Adaptive User Modeling with Long and Short-Term Preferences for Personalized Recommendation

Abstract

User modeling is an essential task for online recommender systems. In the past few decades, collaborative filtering (CF) techniques have been well studied to model users' long term preferences. Recently, recurrent neural networks (RNN) have shown a great advantage in modeling users' short term preference. A natural way to improve the recommender is to combine both long-term and short-term modeling. Previous approaches neglect the importance of dynamically integrating these two user modeling paradigms. Moreover, users' behaviors are much more complex than sentences in language modeling or images in visual computing, thus the classical structures of RNN such as Long Short-Term Memory (LSTM) need to be upgraded for better user modeling. In this paper, we improve the traditional RNN structure by proposing a time-aware controller and a content-aware controller, so that contextual information can be well considered to control the state transition. We further propose an attention-based framework to combine users' long-term and short-term preferences, thus users' representation can be generated adaptively according to the specific context. We conduct extensive experiments on both public and industrial datasets. The results demonstrate that our proposed method outperforms several state-of-art methods consistently.

Sequential Recommender via Time-aware Attentive Memory Network

“…However, the dot-product correlation limits the capability of neural recommenders. To learn deeper non-linear relationships between a target user and candidate items, some more expressive models have been proposed for the ranking stage, such as neural collaborative ltering [14], deep interest network [39], SLi-Rec [37] and user memory network [4]. In this paper, we focus on the candidate generation stage and the purposed MTAM be viewed as a non-linear generalization of factorization techniques.…”

“…Research [40] improves LSTM by proposing some temporal gates to capture both long-term and short term preferences of users. Researches [5,37] further propose two timeaware recurrent units. The main di erence between our proposed T-GRU and previous models is we only use the time interval between adjacent interaction to control how much past information can be transferred to future states, while previous researches apply temporal gates to control both previous information and current content.…”

“…Therefore, classical structures of recurrent and attentive networks need to be upgraded to model the temporal dynamics of sequential data better. Some recent researches improve GRU or LSTM units by adding time gates to capture the temporal information of user behavior sequences [5,37,40]. However, the temporal information is often ignored in attention mechanism.…”

“…Meanwhile, for better short-term preference modeling, we utilize the proposed temporal gate in GRU to control how much past information can be transferred to future states, named T-GRU. Di erent from previous time-aware recurrent recommenders which aim to capture both the short-term and long-term interests [5,37,40], we only aim at short-term intent by ltering out irrelevant past information with the temporal gate. Additionally, both long-term preferences and short-term intents determine the behaviors of users.…”

Pairwise-Based Hierarchical Gating Networks for Sequential Recommendation

“…Convolutional Sequence embedding(Caser) [6] captures sequential patterns and transitions from previous item sequence by convolutional operation with various filters. Another popular method RNN is also used to model user's sequential interactions because RNN is good at capturing transition patterns in sequence [16,17]. Attention Mechanisms have been incorporated into next item recommendation to model complex transitions for better recommendation [18].…”