1. Fusion of GAN and recommendation system:

(From paper APL)

  • Antagonistic Angle: GAN is easy to gradient disappearance, update instability. The sparsity of data will reduce the performance of adversarial training.
  • Recommended Angle: the generation problem of discrete projects. When GAN is generated, it is designed with differentiable values, and the discrete type cannot be optimized by gradient descent. To solve this problem, there are two sides to the problem. SeqGAN (Yu Zhang, Wang, &Yu, 2017) regarded discrete data generation actions as random strategies and applied strategic strategies (Sutton et al. , 2000) to estimate the generator gradient. Another approach (Jang, Gu, Poole, 2017; Maddison, Mnih, Teh, 2017) simulates the discrete project generation process using differentiable processes and directly optimizes the model through back propagation.

2. Development history

  1. 2017 full mark paper IRGAN is the first application of GAN in the recommendation system. At the beginning, the aim was to unify the two models in the field of retrieval. The generator predicts which documents are most relevant to the query, and the discriminator determines the relevance of the two documents sent to the query.
  2. APR published in SIGIR in August 2018 improved BPR (Bayesian Personalized Ranking). The adversarial idea of GAN is integrated into BPR, and perturbation is added into the model (according to FGSM algorithm), which finally improves the robustness of the model.
  3. CFGAN, CIKM, October 2018. Some problems in IRGAN are solved and collaborative filtering is optimized. CFGAN focuses on “purchased” goods, that is, items that the framework has lost part of the generator and the user has not actually purchased. The limitation of Disctete item index genneration is solved. For example, item 1 is marked with no real data, but is marked with no generated data. The discriminator will get interfered, which degrades performance.
  4. APL presented in ScienceDirect in March 2019 improves the stability and convergence effect of adversarial learning. Unlike IRGAN, IRGAN solves discrete item generation using the strategy gradient algorithm mentioned in SeqGAN. APL uses a differentiable program to simulate discrete item generation and optimize the model through back propagation.
  5. Adversarial Attacks on an Oblibious Recommender published at Rescys in 2019 reconsidered the Adversarial Attacks against recommendations from the perspective of machine learning optimization. It is assumed that the recommender system does not know the existence of the attacker, and the attacker knows the loss function and parameter representation of the recommender system, but does not know the gradient. GAN’s generator is used to generate spurious interaction matrices to preserve the user’s rating distribution. Projection gradient descent method is used to solve the gradient problem.
  6. AugCF alleviates the sparsity problem in collaborative filtering by augmentation data in 2019. More subtly, the discriminator of AugCF has two processes: 1) the generator generates data, and the discriminator determines whether the data is true or false; 2) The generator fixes and trains the discriminator to predict what the user likes or dislikes. This method also uses gumbel-Softmax method to solve the discrete sampling problem. (Same as the solution in APL)
  7. In 2019 RAGAN (BT) mitigated the sparsity problem by populating data. The main solution is the impact of rating bias during data filling (OCCF problem: collaborative filtering that only gives positive examples. Only positive feedback and no negative feedback). Generally, people will give scores to the ones they like, so the score data may be relatively high, affecting the filling effect. Therefore, CDAE was used to select negative feedback samples, and then GAN was used for data enhancement. Finally, AutoRec method was used for recommendation. This is also the step of proposing the method.