Kailun Wu

Kailun Wu, Zhangming Chan, Weijie Bian et al.

Exploration-Exploitation (E{\&}E) algorithms are commonly adopted to deal with the feedback-loop issue in large-scale online recommender systems. Most of existing studies believe that high uncertainty can be a good indicator of potential reward, and thus primarily focus on the estimation of model uncertainty. We argue that such an approach overlooks the subsequent effect of exploration on model training. From the perspective of online learning, the adoption of an exploration strategy would also affect the collecting of training data, which further influences model learning. To understand the interaction between exploration and training, we design a Pseudo-Exploration module that simulates the model updating process after a certain item is explored and the corresponding feedback is received. We further show that such a process is equivalent to adding an adversarial perturbation to the model input, and thereby name our proposed approach as an the Adversarial Gradient Driven Exploration (AGE). For production deployment, we propose a dynamic gating unit to pre-determine the utility of an exploration. This enables us to utilize the limited amount of resources for exploration, and avoid wasting pageview resources on ineffective exploration. The effectiveness of AGE was firstly examined through an extensive number of ablation studies on an academic dataset. Meanwhile, AGE has also been deployed to one of the world-leading display advertising platforms, and we observe significant improvements on various top-line evaluation metrics.

Ziang Li, Kailun Wu, Yiwen Guo et al.

Drawing on theoretical insights, we advocate an error-based thresholding (EBT) mechanism for learned ISTA (LISTA), which utilizes a function of the layer-wise reconstruction error to suggest a specific threshold for each observation in the shrinkage function of each layer. We show that the proposed EBT mechanism well disentangles the learnable parameters in the shrinkage functions from the reconstruction errors, endowing the obtained models with improved adaptivity to possible data variations. With rigorous analyses, we further show that the proposed EBT also leads to a faster convergence on the basis of LISTA or its variants, in addition to its higher adaptivity. Extensive experimental results confirm our theoretical analyses and verify the effectiveness of our methods.

Weijie Bian, Kailun Wu, Lejian Ren et al.

Feature interaction has been recognized as an important problem in machine learning, which is also very essential for click-through rate (CTR) prediction tasks. In recent years, Deep Neural Networks (DNNs) can automatically learn implicit nonlinear interactions from original sparse features, and therefore have been widely used in industrial CTR prediction tasks. However, the implicit feature interactions learned in DNNs cannot fully retain the complete representation capacity of the original and empirical feature interactions (e.g., cartesian product) without loss. For example, a simple attempt to learn the combination of feature A and feature B <A, B> as the explicit cartesian product representation of new features can outperform previous implicit feature interaction models including factorization machine (FM)-based models and their variations. In this paper, we propose a Co-Action Network (CAN) to approximate the explicit pairwise feature interactions without introducing too many additional parameters. More specifically, giving feature A and its associated feature B, their feature interaction is modeled by learning two sets of parameters: 1) the embedding of feature A, and 2) a Multi-Layer Perceptron (MLP) to represent feature B. The approximated feature interaction can be obtained by passing the embedding of feature A through the MLP network of feature B. We refer to such pairwise feature interaction as feature co-action, and such a Co-Action Network unit can provide a very powerful capacity to fitting complex feature interactions. Experimental results on public and industrial datasets show that CAN outperforms state-of-the-art CTR models and the cartesian product method. Moreover, CAN has been deployed in the display advertisement system in Alibaba, obtaining 12\% improvement on CTR and 8\% on Revenue Per Mille (RPM), which is a great improvement to the business.

Yuhai Song, Zhong Cao, Kailun Wu et al.

The idea of unfolding iterative algorithms as deep neural networks has been widely applied in solving sparse coding problems, providing both solid theoretical analysis in convergence rate and superior empirical performance. However, for sparse nonlinear regression problems, a similar idea is rarely exploited due to the complexity of nonlinearity. In this work, we bridge this gap by introducing the Nonlinear Learned Iterative Shrinkage Thresholding Algorithm (NLISTA), which can attain a linear convergence under suitable conditions. Experiments on synthetic data corroborate our theoretical results and show our method outperforms state-of-the-art methods.

Guorui Zhou, Kailun Wu, Weijie Bian et al.

Recently, click-through rate (CTR) prediction models have evolved from shallow methods to deep neural networks. Most deep CTR models follow an Embedding\&MLP paradigm, that is, first mapping discrete id features, e.g. user visited items, into low dimensional vectors with an embedding module, then learn a multi-layer perception (MLP) to fit the target. In this way, embedding module performs as the representative learning and plays a key role in the model performance. However, in many real-world applications, deep CTR model often suffers from poor generalization performance, which is mostly due to the learning of embedding parameters. In this paper, we model user behavior using an interest delay model, study carefully the embedding mechanism, and obtain two important results: (i) We theoretically prove that small aggregation radius of embedding vectors of items which belongs to a same user interest domain will result in good generalization performance of deep CTR model. (ii) Following our theoretical analysis, we design a new embedding structure named res-embedding. In res-embedding module, embedding vector of each item is the sum of two components: (i) a central embedding vector calculated from an item-based interest graph (ii) a residual embedding vector with its scale to be relatively small. Empirical evaluation on several public datasets demonstrates the effectiveness of the proposed res-embedding structure, which brings significant improvement on the model performance.