Source-linked AI summary
IntentGC: a Scalable Graph Convolution Framework Fusing Heterogeneous Information for Recommendation
Jun Zhao, Zhou Zhou, Ziyu Guan, Wei Zhao, Wei Ning, Guang Qiu, Xiaofei He
TL;DR
Sparse user-item interactions and incomplete use of heterogeneous auxiliary relationships limit recommendation, while prior linear regularizers require manual parameter tuning. IntentGC combines explicit preferences with heterogeneous relationships using graph convolutions, and IntentNet improves scalability by avoiding unnecessary feature interactions. Across offline experiments and Alibaba online A/B tests, the method outperforms the best baseline, including 95.1% higher offline MRR and 65.4% higher online CTR.
Problem
Sparse explicit user-item preferences and incomplete use of heterogeneous auxiliary relationships limit recommendation, while prior approaches rely on linear regularizers with handcraft parameter tuning.
Method
IntentGC jointly models explicit preferences and heterogeneous relationships with graph convolutions, while IntentNet separates neighborhood propagation from node feature interaction for scalable training.
Results
IntentGC improves offline MRR by 95.1% and online CTR by 65.4% compared to the best baseline, while IntentNet gains 22.1% in MRR and reduces running time by 75.6%.
Takeaways & Limitations
Heterogeneous auxiliary relationships are practically useful and complementary for predicting user preferences, and the framework supports billion-scale recommendation applications.
Takeaways & Limitations
GraphSage is unsuitable in this context because clustered mini-graph sampling can produce very large subgraphs for sparse, heterogeneous user-item graphs.
Abstract
from arXiv · showhide
The remarkable progress of network embedding has led to state-of-the-art algorithms in recommendation. However, the sparsity of user-item interactions (i.e., explicit preferences) on websites remains a big challenge for predicting users' behaviors. Although research efforts have been made in utilizing some auxiliary information (e.g., social relations between users) to solve the problem, the existing rich heterogeneous auxiliary relationships are still not fully exploited. Moreover, previous works relied on linearly combined regularizers and suffered parameter tuning. In this work, we collect abundant relationships from common user behaviors and item information, and propose a novel framework named IntentGC to leverage both explicit preferences and heterogeneous relationships by graph convolutional networks. In addition to the capability of modeling heterogeneity, IntentGC can learn the importance of different relationships automatically by the neural model in a nonlinear sense. To apply IntentGC to web-scale applications, we design a faster graph convolutional model named IntentNet by avoiding unnecessary feature interactions. Empirical experiments on two large-scale real-world datasets and online A/B tests in Alibaba demonstrate the superiority of our method over state-of-the-art algorithms.
1 INTRODUCTION
IntentGC addresses sparse user-item preferences by unifying explicit interactions with heterogeneous auxiliary relationships in graph convolutions. It also introduces IntentNet to make graph convolution more scalable while automatically learning relationship importance.
- Motivation: Sparse explicit user-item preferences limit recommendation, while websites contain heterogeneous auxiliary relationships that can reveal user interests and item semantics.Examples include query words, visited shops, preferred brands, and item properties.
- IntentGC framework: IntentGC captures explicit preferences and heterogeneous auxiliary relationships within a unified graph-convolution framework for recommendation.The paper describes this as its first framework for jointly modeling these information sources.
- Heterogeneous relationships: IntentGC translates first-order auxiliary connections into more robust second-order relationships and automatically learns their importance through a nonlinear neural model.For example, users sharing multiple query words form a stronger relationship than a single user-query connection.
- IntentNet: IntentNet separates neighborhood propagation from node feature interaction, avoiding unnecessary feature interactions for more efficient large-scale graph convolution.Its vector-wise convolution component propagates neighborhood features, while a fully connected component models node feature interactions.
- Evaluation: IntentNet achieves a 22.1% gain in MRR and a 75.6% reduction in running time, while IntentGC improves offline MRR by 95.1% and online CTR by 65.4% over the best baseline.The reported results combine graph-convolution efficiency gains with offline evaluation and Alibaba online A/B testing.
2 RELATED WORK
Prior recommendation research largely focused on network representation learning and graph convolution, while heterogeneous auxiliary information remained underused for user-item recommendation. IntentGC addresses this gap by combining explicit preferences with diverse auxiliary relationships and learning their importance automatically.
- Network embedding methods learn low-dimensional node representations that preserve network structure and properties for tasks including recommendation.
- Heterogeneous information networks provide rich information through multiple node and edge types, motivating methods such as meta-path walks and edge representations.
- Earlier recommendation work generally captured only limited auxiliary information, leaving many heterogeneous relationships unused.
- User-item recommendation directly predicts preferred items but is more challenging because user-item interactions are sparse.
- IntentGC exploits explicit preferences and rich heterogeneous auxiliary relationships while automatically determining the importance of different relationship types through graph convolutions.
3 PROBLEM DEFINITION
The paper formulates recommendation on a heterogeneous information network built from users, items, and other objects connected by typed relationships. The task is to predict user-item preference edges from historical data.
- A heterogeneous information network is an undirected graph with node and edge type mappings, containing multiple node types or edge types.
- Users and items are designated as the first two node types, while other objects include query words, brands, and related entities.
- User-item edges represent explicit preferences, whereas other edges encode auxiliary relationships in the network.
- The input is a historical-data HIN, and the output is a predicted edge set representing inferred preference edges.
4 METHODOLOGY
IntentGC translates heterogeneous auxiliary links into weighted within-type relationships, then uses efficient graph convolution and dual user-item networks to learn recommendation representations. Its IntentNet separates neighborhood propagation from node-level feature interaction, reducing computational cost relative to GraphSage.
- Method overview: IntentGC has three methodological components: network translation, faster convolutional networking, and dual graph convolution for users and items.
- Network Translation: Network translation converts shared auxiliary neighbors into weighted user-user or item-item second-order relationships while retaining user-item preference edges.
- Motivation: GCN computation can become impractical at web scale because stacked neighborhood aggregation requires many convolution operations per node.
- Vector-wise convolution: Each graph-convolution layer aggregates neighborhood features before applying a convolution function to combine self-node and neighborhood information.
- IntentNet: IntentNet uses vector-wise convolution for neighborhood utility and fully connected layers for interactions among node-feature dimensions.
- Complexity: IntentNet has time cost O(ρq−1∗m+m2), compared with GraphSage’s O(ρq−1∗m2), making it more efficient under the stated complexity analysis.
1: Network Translation:
The framework translates the original heterogeneous graph, trains separate user and item IntentNets with sampled positive and negative tuples, and performs approximate nearest-neighbor search after training. The section also reports dataset-statistics tabulation as part of the experimental setup.
- Framework: IntentGC comprises network translation, training, and inference.
- Network Translation: Network translation generates second-order relationships through auxiliary nodes and converts the original HIN into a user-item HIN.
- Training: Training initializes user and item feature matrices, samples positive user-item edges with negative items, computes representations, and updates parameters by triplet loss.
- Inference: Inference applies the trained IntentNets to all users and items, then recommends items using approximate K-nearest-neighbor search.
- Evaluation setup: Table 1 reports statistics for the two datasets used in the evaluation.
5 EXPERIMENTS
Experiments evaluate IntentGC and IntentNet through offline recommendation benchmarks, efficiency comparisons, heterogeneous-relationship studies, and Alibaba online A/B tests. The results report stronger performance, faster training, and benefits from combining heterogeneous auxiliary relationships.
- 5.3 Offline Evaluation: Experiments compare IntentGC with homogeneous embedding, GCN, explicit-preference, heterogeneous embedding, and joint-optimization recommendation methods.The evaluation uses AUC and MRR, with positive user-item pairs defined by clicks in Taobao and high ratings in Amazon.
- 5.3 Offline Evaluation: IntentGC outperforms the compared methods in Taobao, while adding heterogeneous auxiliary relationships improves AUC by 0.012 and MRR by 37.8% over IntentGC(Single).The same passage reports gains from GraphSage to DSPR and from BiNE to IntentGC(Single), while similar results are observed on Amazon.
- 5.3 Offline Evaluation: IntentNet trains IntentGC(Single/All) in 19/21 hours versus 78 hours for IntentGC(Single) with GraphSage on the Taobao setup.The design separates vector-wise graph convolution from dense node processing and avoids unnecessary feature interactions.
- 5.3 Offline Evaluation: IntentNet is more effective than the GraphSage version in both Taobao and Amazon, suggesting better fit without overfitting.The comparison uses the same number of training epochs and evaluates IntentGC(Single) variants.
- 5.3 Offline Evaluation: IntentGC(All) outperforms every single-type auxiliary-relationship version in both Taobao and Amazon, indicating complementary relationship information.The study evaluates relationship types generated from different object categories, including words and item/user links.
- 5.4 Online Evaluation: Online Alibaba A/B tests use scaled CTR because absolute CTR values cannot be exposed under Alibaba business policy.CTR is the advertising evaluation metric, and higher CTR indicates better advertising performance.
6 CONCLUSIONS
IntentGC captures heterogeneous auxiliary relationships for recommendation and shows their practical usefulness. It also includes a faster graph convolutional network for billion-scale applications.
- IntentGC is presented as the first framework for capturing heterogeneous auxiliary relationships in recommendation.
- Experiments show that heterogeneous relationships are practically useful for predicting users’ preferences and complementary to each other.
- The framework includes a faster graph convolutional network designed for billion-scale applications.
A.1 Hyper-parameter Settings
The experiments use shared hyper-parameter settings and provide practical guidance for learning rate, mini-batch size, and parameter initialization.
- Compared algorithms are not quite sensitive to learning rates between 0.001 and 0.0001.
- A mini-batch size range of 100-1000 is recommended, with 200 used in the experiments.
- The standard deviation used to initialize parameters is important because smaller values for network parameters might lead to bad local optima.
- All algorithms use the same hyper-parameter settings to support convergence and fair performance comparison.
A.2 Hardware and Software
The Amazon dataset uses high ratings as explicit preferences because user clicks are unavailable, with training and test data drawn from a broad time range.
- High ratings, defined as rating = 5, serve as explicit item preferences in the Amazon recommendation task.
- Because Amazon provides limited public data, the experiment uses a wide time range for both training and testing.
A.4 Feature Design
Feature design differs by dataset: Taobao uses richer platform records, while Amazon uses a smaller feature set. The experiments run in a distributed 200-machine environment implemented with TensorFlow and Python.
- Taobao features: Taobao user features include categories, brands, query words, profiles, member information, purchase statistics, visited shops, and preferred properties.
- Amazon features: Amazon uses smaller user and item feature sets built from categories, brands, words, ratings, metadata, and consumption statistics.
- Distributed environment: Each algorithm uses a 200-machine cluster per epoch, with 32 CPU cores and 128G memory per instance machine.
- Distributed environment: The models are implemented with TensorFlow v1.7 and Python v2.7, and code and environment instructions are provided for reproduction.
A.3 Dataset Preprocessing
The preprocessing pipeline constructs recommendation data from Taobao interactions, prepares node features, samples hard negatives, and supports scalable graph processing and distributed execution.
- Dataset construction: Taobao clicks serve as explicit preference labels, with October 11–17 used for training and October 19 for testing.October 18 is skipped to mirror the one-day training delay in online usage.
- Feature preprocessing: Co-features match user and item attributes, such as pairing item price with users’ average price.This design follows the link-prediction objective of mapping users closer to preferred items.
- Feature preprocessing: User and item features combine continuous values with embedding vectors obtained from dictionaries for discrete values.Sparse embedding lookup is used for features containing multiple discrete values.
- Negative sampling: Training uses multiple negative items from the same leaf category as each positive item to distinguish harder cases.The process discards sampled positives and repeats negative sampling five times per labeled edge.
- Scalable implementation: The implementation avoids complete second-order proximity computation because high-degree auxiliary nodes can create O(10^12) user-pair counts.Preprocessing runs offline through MapReduce, while training and inference run on a distributed IntentGC platform.