Future Data Helps Training: Modeling Future Contexts for Session-based Recommendation

Abstract

Long session-based recommender systems have attacted much attention recently. For each user, they may create hundreds of click behaviors in short time. To learn long session item dependencies, previous sequential recommendation models resort either to data augmentation or a left-to-right autoregressive training approach. While effective, an obvious drawback is that future user behaviors are always mising during training. In this paper, we claim that users' future action signals can be exploited to boost the recommendation quality. To model both past and future contexts, we investigate three ways of augmentation techniques from both data and model perspectives. Moreover, we carefully design two general neural network architectures: a pretrained two-way neural network model and a deep contextualized model trained on a text gap-filling task. Experiments on four real-word datasets show that our proposed two-way neural network models can achieve competitive or even much better results. Empirical evidence confirms that modeling both past and future context is a promising way to offer better recommendation accuracy.

A Peep into the Future: Adversarial Future Encoding in Recommendation

“…It randomly masks some tokens in a sequence, and then predicts them via the rest contexts. Some models [2,21,26,38,39] adopt the MLM task in sequencebased recommendation, indirectly considering the future behaviors after the current position via MLM in training. However, there are optimization biases between MLM and next-item prediction tasks.…”

“…Recently, with the thriving of pre-training, some works introduce the masked language model (MLM) pre-training task in NLP [7] to sequence-based recommendation, which consider bidirectional information to learn better sequential models via selfsupervised learning (SSL) [26,40]. The MLM task randomly masks some items in the user behavior sequence during training, and then attempts to predict them with the remaining unmasked contexts [38,39]. These models can be viewed as indirectly using a specific future feature (i.e., behaviors after the masked item) under SSL in sequence-based recommendation.…”

Learning Heterogeneous Temporal Patterns of User Preference for Timely Recommendation

“…We used the adaptive fusing method for SLi-Rec, which shows the best performance in [28]. • GRec [29]: GRec is a state-of-the-art sequential recommender system without the time information, which leverages future data in a sequence as well for richer information in dilated convolutional neural networks. • TiSASRec [17]: TiSASRec is a state-of-the-art recommender system that utilizes the time information, which considers the time intervals between consecutive interactions in a user's interaction sequence in a self-attention network.…”

Unsupervised Proxy Selection for Session-based Recommender Systems

“…• CSRM [37] utilizes memory networks to incorporate the neighbor sessions of the input session. • GRec [42] leverages future data in a session as well when learning the preference of the session for richer information in dilated convolutional neural networks. • GCE-GNN [38] is the state-of-the-art SRS that constructs a global graph that models pairwise item-transitions over all sessions as well as the session graphs.…”