, in real-world applications, the system operates
with the entire Buand can be evaluated once the next basket bu
|Bu|+1 is observed.
2.1 Architecture
History Encoding Module. SAFERec processes each user’s purchase history
Buin mini-batches. This begins with the conversion of baskets into sparse multi-
hot vectors qu
k∈R|V|, where the i-th component equals 1 if item iis present
4https://github.com/anon-ecir-nbr/SAFERec
Title Suppressed Due to Excessive Length 3
Fig. 1. SAFERec architecture overview.
in basket bu
kand 0 otherwise. The user’s entire purchase history Buis thus
represented as a sparse matrix Wu∈Rlu×|V|of sparse vectors qu
k, with lu
denoting the number of purchases made by user u.
To handle variations in purchase history length, we introduce a hyperparam-
eter Lrepresenting the maximum number of recent baskets the model considers.
If L > lu, we apply left padding to align all user purchase histories to a uniform
length. Conversely, if L < lu, we select Lmost recent purchase histories.
Given the typically large catalog size |V|, we utilize a series of fully connected
(FC) layers, similar to the approach in Mult-VAE [17] with the tanh activation
function. For simplicity, the hyperparameter dis employed for all hidden dimen-
sions throughout these linear layers and other modules in the SAFERec. The
final layer outputs the matrix Wu,lat ∈RL×d, representing the user’s history,
where each basket is encoded into a d-dimensional latent vector space.
User Representation Module. We employ a Transformer Layer [29] to
capture user-specific representations, following a structure similar to SASRec [11]
In contrast to the fixed positional embeddings used in [29], our approach incor-
porates learned ones, denoted as P∈RL×d, for positions 1through L, resulting
in Wu,p=Wu,lat +P. This change consistently delivered better results.
For hyperparameters, we adopt a setup similar to SASRec but vary the num-
ber of attention heads and stacked Transformer layers. Furthermore, we set inner
hidden and output dimensions to dto reduce the hyperparameter search space.
The Transformer layer’s output, Wu,tr =T r(Wu,p)∈RL×d, encodes the user’s
basket history Bu. Following SASRec’s methodology, the final vector wlast
ufrom
Wu,tr acts as a hidden representation of the user. Finally, stacked FC layers are
applied to wlast
uto obtain the user representation uh.
Approaches to Frequency Module. To address the highly repetitive na-
ture of user purchases in NBR, we introduce the Frequency Module (FM ). Prior
work, such as TIFU-KNN [8], has shown that Recurrent Neural Networks strug-
gle to accurately capture sums of input vectors, raising similar concerns about
the effectiveness of Transformer layers for this task. In NBR, it is essential to rec-
ommend previously purchased items, yet encoding all possible item combinations
within the latent user vector uhcan be challenging for models like SASRec.
4 O. Lashinin et al.
Proposed FM Approach. To tackle this issue, we incorporate item occur-
rence frequencies into the model, as illustrated in Figure 1. SAFERec utilizes two
learned item embedding matrices: I1,I2∈R|V|×d, used for collaborative item
representation and frequency aware representation, respectively. For each user-
item pair, we define a history vector hi
u=1{i∈bk
u}L
k=1, indicating whether user
upurchased item iin each of their past baskets Bu. This history vector helps
capture the frequency of item interactions within the user’s purchasing history.
Item-Specific and User-Specific Patterns. FM utilizes both user his-
tory vectors, hi
u, and item embedding, i2∈I2. For explicit frequency accounting,
we use learned embeddings F E ∈RFmax×dfor each frequency value from 0to
Fmax. All values greater than Fmax are clipped to Fmax . We concatenate these
components to form the vector ci
u=i2+fi;hi
u, which we then process through
a series of FC layers with a tanh activation function. This transformation outputs
an item-specific frequency-based score pi
ui for each user-item pair.
User preferences are incorporated similarly to SASRec through the dot prod-
uct of the item embedding i1and the latent user vectors uhfrom the Transformer
Layer, which encodes patterns from the user’s interactions hi
u. The output pi
uu =
i1·uhis the prediction score for a user-item pair, reflecting user-specific patterns.
The final prediction of the SAFERec model is generated by summing both
scores pi
u=pi
uu +pi
ui. We argue that including frequency information helps the
model generalize better, improving its performance compared to SASRec [11].
Objective Function. In order to optimize the parameters of our neural
network, we employ the Cross-Entropy Loss function, which remains a highly
effective approach according to recent top-nrecommendation benchmarks [26].
Like Mult-VAE [17], which predicts the set of items in a top-n recommendation
setup, SAFERec predicts the entire basket, in contrast to SASRec and other
next-item models that typically predict only one item. Notably, incorporating
item frequencies into variational inference methods is left for future work.
3 Experiments
We design our experiments to answer the following research questions:
RQ1 How do frequency-aware techniques affect Transformer performance?
RQ2 How does SAFERec solve the NBR problem compared to well-established
NBR baselines?
3.1 Offline Experimental setup
Datasets. We run experiments on three public datasets: TaFeng5, Dunnhumby6,
TaoBao7. We apply minimal preprocessing across all datasets: users and items
with fewer than 5 interactions are removed, as well as users with only one basket.
5https://www.kaggle.com/chiranjivdas09/ta-feng-grocery-dataset
6https://www.kaggle.com/datasets/frtgnn/dunnhumby-the-complete-journey
7https://tianchi.aliyun.com/dataset/649
Title Suppressed Due to Excessive Length 5
Table 1. Dataset statistics after preprocessing.
Dataset #users #items #baskets avg #baskets
per user
Dunnhumby 2495 33910 222196 89.06
Tafeng 19112 14713 104257 5.46
TaoBao 144408 82310 820764 5.68
For TaoBao, this threshold is increased to 10 due to computational constraints.
The main statistics are presented in Table 1.
Metrics. We employ two ranking-based metrics: Recall@K and NDCG@K
with binary relevance function. We also compute UserNovelty@K (UN@K),
which indicates the rate of new items for user u, defined as UN@K=Pk
j=1
1
[j∈
Bu]. All metrics are reported at cutoffs 10 and 100.
Baselines. We compare SAFERec to well-established baselines for NBR. In
a recent comprehensive NBR reproducibility study [16], the authors show that
TIFU-KNN [8], UP-CF [6], and DNNTSP [32] - the latter being a deep learning
method - are more stable compared to other proposed deep learning approaches.
Therefore, we selected these models as our baselines. Notably, we exclude novel
methods such as ReCaNet [3], BTBR [15], PerNIR [4] as they address different
issues such as repetition, next-novel, and within-basket recommendations. We
omit the recent TIFU-KNN extension, TAIW [24], which explicitly models time
intervals. Although SAFERec could be extended to incorporate time-awareness,
we leave this for future work. We also train a version of the SAFERec model with-
out the FM, called SASRec* [11], to represent its adaptation for the NBR task.
Evaluation Protocol. We adopt a leave-one-basket protocol, commonly
used in recent studies on this task [6,8]. The last basket for each user is assigned
either for validation or testing, with half of the users allocated to the validation
set and the other to the test set. All remaining baskets are used to train the
models. For DNNTSP and SAFERec, we employ early stopping with a patience
of 5 epochs and a maximum of 100 epochs. Hyperparameter tuning is conducted
with Optuna [2], using 50 trials for each model8. This setup provides a robust
foundation for evaluating SAFERec against the established baselines.
3.2 Results of Offline Experiments
Table 2 presents findings for both RQ1 and RQ2. First, SAFERec consistently
outperforms the SASRec model [11], demonstrating its ability to capture item-
specific patterns of recurrent behavior effectively. Additionally, SASRec*, per-
forms nearly at the P-Pop level, suggesting it struggles to promote previously
consumed items through hidden vectors alone. Although SASRec* shows an ad-
vantage in UN@K, this metric is less emphasized due to its poor ranking quality.
Nevertheless, SAFERec excels in recommending novel items, offering a more di-
8The search grid and optimal hyperparameters are available in the repository: https:
//github.com/anon-ecir-nbr/SAFERec/blob/main/README.md
6 O. Lashinin et al.
Table 2. The best and second-best models are bolded and underlined, respectively.
The symbol †denotes no statistically significant difference (p > 0.05) from the
SAFERec model, as determined by a paired t-test with Bonferroni correction [5].
Dataset
Metric Model
P-Pop GP-Pop TIFU-KNN UP-CF DNNTSP SASRec* SAFERec
TaFeng
Recall@10 0.0633 0.09991 0.0897 0.1214 0.1276†0.0711 0.1256 (▽1%)
Recall@100 0.1845 0.2708 0.1557 0.2701 0.2455 0.1947 0.2761 (△3%)
NDCG@10 0.1589 0.2084 0.2005 0.2417 0.1955 0.1381 0.2504 (△3%)
NDCG@100 0.2143 0.2642 0.2254 0.2854 0.2524 0.0995 0.2927 (△4%)
UN@10 0.9136 0.1373 0.1373 0.1805 0.191 0.9142 0.2865 (△50%)
UN@100 0.9642 0.7518 0.7518 0.7522 0.7576 0.9646 0.7964 (△5%)
TaoBao
Recall@10 0.0033 0.0637 0.0721†0.0648 0.0623 0.0046 0.0707 (▽2%)
Recall@100 0.0229 0.1028†0.085 0.1018 0.0966 0.0225 0.1014 (▽1%)
NDCG@10 0.0052 0.067 0.0843 0.0684 0.0639 0.0065 0.08 (▽4%)
NDCG@100 0.0157 0.0835 0.0896 0.0922†0.0787 0.0159 0.0949 (△1%)
UN@10 0.9934 0.0701 0.0701 0.0704 0.0711 0.9936 0.0794 (△12%)
UN@100 0.9964 0.8627 0.8627 0.8627 0.8629 0.9965 0.9406 (△10%)
Dunnhumby
Recall@10 0.0814 0.1415 0.1472 0.1503 0.1279 0.0840 0.1619 (△8%)
Recall@100 0.1815 0.3312 0.3402 0.3395 0.3023 0.1921 0.3585 (△6%)
NDCG@10 0.2703 0.3805 0.3913 0.4026†0.3492 0.2655 0.4086 (△1%)
NDCG@100 0.2812 0.394 0.4035 0.4063†0.3961 0.2857 0.4184 (△2%)
UN@10 0.4771 0.0033 0.0033 0.0033 0.0108 0.4507 0.0147 (△36%)
UN@100 0.6707 0.0555 0.0555 0.0764 0.0923 0.6659 0.1029 (△11%)
verse user experience with higher accuracy. This balance means users receive
accurate recommendations while also discovering new products.
Furthermore, Table 2 addresses RQ2, showing that SAFERec outperforms
baselines across all datasets, except the TaoBao dataset, where GP-Pop and
TIFU-KNN surpass the model in Recall@100 and NDCG@10, likely due to the
dataset’s emphasis on rare items. GP-Pop and TIFU-KNN recommend rare items
based on user purchase history rather than learned hidden representations. No-
tably, baseline results align with reproducibility papers [16, 18], where TIFU-
KNN, UP-CF, and DNNTSP show comparable performance and outperform
GP-Pop. Despite accuracy differences, UN@Kindicates SAFERec’s higher rate
of recommending new items, which could enhance user satisfaction and engage-
ment, outlined in [7, 10, 15]. Additionally, SAFERec maintains competitive ac-
curacy metrics while exploring novel items. For instance, on TaFeng, SAFERec
achieves better quality metrics than UP-CF, recommending 28% more novel
items in the top 10 compared to UP-CF’s 18%. These results show SAFERec’s
ability to take exploratory risks without compromising ranking quality, a trend
consistent across datasets.
4 Conclusions
Our study introduces SAFERec, a novel model for the next-basket recommen-
dations. Inspired by SASRec, designed for the next-item prediction, we present
Title Suppressed Due to Excessive Length 7
a transformer-based architecture incorporating user frequencies of purchasing
particular items. Our experiments on three open-source datasets demonstrate
that SAFERec outperforms state-of-the-art methods for the next-basket recom-
mendations, achieving improvements of up to 8% in Recall@10 and up to 50% in
the UN@10 metric. Additionally, SAFERec surpasses TIFU-KNN by as much as
56% in Recall@100, all while maintaining scalability in large datasets through an
efficient mini-batch training approach. We hope our SAFERec model will encour-
age researchers and practitioners to explore the potential of transformer-based
methods in next-basket recommendations.
References
1. Abbattista, D., Anelli, V.W., Di Noia, T., Macdonald, C., Petrov, A.V.: En-
hancing sequential music recommendation with personalized popularity aware-
ness. In: Proceedings of the 18th ACM Conference on Recommender Systems.
p. 1168–1173. RecSys ’24, Association for Computing Machinery, New York, NY,
USA (2024). https://doi.org/10.1145/3640457.3691719,https://doi.org/10.
1145/3640457.3691719
2. Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.: Optuna: A next-
generation hyperparameter optimization framework. In: Proceedings of the 25th
ACM SIGKDD International Conference on Knowledge Discovery and Data Min-
ing (2019)
3. Ariannezhad, M., Jullien, S., Li, M., Fang, M., Schelter, S., de Rijke, M.: Recanet:
A repeat consumption-aware neural network for next basket recommendation in
grocery shopping. In: Proceedings of the 45th International ACM SIGIR Confer-
ence on Research and Development in Information Retrieval. pp. 1240–1250 (2022)
4. Ariannezhad, M., Li, M., Schelter, S., De Rijke, M.: A personalized neighborhood-
based model for within-basket recommendation in grocery shopping. In: Proceed-
ings of the Sixteenth ACM International Conference on Web Search and Data
Mining. pp. 87–95 (2023)
5. Armstrong, R.A.: When to use the b onferroni correction. Ophthalmic and Physi-
ological Optics 34(5), 502–508 (2014)
6. Faggioli, G., Polato, M., Aiolli, F.: Recency aware collaborative filtering for next
basket recommendation. In: Proceedings of the 28th ACM Conference on User
Modeling, Adaptation and Personalization. pp. 80–87 (2020)
7. Herlocker, J.L., Konstan, J.A., Terveen, L.G., Riedl, J.T.: Evaluating collaborative
filtering recommender systems. ACM Transactions on Information Systems (TOIS)
22(1), 5–53 (2004)
8. Hu, H., He, X., Gao, J., Zhang, Z.L.: Modeling personalized item frequency infor-
mation for next-basket recommendation. In: Proceedings of the 43rd international
ACM SIGIR conference on research and development in information retrieval. pp.
1071–1080 (2020)
9. Jannach, D., Jesse, M., Jugovac, M., Trattner, C.: Exploring multi-list user inter-
faces for similar-item recommendations. In: Proceedings of the 29th ACM Confer-
ence on User Modeling, Adaptation and Personalization. pp. 224–228 (2021)
10. Kaminskas, M., Bridge, D.: Diversity, serendipity, novelty, and coverage: a survey
and empirical analysis of beyond-accuracy objectives in recommender systems.
ACM Transactions on Interactive Intelligent Systems (TiiS) 7(1), 1–42 (2016)
8 O. Lashinin et al.
11. Kang, W.C., McAuley, J.: Self-attentive sequential recommendation. In: 2018 IEEE
international conference on data mining (ICDM). pp. 197–206. IEEE (2018)
12. Katz, O., Barkan, O., Koenigstein, N., Zabari, N.: Learning to ride a buy-cycle: A
hyper-convolutional model for next basket repurchase recommendation. In: Pro-
ceedings of the 16th ACM Conference on Recommender Systems. pp. 316–326
(2022)
13. Kou, Z., Manchanda, S., Lin, S.T., Xie, M., Wang, H., Zhang, X.: Modeling se-
quential collaborative user behaviors for seller-aware next basket recommendation.
In: Proceedings of the 32nd ACM International Conference on Information and
Knowledge Management. pp. 1097–1106 (2023)
14. Lerche, L., Jannach, D., Ludewig, M.: On the value of reminders within e-commerce
recommendations. In: Proceedings of the 2016 Conference on User Modeling Adap-
tation and Personalization. pp. 27–35 (2016)
15. Li, M., Ariannezhad, M., Yates, A., de Rijke, M.: Masked and swapped sequence
modeling for next novel basket recommendation in grocery shopping. In: Proceed-
ings of the 17th ACM Conference on Recommender Systems. pp. 35–46 (2023)
16. Li, M., Jullien, S., Ariannezhad, M., de Rijke, M.: A next basket recommendation
reality check. ACM Transactions on Information Systems 41(4), 1–29 (2023)
17. Liang, D., Krishnan, R.G., Hoffman, M.D., Jebara, T.: Variational autoencoders
for collaborative filtering. In: Proceedings of the 2018 world wide web conference.
pp. 689–698 (2018)
18. Liu, Y., Li, M., Ariannezhad, M., Mansoury, M., Aliannejadi, M., de Rijke, M.:
Measuring item fairness in next basket recommendation: A reproducibility study.
In: European Conference on Information Retrieval. pp. 210–225. Springer (2024)
19. Liu, Z., Liu, S., Zhang, Z., Cai, Q., Zhao, X., Zhao, K., Hu, L., Jiang, P., Gai, K.:
Sequential recommendation for optimizing both immediate feedback and long-term
retention. arXiv preprint arXiv:2404.03637 (2024)
20. Petrov, A., Macdonald, C.: Aligning gptrec with beyond-accuracy goals with rein-
forcement learning. arXiv preprint arXiv:2403.04875 (2024)
21. Petrov, A.V., Macdonald, C.: gsasrec: Reducing overconfidence in sequential rec-
ommendation trained with negative sampling. In: Proceedings of the 17th ACM
Conference on Recommender Systems. pp. 116–128 (2023)
22. Ren, P., Chen, Z., Li, J., Ren, Z., Ma, J., De Rijke, M.: Repeatnet: A repeat aware
neural recommendation machine for session-based recommendation. In: Proceed-
ings of the AAAI conference on artificial intelligence. vol. 33, pp. 4806–4813 (2019)
23. Rendle, S., Freudenthaler, C., Schmidt-Thieme, L.: Factorizing personalized
markov chains for next-basket recommendation. In: Proceedings of the 19th in-
ternational conference on World wide web. pp. 811–820 (2010)
24. Romanov, A., Lashinin, O., Ananyeva, M., Kolesnikov, S.: Time-aware item weight-
ing for the next basket recommendations. In: Proceedings of the 17th ACM Con-
ference on Recommender Systems. pp. 985–992 (2023)
25. Shao, Z., Wang, S., Zhang, Q., Lu, W., Li, Z., Peng, X.: A systematical evalua-
tion for next-basket recommendation algorithms. In: 2022 IEEE 9th International
Conference on Data Science and Advanced Analytics (DSAA). pp. 1–10. IEEE
(2022)
26. Shevchenko, V., Belousov, N., Vasilev, A., Zholobov, V., Sosedka, A., Semenova,
N., Volodkevich, A., Savchenko, A., Zaytsev, A.: From variability to stability: Ad-
vancing recsys benchmarking practices. arXiv preprint arXiv:2402.09766 (2024)
27. Singh, A., Hosein, P.: On the prediction of possibly forgotten shopping basket
items. In: Artificial Intelligence and Applied Mathematics in Engineering Problems:
Title Suppressed Due to Excessive Length 9
Proceedings of the International Conference on Artificial Intelligence and Applied
Mathematics in Engineering (ICAIAME 2019). pp. 559–571. Springer (2020)
28. Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W., Jiang, P.: Bert4rec: Sequential
recommendation with bidirectional encoder representations from transformer. In:
Proceedings of the 28th ACM international conference on information and knowl-
edge management. pp. 1441–1450 (2019)
29. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser,
Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro-
cessing systems 30 (2017)
30. Wang, B.L., Schelter, S.: Efficiently maintaining next basket recommendations un-
der additions and deletions of baskets and items. arXiv preprint arXiv:2201.13313
(2022)
31. Xin, X., Karatzoglou, A., Arapakis, I., Jose, J.M.: Supervised advantage actor-critic
for recommender systems. In: Proceedings of the Fifteenth ACM International
Conference on Web Search and Data Mining. pp. 1186–1196 (2022)
32. Yu, L., Sun, L., Du, B., Liu, C., Xiong, H., Lv, W.: Predicting temporal sets with
deep neural networks. In: Proceedings of the 26th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining. pp. 1083–1091 (2020)
