PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning from History and Present: Next-item Recommendation via Discriminatively Exploiting User Behaviors

Zhi Li, Hongke Zhao, Qi Liu, Zhenya Huang, Tao Mei, Enhong Chen

arXiv:1808.01075v1cs.IR

TL;DR

Existing recommenders do not fully capture both users’ long-term preferences and present motivations from interactive behavior sequences. BINN combines interaction-based neural item embedding with discriminative learning of session and preference behaviors, and experiments report that it outperforms state-of-the-art methods on two real-world datasets.

  • Problem

    Existing works mainly exploit short-term behaviors without fully accounting for users’ long-term stable preferences and their evolutions.

  • Method

    BINN uses interaction-based neural item embedding and discriminative learning of users’ session behaviors and preference behaviors.

  • Results

    BINN outperforms several state-of-the-art methods on two real-world datasets, Tianchi and JD.

  • Takeaways & Limitations

    Next-item recommendation can integrate historical preferences and present motivations learned from users’ interactive behaviors.

  • Takeaways & Limitations

    The paper identifies that many existing recommenders focus on static user–item relevance and do not specially attend to preference dynamics or present motivations.

Abstract

from arXiv · show

In the modern e-commerce, the behaviors of customers contain rich information, e.g., consumption habits, the dynamics of preferences. Recently, session-based recommendations are becoming popular to explore the temporal characteristics of customers' interactive behaviors. However, existing works mainly exploit the short-term behaviors without fully taking the customers' long-term stable preferences and evolutions into account. In this paper, we propose a novel Behavior-Intensive Neural Network (BINN) for next-item recommendation by incorporating both users' historical stable preferences and present consumption motivations. Specifically, BINN contains two main components, i.e., Neural Item Embedding, and Discriminative Behaviors Learning. Firstly, a novel item embedding method based on user interactions is developed for obtaining an unified representation for each item. Then, with the embedded items and the interactive behaviors over item sequences, BINN discriminatively learns the historical preferences and present motivations of the target users. Thus, BINN could better perform recommendations of the next items for the target users. Finally, for evaluating the performances of BINN, we conduct extensive experiments on two real-world datasets, i.e., Tianchi and JD. The experimental results clearly demonstrate the effectiveness of BINN compared with several state-of-the-art methods.

1 INTRODUCTION

BINN addresses next-item recommendation by combining users’ long-term preferences with short-term consumption motivations learned from interactive behaviors. It uses interaction-based item embedding and discriminative behavior learning, with experiments on two real-world datasets showing improved performance over state-of-the-art methods.

  • Customer interaction logs contain information about historical stable preferences and present consumption motivations.
  • Static recommenders overlook preference dynamics, while session-based models mainly use short-term behaviors.
  • BINN combines neural item embedding with discriminative behavior learning for next-item recommendation.
  • BINN integrates historical preferences and present motivations by discriminatively exploiting different user behavior types.
  • BINN’s experiments on the Tianchi and JD datasets outperform several state-of-the-art methods from various aspects.

2 RELATED WORKS

Related work comprises general recommenders that model static user–item relevance and sequential recommenders that model interaction sequences. The paper argues that both paradigms insufficiently integrate historical preference dynamics with present motivations.

  • General Recommenders: General recommenders commonly use collaborative filtering to mine static user–item relevance from interactions.
  • General Recommenders: Neighborhood methods recommend nearby items using precomputed entity similarities.
  • General Recommenders: Factorization models reconstruct user–item interaction matrices using dot products of latent vectors.
  • Sequential Recommenders: Sequential recommendation includes next-basket, session-based, and next-item scenarios, with session-based models receiving increasing attention.
  • Sequential Recommenders: Existing sequential models use sequential information but leave coherence and historical preference dynamics largely unexploited.
  • Sequential Recommenders: The paper proposes integrating historical preferences and present motivations learned from users’ interactive behaviors.

3 BINN: BEHAVIOR-INTENSIVE NEURAL NETWORK

BINN predicts a target user’s next item from sequential interactions by combining interaction-based item representations with separate session- and preference-behavior learning. Its w-item2vec component weights item frequency when learning sequential similarities.

  • Preliminaries: Personalized next-item recommendation predicts the item a target user is most likely to access next from sequential interactions.
  • BINN Overview: BINN contains Neural Item Embedding and Discriminative Behaviors Learning as its two main components.
  • Discriminative Behaviors Learning: Session Behaviors Learning models present motivations, while Preference Behaviors Learning models historical stable preferences.
  • Discriminative Behaviors Learning: BINN jointly learns the two behavior alignments over item representations and recommends top-k potentially preferred items.
  • Neural Item Embedding: w-item2vec generates item representations from users’ collaborative interactions and uses item frequency as a weighting factor.
  • Neural Item Embedding: The w-item2vec objective is based on Skip-gram with Negative Sampling for item sequences.

Preference Behaviors Learning...

The supplied passage contains only fragmentary notation associated with Session Behaviors Learning.

  • The passage includes item-index notation associated with the session-learning component.
  • The supplied text does not state the operation or interpretation of the displayed notation.

B. Discriminative Behaviors Learning

Discriminative Behaviors Learning separates present motivations from historical preferences and models them with complementary sequential architectures. These learned representations are jointly used to predict the next item.

  • Neural Item Embedding: w-item2vec converts sequentially interacting items into a unified embedding space whose vectors capture item similarities and sequential relationships.The method weights negative sampling by item frequency within interaction sequences.
  • Session Behaviors Learning: Session Behaviors Learning captures short-term present motivations from recent interaction behaviors, including clicks, collections, carts, and purchases.The session alignment models behavioral fluctuations over a recent sequence.
  • Preference Behaviors Learning: Preference Behaviors Learning extracts longer-term historical preferences using selected preference behaviors rather than every interaction type.The paper gives purchase as a behavior that can clearly depict underlying preference and defines a preference set containing collect, cart, and purchase.
  • Neural Architectures: BINN uses a Contextual LSTM for session behaviors and a bidirectional Contextual LSTM for global historical preferences.The bidirectional architecture uses contextual representations in both forward and backward directions, while the session model represents present motivations with its final hidden state.
  • Joint Prediction: The two learned behavior alignments are combined through a fully connected layer to generate the next-item representation, trained jointly with mean squared error.Optimization uses Adagrad over the sequential interaction history.

4 EXPERIMENTS

The experiments evaluate BINN through embedding visualization, recommendation performance, cold-start behavior, and interaction-parameter sensitivity on Tianchi and JD datasets. Both datasets are temporally split after filtering sparse users and items.

  • Experiments assess embedding quality, overall recommendation performance, cold-start scenarios, and sensitivity to user interactions.
  • The evaluation uses two real-world datasets: Tianchi and JD.
  • Tianchi2 contains 23,291,027 interactions from 20,000 customers across 4,758,484 items within one month.
  • JD contains 37,087,895 interactions from 105,180 customers across 28,710 items over 75 days.
  • Users with fewer than 10 interactions and items appearing fewer than 5 times are filtered before temporal train-test splitting.

4.2 Baseline Methods

BINN is compared with traditional recommenders and RNN-based next-item models under shared experimental settings. The baselines span popularity, factorization, neighborhood, recurrent, and hierarchical recurrent approaches.

  • BINN is compared with S-POP, BPR-MF, Item-KNN, GRU4Rec, and HRNN variants.
  • S-POP recommends the target user’s most-interacted item and is suited to settings with high repetitiveness.
  • BPR-MF optimizes a pairwise ranking objective through stochastic gradient descent, while Item-KNN recommends items similar to previously accessed items.
  • GRU4Rec uses a basic GRU with TOP1 loss and session-parallel minibatching, whereas GRU4Rec Concat feeds each full interaction sequence independently.
  • HRNN Init models cross-session interest evolution with an additional GRU layer, while HRNN All also propagates its user representation at each step.
  • All RNN baselines use 100 hidden units, 0.1 dropout, and 0.1 learning rate; BINN uses 64-dimensional item embeddings.

4.3 Evaluation Metrics

Next-item recommendation quality is evaluated with Recall@20 and MRR@20. Recall measures whether the desired item appears in the top-20 list, while MRR additionally accounts for its rank.

  • The experiments evaluate personalized next-item recommendation quality using Recall@20 and MRR@20.
  • Recall@20 is the proportion of test cases where the desired item appears among the top-20 recommendations.
  • Recall@20 does not distinguish among rankings within the top-20 candidate set.
  • MRR@20 averages reciprocal ranks, assigning zero when the desired item ranks below 20.
  • Higher values on both metrics indicate better recommendation performance.

4.4 Experimental Results

BINN is evaluated through embedding visualization, overall next-item recommendation, cold-start testing, and history-length analysis on Tianchi and JD. It achieves the best overall performance and benefits from modeling both personalized long-term preferences and short-term sequential behavior.

  • Item Embedding Visualization: w-item2vec produces clearer category clustering than item2vec on Tianchi and better clustering on JD.The authors attribute this partly to accounting for item frequencies.
  • Cold Start of New Users: BINN performs better than other models for most new-user cold-start cases, while all deep-learning models show capacity to handle new users.BINN initially deteriorates to CLSTM when personalized preferences are absent, then improves as interactions accumulate.
  • Recommendation Performances: BINN achieves the best overall next-item recommendation performance on both Tianchi and JD.On Tianchi, it performs significantly better than other methods on Recall@20 and MRR@20.
  • Recommendation Performances: Personalized models outperform non-personalized alternatives on Tianchi MRR@20, highlighting the value of tracking long-term user preferences.BINN, HRNN Init, and HRNN All outperform GRU4Rec and GRU4Rec Concat on this metric.
  • Recommendation Performances: On JD, all RNN-based models outperform traditional methods, while GRU4Rec Concat surpasses HRNN All.The result indicates that improper personalization can reduce recommendation performance and that short-term community trends matter.
  • Analysis on the User History Length: 3.92% is BINN's minimum improvement over other RNN-based models for JD users with more than 500 interactions.BINN improves with increasing history length, whereas GRU4Rec and GRU4Rec Concat do not improve continuously across the evaluated ranges.

5 CONCLUSIONS AND FUTURE WORKS

The paper concludes that BINN addresses personalized next-item recommendation by combining sequential item representations with discriminative modeling of historical preferences and present motivations. Future work will examine different behavior types and extend the model to other domains.

  • Conclusions: BINN combines w-item2vec item representations with LSTM-based alignments for historical preferences and present consumption motivations.The framework was evaluated on two industrial datasets for personalized next-item recommendation.
  • Future Works: Future work will study how different user-behavior types affect user representations and investigate applications in domains such as advertising.
Loading 1808.01075v1…