All transductive recommender systems are unable to make predictions for users who were not included in the training sample due to the process of learning user-specific embeddings. In this paper, we propose a new method for replacing identity-based user embeddings in existing sequential models with interaction-based user vectors trained purely on interaction sequences. Such vectors are composed of user interactions using GRU layers with adjusted dropout and maximum item sequence length. This approach is substantially more eficient and does not require retraining when new users appear. Extensive experiments on three open-source datasets demonstrate noticeable improvement in quality metrics for the most of selected state-of-the-art sequential recommender models.
Recommender systems are widely used in various online tainment platforms. These services gather large amounts of sequential data, including the history of interactions quire learning the ID-based latent user vectors, which are supposed to represent both short-term and long-term preferences based on user-specific information and previous history of interactions. However, there are several drawbacks to this approach.
to users from the training set. The predictions cannot be obtained only based on previous interactions of outof-sample users because the model’s user embeddings depend on users’ IDs and additional features (if provided). The problem of making recommendations for new users is solved either by fully retraining the model on the updated data or iterative training on new batches [1]. For industrial purposes, the retraining process on large-scale data is time- and space-consuming, constantly afecting the user coverage with recommendations and the quality of service for new users. nEvelop-O In addition, it does not require storage of peruser embeddings and is therefore more storage eficient and scalable. • We have comprehensively reviewed existing works in three A and B-ranked conference series (RecSys, CIKM, and SIGIR) in 2019-2021 that use identity-based user embeddings in architectures.
used for recommendation tasks on serial data. Most of them are based on recurrent neural networks (RNNs), for instance, GRU4Rec [36], SASRec [37], and SHAN [38].
Additionally, Transformers4Rec [39] is gaining popularity in usage for sequential and session-based tasks.
Some architectures attempt to model temporal decay efects in user interaction history in order to improve the relevance of recommendations. Customer needs as well as both short-term and long-term preferences change over time, which should be taken into account in the predictions. Intuitively, the most recent interactions should have greater weight than older ones in deciding on the next item. Additionally, users may require substitutions or supplements for an already acquired item. These assumptions have been incorporated into the design of the 3. METHODOLOGY SLRC [40], Chorus [41] and KDA [42] models.
The approaches mentioned above have serious lim- 3.1. The rationale for Using User-specific itations for industrial applications: item recommenda- Vectors tions are made based on user-specific embeddings, which can be trained only for users that were included in the training set. In contrast, inductive learning models can provide recommendations for out-of-sample users, who have interactions but are not included in the training process. For instance, Mult-VAE [43] and CF-LGCN-E [44], which is modified version of LightGCN [ 45] for inductive learning mode, can provide predictions for users outside of the training sample. Nevertheless, the quality of inductive models is often lower than that of transductive ones.
Thus, one of the open challenges for transductive models which show high performance is to overcome the problem of making predictions for out-of-sample users.
Additionally, the efect of user-specific embeddings on the quality of recommendations is not yet suficiently studied. In this research, we propose researching whether we really need user-specific embeddings or if it is better to train ID- and feature-free user vectors based solely on previous item interactions.
To determine how frequently trainable user-specific vectors are used in existing sequential recommender models and to systematize the reasons for their use, we examined the proceedings of scientific conferences with relevant store pre-computed embeddings in a look-up matrix with articles. The summary in Table 1 shows that our anal- users’ IDs as each vector can be derived on-the-fly from ysis research includes articles that were presented be- the input interaction sequence using the learned neural tween 2019 and 2021 in three conference series: RecSys, network weights. It addresses the scalability issues for CIKM, and SIGIR. A paper was considered relevant if it commercial applications. Secondly, it can be regarded as proposed a sequential recommendation model, including a step toward users’ privacy and confidentiality, because session-based and POI recommendation tasks, and if their a user identifier is redundant information, and without usperformance was compared to sequential recommender ing it we can not map it back to personal data. Lastly, our model baselines. As a result, we compiled a list of 34 approach allows adapting previously introduced sequenrelevant studies, 12 of which contain applications of user tial recommender models to inductive learning scenarios, embeddings. According to the authors, the primary pur- when we can infer the recommendations for the users, pose for including user vector processing in the proposed who were not included in a training sample. methods, which appeared in five studies, was to represent long-term preferences. Other objectives included model- 3.3. Models ing both short-term and long-term preferences, learning user-specific vectors from mixed representations of all users sharing the same account, and modeling the impact of friends’ behavior.
popular frameworks for sequential recommendation models - ReChorus2 and RecBole3. For the experimental setup, we have selected state-of-the-art models that 3.2. Initialization of User Vectors have proven themselves in many new research papers as reliable baselines for comparison with new models.
Each sequential model processes the history of users’ Thus, we selected three models from the ReChorus interactions in order to represent relationships between framework - KDA, Chorus, and SLRC - and two modinteractions and then model the user’s behavioral pat- els from RecBole - SHAN and HGN - in order to study terns [36, 37]. how diferent user vector initialization techniques afect
In our work, we investigate the feasibility of employing model performance on three open-source datasets. this method to obtain vector representations that reflect Two RecBole models have been implemented in Reusers’ interests, as well as how it influences the quality of Chorus to ensure a fair comparison of the models. sequential models. All of the selected models transform the user ID into a low-dimensional real-valued dense • Sequential Hierarchical Attention Network vector representation u ∈ ℝ , where d is the dimension (SHAN) [38] is a two-layer hierarchical attention of the user embedding. The embedding is then processed network. The attention mechanism is needed to in accordance with the architecture of each model. assign altered weights of items for the user to
Instead of using this technique, we propose to re- capture the dynamic property, while the hierarplace the ID-based user embedding initialization with chical structure integrates the user’s long- and interaction-based user embedding initialization, suggest- short-term preferences. User embedding vector ing that the user ID be discarded as limiting information is used as context information to obtain various for eficiency and scalability. weights for diferent users.
Let Su be the input representations of previous interac- • Hierarchical Gating Network (HGN) [47] contions Su ∈ ℝ× , where L is the maximum history length. sists of three parts: feature gating, instance gatFirst, we apply a Dropout layer [46] to the matrix Su . ing, and item-item product modules. The feaThe sequence representation is then processed by GRU ture gating module allows the adaptive selection layers. Note that our goal is to show that our approach is of efective latent features based on user interefective even with a simple recurrent layer like GRU. The ests. At the instance gating module, items that use of more advanced layers is left for future improve- reflect short-term user preferences are selected ments. The final step is to use a linear layer to reduce and passed down to lower layers along with item the embedding dimension to its initial size d and take the features. User embedding is used in both feature last known vector of interaction history. As seen on Fig- gating and instance gating modules. ure 1, we obtain a user embedding u ∈ ℝ as the output • Chorus [41] incorporates the representation of of successively applied layers: Dropout layer, two GRU diferent sequence contexts by knowledge and layers, and Dense layer, the input of which is a sequence time-aware item modeling. The constructed temof each user’s historical data. poral kernel functions modify the temporal dy
This approach has several significant advantages. First, namics of relations by representing two sorts of the space complexity is optimized from () to (1) , 2https://github.com/THUwangcy/ReChorus where is the number of users. There is no need to 3https://recbole.io/ 4.1. Datasets
We chose the three datasets most commonly used for #items #actions #density sequential recommendation: MovieLens-1M 4, Amazon3,416 1M 4.84% Grocery and Gourmet Food and Amazon-Electronics5. 41,280 1,1M 0.022% These open-source datasets have diferent domains, sizes, 63,001 1,7M 0.014% and sparsity. They contain user interaction sequences with timestamps and item metadata, including the list of also view and also buy relations in Amazon datasets and items - substitutes and complements - and al- the list of genres in the MovieLens data set. We use a lowing relational representations to contribute common leave-one-out strategy with 99 negative items, diferentially to the final item embedding. User similar to [42]. For SHAN, HGN, and SLRC, we only need embeddings are used in both the BPR and GMF user interaction sequences, while Chorus and KDA are approaches for making predictions. based on knowledge graphs, so we use metadata to build • Short-Term and Life-Time Repeat Consump- them. In Amazon datasets, we simply use the relations of tion (SLRC) [40] model uses the Hawkes Process also view and also buy, provided in the metadata data set, and Collaborative Filtering, which requires learn- as was done in [41]. We chose the most popular movies ing user embeddings to distinguish between user of the same genre as the equivalent of also view items for interests and help explore new items. Consider- the MovieLens data set, and the most popular items in ing the lack of recurrent interactions in the Ama- the set of movies that the user has watched right after the zon and MovieLens datasets, we use this model to ground-truth item as the equivalent of also buy items. derive substitutive and complementary types of relations between items, as implemented in the
SLRC model in Chorus. 4.2. Evaluation Metrics • Knowledge-aware Dynamic Attention (KDA) Hit Ratio (HR@k) and Normalized Discounted Cumula[42] takes both item relations and their temporal tive Gain (NDCG@k) were used as evaluation metrics, evolution into account. The core idea of KDA is to where k = [5, 10, 20, 50]. HR@k measures whether at aggregate the sequence of interactions into multi- least one ground-truth item appears in the top-k recomple relation-specific embeddings via an attention mendation list, whereas NDCG@k considers both the mechanism. Fourier transform with trainable fre- position and relevance of the item in the recommendaquency domain embeddings was used in a novel tion list. The values of NDCG@10 and HR@10 for 5 way to simulate the diverse temporal efects of original and 5 modified models are presented in Table 3. various relational interactions. User vectors, as well as item vectors and interaction representa- 4.3. Experiment Settings tions, are used in the final ranking score.
Overall, we selected all models stated above and com- work [48]. For a fair comparison, we set the embedding pared the original architectures with the architectures size to 64, batch size to 256, and the maximum history without user-specific vectors, based on our approach of length to 20 for all models and datasets, similar to experlearning only from interaction sequences. iments in [41]. Additionally, we demonstrate the results of experiments for other values of the maximum history 4. EXPERIMENTS length: 10, 30, and 50. Other hyperparameters are dependent on the model and are set to their default values the In this section, we introduce our experimental setup and same as in the original implementations. The tuning of compare the performance of original models with modi- the hyperparameters across all methods and datasets is ifed ones. Our experiments are designed to answer the left for future work. following research questions: RQ1: Does the proposed method have a positive efect on 4.4. Baselines the quality of existing sequential recommender models? RQ2: How does the maximum sequence length afect the models’ performance?
performance of non-sequential methods. Specifically, we include the POP method [49] which is a common nonpersonal baseline that recommends the most popular
ℎ ℎ
ML-1M
Grocery&Gourmet
Electronics NDCG@10
NDCG@10
NDCG@10 0.2513 0.4074 items. Additionally, we add a BPR-MF [50] approach that NDCG@10 has decreased by 16%. As the authors of is often adopted as a classic matrix factorization-based HGN observed, the predictions of this model are highly method. dependent on the last items. When our approach constructs a user vector from long sequences, the impact of 4.5. Performance Comparison the last items may be reduced. On Amazon Electronics we can see decreasing of metric values by 1% for As can be seen, Table 3 shows the recommendation per- and by 5% for , while for , and formance of original architectures and modified models ℎ the quality improved in range of 7% to 12%. on three datasets (RQ1). The proposed strategy has a The overall performance of the four models improved significant impact on model quality across all datasets. dramatically, but changing user-specific vectors had alFor instance, on MovieLens-1M we can see increases in most no influence on the model. According to a both NDCG@10 and HR@10 for 4 modified models, com- research article on KDA, one possible explanation is that pared to the original ones: , , the architecture of KDA [42] is not highly sensitive to and ℎ , while for quality remains nearly the presence of user vectors at all. demonstrates the same. The quality improvement varies widely, rang- significant improvement in quality for all datasets. It ing from 1% for to 81% for . We even ob- is explained by the fact that the SLRC algorithm’s core serve a slight boost in evaluation metrics for the strongest component is collaborative filtering (CF), which is good baseline, KDA, on Amazon Grocery&Gourmet Food. for modeling long-term user preferences. Our technique
However, the quality of has deteriorated: allows us to evaluate short-term preferences in CF, which the original model may have overlooked. If we consider sequential models instead of user-specific embeddings. each of the model-dataset pairs as a separate experiment, Our method does not require constant retraining of the our approach dramatically increases the quality metrics model as the number of users increases, and is memoryin 11 out of 15 cases. eficient. Extensive experiments on 3 real-world datasets
Summing up, comparative experiments on three real- reveal that the majority of evaluated models were imworld datasets show the efectiveness of our approach proved in quality. Additionally, we studied the relationand significant improvement of quality for the majority ship between the model’s relative improvement and item of examined models. A new method with replaced user- sequence length when our method is applied. Thus, we specific embeddings provides a significant relative gain suggest researchers experiment with our approach in in performance (e.g., 0.6% − 12.1% for SHAN [38], 1.1% − their studies by using ID-based user-specific embeddings. 38.7% for Chorus [41], 6.6% − 80% for SLRC[40]). Our results can open up a new research area for ablation
Figure 2 shows how the maximum history length influ- studies on the use of user-specific embeddings in recomences quality improvement when our approach is applied mender systems. In the future, we are going to apply our (RQ2). The smaller the maximum sequence length, the approach to more modern models and try more complex better the model captures user short-term preferences, architectures than GRU. In addition, it is essential to inwhile long-term efects outweigh short-term efects for vestigate how high-quality and stable this approach is larger lengths. When the length of the sequence shrinks, with an extremely small number of user interactions. the long-term influence of modeling, which is the primary reason for using user embeddings in a model, disappears.
As a result, replacing user-specific vectors works efec- Acknowledgements tively for both short (l = 10) and long (l = 50) sequences.
and the Laboratory for Model and Methods of
Computational Pragmatics at the National Research University
Higher School of Economics (HSE). The contribution of
Dmitry I. Ignatov to the article was done within the
framework of the HSE University Basic Research Program.
[22] M. M. Tanjim, Dynamicrec: A dynamic convolu- ence on Research and Development in Information
tional network for next item recommendation, in: Re
Proceedings of the 29th ACM International Confer- [32] L. Zheng, N. Guo, W. Chen, J. Yu, D. Jiang,
ence on Information and Knowledge Management Sentiment-guided sequential recommendati
ment, 2019, pp. 2301–2304. [35] M. Ma, P. Ren, Y. Lin, Z. Chen, J. Ma, M. d.
Ri[26] F. Lv, T. Jin, C. Yu, F. Sun, Q. Lin, K. Yang, W. Ng, jke, -net: A parallel information-sharing network
Sdm: Sequential deep matching model for online for shared-account cross-domain sequential
recomlarge-scale recommender system, in: Proceedings mendations, in: Proceedings of the 42nd
Interof the 28th ACM International Conference on In- national ACM SIGIR Conference on Research and
formati
mation Re
Wen, Lighter and better: low-rank decomposed [40] C. Wang, M. Zhang, W. Ma, Y. Liu, S. Ma,
Modelself-attention networks for next-item recommenda- ing item-specific temporal dynamics of repeat
contion, in: Proceedings of the 44th International ACM sumption for recommender systems, in: The World
SIGIR Conference on Research and Development Wide Web C