PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning to Recommend with Multiple Cascading Behaviors

Chen Gao, Xiangnan He, Dahua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat-Seng Chua, Lina Yao, Yang Song, Depeng Jin

arXiv:1809.08161v4cs.IR

TL;DR

Traditional recommenders often use only one behavior type, although other behaviors can provide valuable preference signals. NMTR combines neural collaborative filtering with shared embeddings, behavior-specific interaction functions, and joint multi-task training. Experiments on two real-world datasets report that NMTR outperforms the best existing methods on hit-ratio results.

  • Problem

    Traditional recommenders often use only one behavior type, despite other behaviors also providing valuable signals about user preference.

    Source: §1 · p. 1

  • Method

    NMTR combines neural collaborative filtering and multi-task learning, sharing embeddings across behaviors while learning separate interaction functions for each behavior type.

    Source: §1 · p. 2, §3.4 · p. 5

  • Results

    NMTR outperformed the best existing methods by 6.08% and 30.76% on the hit-ratio effect across two datasets.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper presents NMTR as a first step toward integrating the rich semantics of users' multiple behaviors into recommender systems.

    Source: §6 · p. 11

    Online evaluation through A/B tests remains future work, as does extending the approach to behaviors without a full-order cascade.

    Source: §6 · p. 11

Abstract

from arXiv · show

Most existing recommender systems leverage user behavior data of one type only, such as the purchase behavior in E-commerce that is directly related to the business KPI (Key Performance Indicator) of conversion rate. Besides the key behavioral data, we argue that other forms of user behaviors also provide valuable signal, such as views, clicks, adding a product to shop carts and so on. They should be taken into account properly to provide quality recommendation for users. In this work, we contribute a new solution named NMTR (short for Neural Multi-Task Recommendation) for learning recommender systems from user multi-behavior data. We develop a neural network model to capture the complicated and multi-type interactions between users and items. In particular, our model accounts for the cascading relationship among different types of behaviors (e.g., a user must click on a product before purchasing it). To fully exploit the signal in the data of multiple types of behaviors, we perform a joint optimization based on the multi-task learning framework, where the optimization on a behavior is treated as a task. Extensive experiments on two real-world datasets demonstrate that NMTR significantly outperforms state-of-the-art recommender systems that are designed to learn from both single-behavior data and multi-behavior data. Further analysis shows that modeling multiple behaviors is particularly useful for providing recommendation for sparse users that have very few interactions.

1 INTRODUCTION

The paper argues that recommendation should use multiple behavior types, whose semantics and ordinal relationships existing methods often overlook. NMTR combines neural interaction functions, cascading predictions, and multi-task learning; experiments on two real-world datasets report higher hit-ratio results than the best existing methods.

1 INTRODUCTION

  • Traditional recommenders often optimize one behavior, but other behaviors can also provide valuable signals about user preferences.Source: §1 · p. 1
  • Existing multi-behavior models largely ignore behavior semantics, including ordinal relations and informative intermediate feedback such as viewed-but-not-purchased products.Source: §1 · p. 1
  • The paper identifies CMF embedding designs as unreasonable because user interests should remain represented by shared embeddings across behavior types.It also criticizes fixed inner-product interactions as insufficient for complicated, multi-type interactions.Source: §1 · p. 2
  • NMTR shares user and item embeddings across behavior types, learns a separate data-dependent interaction function for each, and jointly trains the behavior tasks.The design combines neural collaborative filtering with multi-task learning.Source: §1 · p. 2
  • Cascaded predictions encode ordinal behavior relations by deriving a higher-level behavior prediction, such as purchase, from a lower-level one, such as view.The paper motivates this design with the relationship that a user must view a product to purchase it.Source: §1 · p. 2
  • NMTR outperformed the best existing methods by 6.08% and 30.76% on the hit-ratio effect across two datasets.The authors also report that further studies demonstrate the effectiveness of multi-task learning.Source: §1 · p. 2

2 PRELIMINARIES

The preliminaries formulate multi-behavior recommendation around predicting a target behavior from interaction data across behavior types. They review neural collaborative filtering and collective matrix factorization as relevant modeling approaches, noting limitations of CMF for this setting.

2 PRELIMINARIES

  • The preliminaries introduce the problem formulation, neural collaborative filtering, and collective matrix factorization as background for the proposed method.Source: §2 · p. 2

2.1 Problem Formulation

  • The task uses target-behavior interactions and other behavior data to estimate a user's likelihood of interacting with an item under the target behavior.Behavior types are assumed to have a total order, with the target behavior at the highest level; recommendations rank items by predicted scores.Source: §2.1 · pp. 2–3

2.2 Neural Collaborative Filtering (NCF)

  • NCF represents users and items with embeddings and feeds them into a neural interaction function to predict preference.MF is a special case using a fixed inner product; NCF also includes GMF, MLP, and NeuMF interaction designs.Source: §2.2 · p. 3

2.3 Collective Matrix Factorization

  • CMF correlates factorizations by sharing embeddings for entities common to multiple data matrices, but multi-behavior recommendation shares both users and items.The paper argues that behavior-specific user embeddings are irrational for a latent-interest representation, and notes fixed inner products and squared regression loss as further limitations.Source: §2.3 · p. 3

3 METHODS

NMTR predicts multiple behavior types from user-item pairs using shared embeddings, behavior-specific interaction functions, cascaded predictions, and joint multi-task training. The design aims to model behavior relationships while learning from all behavior types together.

3 METHODS

  • NMTR predicts behavior likelihoods from user-item pairs using four designs: shared embeddings, separated interaction functions, cascaded predictions, and multi-task learning.The model outputs predictions for each behavior type.Source: §3 · p. 3

3.1 Shared Embedding Layer

  • One embedding layer maps user and item IDs to latent vectors used to predict all behavior types.The paper uses static user embeddings because its focus is multiple interactions with one item over a relatively short period.Source: §3.1 · pp. 3–4

3.2 Separated Interaction Function

  • Each behavior type has a separate interaction function, built from neural collaborative filtering units, to model its likelihood from the same user-item input.The paper considers GMF, MLP, and NeuMF units; the choice may depend on the dataset.Source: §3.2 · p. 4

3.3 Cascaded Predictions

  • Cascaded predictions encode ordinal relations by making each behavior prediction depend on precedent behaviors; item-specific bias captures differences across behavior types.This design is intended to help predict target behavior for users with insufficient target-behavior data by using precedent behaviors that may be more plentiful.Source: §3.3 · pp. 4–5

3.4 Multi-Task Learning

  • Multi-task learning trains the behavior-specific models simultaneously, treating each behavior as a task so that the models can reinforce one another.Joint optimization updates parameters across behavior types together.Source: §3.4 · p. 5
  • The model assumes a full-order cascade, although some real-world behavior relations may instead be partial-order.The authors suggest ordering behaviors by strength in reflecting user preference as a possible adaptation and leave that exploration for future work.Source: §3.4 · p. 5
  • The joint objective weights each behavior's loss with λ_r, whose values are tuned by dataset, and constrains the weights to sum to one.The loss is minimized using stochastic gradient descent; mini-batches pair observed interactions with sampled negative instances.Source: §3.4 · p. 5

4 EXPERIMENTS

Across two real-world e-commerce datasets, the experiments compare NMTR with single- and multi-behavior baselines, examine auxiliary behaviors and training choices, and test performance for users with sparse purchase histories. NMTR outperforms the best baseline, with additional gains reported for sparse users.

4 EXPERIMENTS

  • The experiments address comparative performance, the effects of hyperparameters and multi-task learning, and whether NMTR helps users with sparse target-behavior data.The authors use two real-world datasets containing multiple types of e-commerce behaviors.Source: §4 · p. 6, §4.1.1 · p. 6

4.1.1 Datasets and Evaluation Protocol

  • Beibei includes views, carts, and purchases, while Tmall includes views and purchases; evaluation uses leave-one-out testing and ranks all previously unobserved items.The datasets contain strict behavior cascades, such as viewing before carting and carting before purchasing on Beibei.Source: §4.1.1 · p. 6

4.1.2 Baselines

  • The comparison includes single-behavior and multi-behavior methods, including BPR, neural collaborative filtering models, CMF, and MC-BPR.MC-BPR accounts for feedback levels when sampling negatives; the study also evaluates MC-NCF and NMTR variants using GMF, MLP, and NeuMF interactions.Source: §4.1.2 · pp. 6–7

4.1.3 Parameter Settings

  • All methods use 64-dimensional embeddings and a negative-sampling ratio of 4, with learning rates and other settings tuned on validation data.Neural models use a tower structure when they have multiple hidden layers; the number of layers is tuned from 1 to 5.Source: §4.1.3 · p. 7

4.2 Performance Comparison (RQ1)

  • NMTR achieves the best HR@K and NDCG@K across the baselines; versus MC-NeuMF, NMTR-NeuMF improves HR@100 by 6.08% on Beibei and 30.76% on Tmall.The reported improvements are statistically significant at p < 0.05; Table 3 covers HR@K and NDCG@K for K=50, 80, 100, and 200.Source: §4.2 · p. 7, Table 3 · p. 8

4.3 Impact of Auxiliary Behaviors and Parameters (RQ2)

  • View data improves purchase recommendations more than cart data, while using 50% of views remains better than purchase-only data and using 50% of carts is worse.The authors attribute this pattern to views adding intermediate feedback, whereas carts often carry signal similar to purchases.Source: §4.3 · p. 9, Table 5 · p. 9
  • NMTR’s best loss-coefficient settings are approximately (2/6,2/6,2/6) on Beibei and (0.4, 0.6) on Tmall; multi-task training also outperforms sequential training.The coefficient results are based on HR@100, and the multi-task versus sequential comparison uses HR@100 and NDCG@100.Source: §4.3 · pp. 9–10, Figure 4 · p. 10, Table 6 · p. 10

4.4 Impact of Data Sparsity (RQ3)

  • For users with 5–8 purchase records, NMTR exceeds the best baseline by 11.23% in HR@100 and 15.35% in NDCG@100.Across purchase-history groups, NMTR-NeuMF’s NDCG decreases 27.56% from the fifth to the first group, versus 40.09% for MC-BPR and 38.62% for MC-NeuMF.Source: §4.4 · p. 10, Figure 5 · p. 10

5 RELATED WORK

Prior work adapted matrix factorization and pairwise learning to multi-behavior recommendation, while other research used neural networks and multi-task learning in recommendation. NMTR extends neural collaborative filtering into a multi-task framework for learning from multiple behaviors.

5.1 Multi-Behavior Recommendation.

  • Multi-behavior approaches include collective matrix factorization and learning-based methods that adapt BPR sampling to different feedback types.Other learning-based methods use co-occurrence, item similarity, or margin-based pairwise learning.Source: §5.1 · p. 11

5.2 Neural Network Based Recommendation.

  • NMTR extends Neural Collaborative Filtering to a multi-task learning framework for recommendation from multi-behavior data.Prior neural recommendation work also applied NCF to other recommendation scenarios and used graph neural networks.Source: §5.2 · p. 11

5.3 Multi-task Learning for Recommendation.

  • Multi-task learning shares common representations across related tasks trained in parallel; NMTR applies this framework to multiple user behaviors.Earlier recommender-system work applied multi-task learning to user and item similarity, while another network shared text representations across users.Source: §5.3 · p. 11

6 CONCLUSION AND FUTURE WORK

The paper presents NMTR, combining neural collaborative filtering and multi-task learning to exploit multiple user behaviors, and reports experiments on two real-world datasets. It identifies online evaluation and settings without a full-order behavior cascade as directions for future work.

6 CONCLUSION AND FUTURE WORK

  • NMTR combines neural collaborative filtering and multi-task learning to exploit multiple user behaviors, with effectiveness demonstrated on two real-world datasets.The authors describe this as a first step toward integrating the semantics of users’ multiple behaviors into recommender systems.Source: §6 · p. 11
  • Future work includes online A/B evaluation, online and incremental learning, and recommendation when behaviors lack a full-order cascade.The authors also propose studying heterogeneous social behaviors and time-aware models of evolving user preferences.Source: §6 · p. 11

Related papers

Loading 1809.08161v4…