PaperlayerView exact revision on arXiv

Source-linked AI summary

RepeatNet: A Repeat Aware Neural Recommendation Machine for Session-based Recommendation

Pengjie Ren, Zhumin Chen, Jing Li, Zhaochun Ren, Jun Ma, Maarten de Rijke

arXiv:1812.02646v1cs.IR

TL;DR

Session-based recommendation must account for repeated consumption, yet prior neural approaches did not explicitly emphasize when to repeat versus explore. RepeatNet adds a repeat-explore mechanism to an encoder-decoder model and outperforms state-of-the-art methods across three datasets, with larger gains as dataset size and repeat ratio increase.

  • Problem

    Repeat consumption is common in recommendation scenarios, but prior studies did not emphasize it with neural networks or provide an effective neural approach for deciding when to repeat.

  • Method

    RepeatNet encodes a session, learns repeat-versus-explore switch probabilities, and uses separate decoders for historical-item repeats and new-item exploration.

  • Results

    RepeatNet outperforms state-of-the-art baselines in MRR and Recall on all three benchmark datasets, with improvements increasing as dataset size and repeat ratio increase.

  • Takeaways & Limitations

    Explicitly modeling repeat consumption gives RepeatNet advantages in handling repeat recommendation scenarios.

  • Takeaways & Limitations

    RepeatNet can overproduce repeat recommendations when mode probabilities are learned entirely from data, suggesting a need for prior knowledge.

Abstract

from arXiv · show

Recurrent neural networks for session-based recommendation have attracted a lot of attention recently because of their promising performance. repeat consumption is a common phenomenon in many recommendation scenarios (e.g., e-commerce, music, and TV program recommendations), where the same item is re-consumed repeatedly over time. However, no previous studies have emphasized repeat consumption with neural networks. An effective neural approach is needed to decide when to perform repeat recommendation. In this paper, we incorporate a repeat-explore mechanism into neural networks and propose a new model, called RepeatNet, with an encoder-decoder structure. RepeatNet integrates a regular neural recommendation approach in the decoder with a new repeat recommendation mechanism that can choose items from a user's history and recommends them at the right time. We report on extensive experiments on three benchmark datasets. RepeatNet outperforms state-of-the-art baselines on all three datasets in terms of MRR and Recall. Furthermore, as the dataset size and the repeat ratio increase, the improvements of RepeatNet over the baselines also increase, which demonstrates its advantage in handling repeat recommendation scenarios.

1 Introduction

Session-based recommendation applies to interactions grouped within time windows, where repeat consumption is common and can represent a large share of activity. RepeatNet addresses this pattern with a repeat-explore encoder-decoder mechanism and outperforms state-of-the-art baselines across three benchmark datasets.

  • Sessions group user interactions within a given time frame and may span days, weeks, or months.
  • Repeat consumption is common across e-commerce, restaurants, music, and other recommendation scenarios, sometimes accounting for a large proportion of interactions.
  • RepeatNet uses an encoder-decoder structure with separate repeat and explore modes to recommend historical or new items.
  • The model learns mode-switch probabilities and item probabilities jointly through end-to-end back-propagation.
  • RepeatNet outperforms state-of-the-art baselines on three benchmark datasets in both MRR and Recall.
  • The paper presents RepeatNet as a neural model that explicitly incorporates repeat consumption into session-based recommendation.

2 Related Work

Prior session-based recommenders include Markov models and recurrent neural networks, while repeat-recommendation studies examine consumption patterns and reminders. RepeatNet differs by explicitly modeling repeat consumption with a neural recommendation model.

  • Session-based recommendation: Related work on session-based recommendation includes Markov-chain methods that predict subsequent actions from sequential patterns or item transitions.
  • Session-based recommendation: Recurrent neural networks were applied to session-based recommendation and improved on conventional methods, with later work incorporating clicks, item features, and cross-session information.
  • Repeat recommendation: RepeatNet emphasizes repeat consumption in its neural models, unlike the cited preceding studies.
  • Repeat recommendation: Repeat-consumption research studies repeated behavior across domains, including the role of recency and contextual features in predicting short-term repeats.
  • Repeat recommendation: Recommendation systems support both discovering new items and reminding users about previously viewed or consumed items.

3 RepeatNet

RepeatNet predicts the next session item by combining repeat and explore modes within an encoder-decoder architecture. It encodes the session, learns mode probabilities, and uses separate decoders for historical and new items.

  • Session-based recommendation predicts the next event from an interaction session, with P(i_t+1 | I_S) denoting the recommendation probability.
  • 3.1 Framework: RepeatNet models the next-item probability as a mixture of repeat and explore mode probabilities and their corresponding item probabilities.The modes are represented by P(r | I_S) and P(e | I_S), while the conditional item probabilities select recommendations within each mode.
  • 3.1 Framework: The architecture contains a session encoder, repeat-explore mechanism, repeat decoder, and explore decoder.The encoder produces latent session representations; the mechanism predicts mode probabilities, and the decoders predict items for each mode.
  • 3.2 Session encoder: A GRU encodes each session into hidden states, using item embeddings and recurrent state updates before producing the sequence representation H.The initial GRU state is a zero vector.
  • 3.3 Repeat-explore mechanism: The repeat-explore mechanism applies attention over encoder states and softmax regression to estimate the repeat and explore mode distribution.The last hidden state supplies the matching signal for attention, whose normalized scores form a fixed-length session context vector.
  • 3.4 Repeat recommendation decoder: The repeat decoder assigns probabilities to re-clicking items in the session, while the explore decoder uses item-level attention to recommend items absent from the session.The explore decoder sets the probability of an existing session item to zero; repeated occurrences are aggregated in the repeat decoder.
  • 3.6 Objective function: RepeatNet jointly optimizes recommendation and mode-prediction negative log-likelihood losses, with the final loss formed as their linear combination.The mode assumption labels an item as repeat when it exists in the input session and explore otherwise; parameters and item embeddings are learned end-to-end.

4 Experiments

Experiments use three benchmark datasets spanning e-commerce and music, compare RepeatNet with conventional and neural baselines, and report percentage results in Table 3.

  • Datasets: The experiments use YOOCHOOSE, DIGINETICA, and LASTFM, covering e-commerce and music artist recommendation.YOOCHOOSE and DIGINETICA are e-commerce datasets; LASTFM is used for music artist recommendation.
  • Datasets: YOOCHOOSE and DIGINETICA filter out sessions of length 1 and items appearing fewer than 5 times.LASTFM keeps the 40,000 most popular artists and filters sessions longer than 50 or shorter than 2 items.
  • Implementation: The model uses 100-dimensional item embeddings and GRU hidden states, dropout with p = 0.5, Adam optimization, and mini-batches of 1024.Parameters are initialized with Xavier initialization, and gradient clipping uses the range [−5, 5].
  • Methods used for comparison: The comparison includes conventional baselines such as POP, S-POP, Item-KNN, BPR-MF, and FPMC.POP uses globally popular training-set items, while S-POP recommends popular items from the current session.
  • Methods used for comparison: Neural baselines include GRU4REC, Improved-GRU4REC, and NARM, with NARM adding a neural attention mechanism.These models are selected as recent state-of-the-art neural session-based recommendation baselines.
  • Evaluation: Table 3 reports experimental results as percentages across the three datasets.The table caption identifies the results as experimental results (%) on the three datasets.

5 Results and Analysis

RepeatNet consistently outperforms conventional and neural baselines, with advantages that generally increase as repeat consumption becomes more prevalent. Its repeat mechanism drives gains mainly on repeated sessions, while attention contributes differently and joint mode supervision can worsen performance.

  • 5.1 Results: RepeatNet outperforms conventional methods and neural baselines, including GRU4REC-TOPK and NARM, on the reported evaluation metrics.The authors attribute these improvements to explicitly modeling repeat consumption.
  • 5.1 Results: As the repeat ratio increases, RepeatNet’s performance generally increases, with a larger advantage on YOOCHOOSE than DIGINETICA.Both datasets are from e-commerce, but YOOCHOOSE has the higher repeat ratio.
  • 5.1 Results: RepeatNet has a bigger advantage in the music domain than in the e-commerce domain, while conventional baselines vary in relative strength across datasets.S-POP performs better than Item-KNN on LASTFM, whereas Item-KNN performs better than S-POP on YOOCHOOSE.
  • 5.2 Analysis of the repeat mechanism: 33.91% and 24.16% are RepeatNet’s Recall@20 improvements on repeated sessions for DIGINETICA and LASTFM, respectively.The improvements mainly come from repeated sessions, while performance drops slightly on non-repeated sessions.
  • 5.3 Analysis of the attention vs repeat mechanism: Repeat and attention mechanisms contribute complementarily: their combination improves all metrics and datasets, with attention helping Recall and repeat helping MRR.Both mechanisms improve results over Improved-GRU4REC.
  • 5.4 Analysis of joint learning: Jointly training Lrec and Lmode slightly reduces overall performance because Lrec already supervises mode prediction and Lmode can intensify repeat recommendations.The mode-label assumption that an item appearing in the session history is generated in repeat mode is not always reasonable.

6 Conclusion and Future Work

RepeatNet addresses repeat consumption in session-based recommendation with an encoder-decoder architecture and a repeat-explore mechanism. Experiments on three datasets show improved MRR and Recall, while future work includes adding prior knowledge, richer information, and applying variants to other recommendation tasks.

  • RepeatNet uses an encoder-decoder architecture with a repeat-explore mechanism for session-based recommendation.
  • RepeatNet better captures repeat-or-explore recommendation intent within a session.
  • Experiments on three datasets show that RepeatNet outperforms state-of-the-art methods in MRR and Recall.
  • Future extensions include incorporating prior knowledge, adding metadata and collaborative-filtering factors, and applying variants to content-based recommendation.

Code

The paper shares the code used to run its experiments to facilitate reproducibility.

  • The authors share the code used to run the experiments.
  • The code is available at the RepeatNet GitHub repository.
Loading 1812.02646v1…