Disparity-preserved Deep Cross-platform Association for Cross-platform Video Recommendation

Shengze Yu, Xin Wang, Wenwu Zhu, Peng Cui, Jingdong Wang

Abstract

Cross-platform recommendation aims to improve recommendation accuracy by associating information from different platforms. However, existing approaches often assume cross-platform information is perfectly consistent and can be easily aligned. This paper identifies two major challenges: platform-specific disparity and differences in semantic granularity. We propose the Disparity-preserved Deep Cross-platform Association (DCA) model, which utilizes a partially-connected multi-modal autoencoder to explicitly capture platform-specific information while handling granularity differences through nonlinear mapping. Extensive experiments on real-world datasets demonstrate that DCA significantly outperforms existing methods in cross-platform video recommendation, achieving notable improvements in precision and F-score.

Introduction: The Cross-Platform Challenge

In the era of information explosion, cross-platform recommendation has emerged as a promising solution to the data sparsity and "cold-start" problems. By linking user data across different social media environments—such as Twitter and YouTube—recommender systems can gain a more holistic view of user preferences. For instance, if a user is new to a video platform, their historical interactions on a microblogging site can provide valuable cues for initial suggestions.

However, the authors of this study argue that current cross-platform association models are built on a flawed premise: the Consistency Fallacy. Most existing models assume that a user's interests are uniform across all platforms and can be mapped via simple linear transformations. This oversight ignores the reality of platform-specific personas and the varying levels of semantic detail inherent to different media types.

The Consistency Fallacy and Granularity Mismatch

The researchers identify two primary obstacles to effective cross-platform association:

  1. Platform-Specific Disparity: Users often exhibit different behaviors depending on the platform's purpose. A user might engage with professional news and political discourse on Twitter while consuming entertainment or hobbyist content on YouTube. Forcing these distinct interest sets into a single "unified" latent representation introduces noise and degrades recommendation quality.
  2. Semantic Granularity Mismatch: Different platforms operate at different "resolutions." Twitter tags and hashtags are often hyper-specific to trending events (fine-grained), whereas YouTube categories and genres tend to be broader (coarse-grained). Linear models are often too rigid to capture the complex, nonlinear relationships between these different levels of semantic meaning.
Visualization of user interest distribution across Twitter and YouTube, highlighting the inconsistency in cross-platform clustering.

As illustrated in the figure above, the researchers conducted a measurement study using t-SNE visualizations. They found that users who form tight, coherent clusters on one platform (e.g., Twitter) often appear scattered and disconnected when their interests are mapped to another (e.g., YouTube). This empirical evidence—quantified as a "Concentration Ratio"—validates the need for a model that explicitly accounts for platform-specific disparity.

The DCA Model: Preserving Disparity through Architecture

To address these challenges, the paper introduces the Disparity-preserved Deep Cross-platform Association (DCA) model. The core of DCA is a partially-connected multi-modal autoencoder designed to separate shared interests from platform-specific noise.

Partially-Connected Architecture

Unlike standard autoencoders where hidden layers are fully connected, DCA employs a split architecture. The hidden layer is divided into three distinct parts:

  • $h^T$ (Twitter-specific): Captures latent features unique to the user's Twitter persona.
  • $h^Y$ (YouTube-specific): Captures latent features unique to the user's YouTube persona.
  • $h^C$ (Common): Captures the overlapping interests shared across both platforms.

By strategically restricting the connections in the neural network, the model ensures that platform-specific "noise" is isolated in $h^T$ and $h^Y$, preventing it from polluting the common association space $h^C$. This allows $h^C$ to serve as a much cleaner and more accurate bridge for cross-platform transfer.

Nonlinear Mapping and Missing Data

To handle the granularity mismatch, DCA utilizes nonlinear activation functions (such as Sigmoid). This allows the model to learn complex mappings, such as translating a specific tweet about a new gadget into a broader interest in "Technology" videos. Furthermore, to make the model robust for real-world applications where a user might only have data on one platform, the researchers trained the network using augmented data. By replacing missing inputs with average values during training, the model learns to reconstruct or predict the missing platform's interests effectively.

Experimental Validation

The researchers evaluated the DCA model using the CASIA-crossOSN dataset, which contains linked accounts between Twitter and YouTube. After filtering for active users, the final dataset comprised 7,338 users and over 34,000 videos. Features were extracted using Latent Dirichlet Allocation (LDA) to generate topic distributions (60 dimensions for Twitter, 80 for YouTube).

Performance Metrics

DCA was tested against several baselines, including Linear Regression (LR), Latent Attribute-based Association (LA), and standard Multi-Layer Perceptrons (MLP).

  • Association Accuracy: DCA significantly reduced error rates (MAE and RMSE) compared to all baselines when predicting a user's YouTube interests from their Twitter data.
  • Recommendation Performance: In a top-10 video recommendation task, DCA achieved a Precision of 0.367 and an F-score of 0.296. This represents a 5.3% improvement in F-score over the next best nonlinear model (MLP) and a 6.1% improvement over the state-of-the-art LA method.
  • Ablation Study: To confirm the importance of the architecture, the researchers tested a version of the model without the split hidden layer (called MA). The MA model performed significantly worse than DCA, proving that the "disparity-preserved" structure is the key driver of the model's success.

Conclusion

The findings of this paper suggest that the key to better cross-platform recommendation is not to force consistency, but to embrace disparity. By providing a dedicated latent space for platform-specific behaviors, the DCA model creates a more robust and accurate association for shared interests. While the current study focuses on the Twitter-YouTube link, the underlying principles of partially-connected architectures and nonlinear mapping offer a promising path forward for multi-platform recommendation systems.

Summary updated

Abstract

Cross-platform recommendation aims to improve recommendation accuracy through associating information from different platforms. Existing cross-platform recommendation approaches assume all cross-platform information to be consistent with each other and can be aligned. However, there remain two unsolved challenges: i) there exist inconsistencies in cross-platform association due to platform-specific disparity, and ii) data from distinct platforms may have different semantic granularities. In this paper, we propose a cross-platform association model for cross-platform video recommendation, i.e., Disparity-preserved Deep Cross-platform Association (DCA), taking platform-specific disparity and granularity difference into consideration. The proposed DCA model employs a partially-connected multi-modal autoencoder, which is capable of explicitly capturing platform-specific information, as well as utilizing nonlinear mapping functions to handle granularity differences. We then present a cross-platform video recommendation approach based on the proposed DCA model. Extensive experiments for our cross-platform recommendation framework on real-world dataset demonstrate that the proposed DCA model significantly outperform existing cross-platform recommendation methods in terms of various evaluation metrics.
BibTeX
@inproceedings{yu2019disparity,
  author = {Shengze Yu and Xin Wang and Wenwu Zhu and Peng Cui and Jingdong Wang},
  title = {{Disparity-preserved Deep Cross-platform Association for Cross-platform Video Recommendation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=wBACYj8MXxK},
}
Figure