Adversarial for Sequential Recommendation Walking in the Multi-Latent Space

Abstract

Recently, sequential recommendation plays a critical role in our daily life, since it serves as personalized information filters to dis- cover popular users’ preferred products over time. Due to the success of the adversarial learning, a mass of research efforts start to strengthen sequential recommendation by the adversarial learning, which is able to learn complex underlying data distribution.

However, existing adversarial sequential recommendation methods suffer from mode collapse and unexplained prediction. To boost the diversity, performance, and interpretability of sequential recommendation system, we propose a novel temporal-aware adversarial framework, namely TSRGAN.

In principle, the input of traditional adversarial-based recommendation system is a noise variable sampled from normal distribution. We argue that it is hard to generate an item cover complex users’ preferences(e.g. price, brand and item style) using a single latent space. Therefore, our model employs multiple latent space to generate plausible item which matches user’ preferences from multiple views(e.g. Movie style, Movie release date).

Besides, previous adversarial-based recommenders focus on generating active item, but they omits that user’s favour is not in- variable. With GANs terminology, the recommenders only will be rewarded when seeking the peak mode, but it neglects minor mode, in other word mode collapse. In order to alleviate this issue, we design a novel diversity reward function and diversify regularization to encourage the model exploring minor mode over time and guarantee generating diversity item with reasonable.

Concretely, we propose multiple learnable latent codes to generate item matching user’s preferences from different views, then we leverage the diversity reward signal to shape the distribution of multiple latent space over time. It means that the multiple latent space are sampled form different distribution instead of Gaussian distribution. Such a manipulation of the latent space can be treated as walking from plain distribution latent space to diversity distributions latent space. Further, the reward signal is modified over time, therefore, our methods names "Temporal-aware" adversarial framework.

In short, our model has two sequential stages: encode the user’ characteristics and historical behaviours under multiple latent space with the Self Attention-based generator(G), and discriminator(D) try to distinguish the generator’s output item from the ground ruth. Besides, discriminator attempt to apply reward signal to shape the latent space distribution time by time. Extensive experiments demonstrate remarkable performance with interpretability improvement against the state-of-the-art baselines.