GAN-based Cold-Start Solution for New Video Recommendation in Short Video Systems

Due to the cold‑start problem in recommendation systems, recommending newly uploaded videos is a highly challenging task. The effectiveness of new‑video recommendation directly influences the stability of the system’s “metabolism” and the health of the content ecosystem. This article introduces the practical solution adopted by iQIYI’s SuiKe recommendation team for new‑video cold start in short‑video recommendation, based on Generative Adversarial Networks (GAN).

Background

The cold‑start problem refers to the difficulty of providing good recommendations when a new item or a new user enters the system without any historical interaction data. This issue can degrade new‑user experience and retention, and is a key challenge in personalized short‑video recommendation.

Cold start can be divided into three categories: user cold start, item (video) cold start, and system cold start. In a user‑item rating matrix, both users and items can be “cold” or “warm”, forming four quadrants; the first quadrant corresponds to the item‑cold‑start problem.

New‑video recommendation is a typical item‑cold‑start problem. On iQIYI’s platform, massive numbers of new videos (especially UGC and PUGC short videos) are produced continuously. These videos lack historical interaction data and may have missing or inaccurate meta attributes, making it difficult to describe them accurately. Moreover, the distribution path of a new video involves an initial exploratory phase followed by large‑scale free distribution; optimizing this path and efficiently reaching target users is essential.

From a system perspective, new‑video recommendation drives the “metabolism” of the recommendation pool. Because short videos are consumed quickly, the material turnover rate is high, and the system must continuously introduce fresh content while keeping the overall recommendation efficiency stable.

Therefore, solving new‑video recommendation requires addressing two fundamental questions: (1) how to represent a new video, and (2) how to connect the new video with interested users.

Model Framework Design

To tackle the above challenges, we propose a GAN‑based cold‑start solution. The generator creates a latent user‑feature vector from video attributes, while the discriminator judges whether a user‑video pair is real or generated. The overall objective follows the classic min‑max formulation of GANs.

The GAN consists of two parts:

Generator : receives the k‑dimensional attribute vector of each candidate video and generates a potential user feature vector. It contains multiple sub‑generators, each handling a specific attribute, followed by a multi‑layer fully‑connected network that merges the sub‑features into a final user vector.

The generator is trained by minimizing its part of the GAN objective while keeping the discriminator fixed.

Multiple generators help alleviate mode collapse, a common issue where the generator produces overly similar samples.

Discriminator : a single network that receives either (generated‑user, video) pairs or (real‑user, video) pairs. The former are labeled 0 (negative), the latter 1 (positive). The discriminator learns to output the probability that a pair is real.

The discriminator is trained by maximizing the same GAN objective, thereby learning to distinguish real from generated user features while ensuring relevance to the given video.

Recommendation Generation

After the generator produces a user feature vector that is likely to like the target video, the system computes cosine similarity between this generated vector and all real users. The video is then recommended to the most similar real users. Approximate Nearest Neighbor (ANN) techniques are employed to accelerate similarity search.

Online Effect

The solution has been deployed in iQIYI’s main short‑video recommendation app. Online A/B tests show significant improvements:

Exposure share of new videos increased, with video freshness up by 12.4%.

Click‑through rate (CTR) improved by 11.9%.

Average watch time per user rose by 56.5%.

Conclusion and Outlook

Our GAN‑based cold‑start solution has demonstrated substantial gains in new‑video recommendation performance. Future work includes:

Enhancing the fusion of attribute features from multiple generators, e.g., by introducing attention mechanisms.

Exploring multi‑discriminator architectures to increase feature diversity and avoid model degeneration.

References

[1] Sun, Changfeng, et al. "LARA: Attribute‑to‑feature adversarial learning for new‑item recommendation." Proceedings of the 13th International Conference on Web Search and Data Mining. 2020.

[2] Goodfellow, Ian, et al. "Generative adversarial networks." Communications of the ACM 63.11 (2020): 139‑144.

[3] Arjovsky, Martin, Soumith Chintala, and Léon Bottou. "Wasserstein generative adversarial networks." International conference on machine learning. PMLR, 2017.

[4] Ebesu, Travis, and Yi Fang. "Neural semantic personalized ranking for item cold‑start recommendation." Information Retrieval Journal 20.2 (2017): 109‑131.

[5] Gantner, Zeno, et al. "Learning attribute‑to‑feature mappings for cold‑start recommendations." 2010 IEEE International Conference on Data Mining. IEEE, 2010.