Online display advertising platforms rely on pre-ranking systems to efficiently filter and prioritize candidate ads from large corpora, balancing relevance to users with strict computational constraints. The prevailing two-tower architecture, though highly efficient due to its decoupled design and pre-caching, suffers from cross-domain interaction and coarse similarity metrics, undermining its capacity to model complex user-ad relationships. In this study, we propose the Hierarchical Interaction-Enhanced Two-Tower (HIT) model, a new architecture that augments the two-tower paradigm with two key components: generators that pre-generate holistic vectors incorporating coarse-grained user-ad interactions through a dual-generator framework with a cosine-similarity-based generation loss as the training objective, and multi-head representers that project embeddings into multiple latent subspaces to capture fine-grained, multi-faceted user interests and multi-dimensional ad attributes. This design enhances modeling effectiveness without compromising inference efficiency. Extensive experiments on public datasets and large-scale online A/B testing on Tencent's advertising platform demonstrate that HIT significantly outperforms several baselines in relevance metrics, yielding a 1.66% increase in Gross Merchandise Volume and a 1.55% improvement in Return on Investment, alongside similar serving latency to the vanilla two-tower models. The HIT model has been successfully deployed in Tencent's online display advertising system, serving billions of impressions daily. The code is available at https://github.com/HarveyYang123/HIT_model.

The incidence of comorbidities exacerbates the affliction of patients with Type 2 Diabetes Mellitus (T2DM) and increases the complexity of therapeutic interventions for physicians, resulting in adverse patient outcomes. Accurate identification and prediction of the comorbidity risk in T2DM facilitates improved clinical diagnostic decisions and enhanced patient prognoses. Existing studies frequently utilize single-task learning (STL) models to assess individual disease risks in T2DM comorbidity, overlooking the intricate relationships between various diseases and thus failing to capture their inherent connections, which leads to inaccurate risk predictions. To this end, we propose an end-to-end multi-gate mixture-of-self-attention-based-experts (MMAE) model under the multi-task learning (MTL) scheme. The proposed MMAE model is based on the self-attention mechanism and incorporates a new comorbidity diffusion coefficient (CDC) index to characterize correlations between diseases. We conduct comparative experiments against eight baselines on a real-world dataset, and the results have demonstrated the superiority of the proposed MMAE. Our ablation experiments also show that the incorporation of the CDC index helps accurately identify high-risk patients. Furthermore, the proposed MMAE model can provide clinical decision support for physicians in assisting them in making more effective comorbidity diagnoses. This study has significant implications for clinical decision support and improving patient prognostic outcomes.

Full-complexity machine learning models, such as the deep neural network, are non-traceable black-box, whereas the classic interpretable models, such as linear regression models, are often over-simplified, leading to lower accuracy. Model interpretability limits the application of machine learning models in management problems, which requires high prediction performance, as well as the understanding of individual features’ contributions to the model outcome. To enhance model interpretability while preserving good prediction performance, we propose a hybrid interpretable model that combines a piecewise linear component and a nonlinear component. The first component describes the explicit feature contributions by piecewise linear approximation to increase the expressiveness of the model. The other component uses a multi-layer perceptron to increase the prediction performance by capturing the high-order interactions between features and their complex nonlinear transformations. The interpretability is obtained once the model is learned in the form of shape functions for the main effects. We also provide a variant to explore the higher-order interactions among features. Experiments are conducted on synthetic and real-world datasets to demonstrate that the proposed models can achieve good interpretability by explicitly describing the main effects and the interaction effects of the features while maintaining state-of-the-art accuracy.

In real-time bidding systems, ad exchanges and supply-side platforms (SSP) are switching from the second-price auction (SPA) to the first-price auction (FPA), where the advertisers should pay what they bid if they win the auction. To avoid overpaying, advertisers are motivated to conceal their truthful evaluations of impression opportunities through bid shading methods. However, advertisers are consistently facing a trade-off between the probability and cost-saving of winning, due to the information asymmetry, where advertisers lack knowledge about their competitors' bids in the market. To address this challenge, we propose a Bayes ian Multi-Armed Bandit (BayesMAB) algorithm for bid shading when the winning price is unknown to advertisers who lose the impression opportunity. BayesMAB incorporates the mechanism of FPA to infer each price interval's winning rate by progressively updating the market price hidden by SSP. In this way, BayesMAB better approximates the winning rates of price intervals and thus is able to derive the optimal shaded bid that balances the trade-off between the probability and cost-saving of winning the impression opportunity. We conducted large-scale A/B tests on Tencent's online display advertising platform. The cost-per-mile (CPM) and cost-per-action (CPA) decreased by 13.06% and 11.90%, respectively, whereas the return on investment (ROI) increased by 12.31% with only 2.7% sacrifice of the winning rate. We also validated BayesMAB's superior performance in an offline semi-simulated experiment with SPA data sets. BayesMAB has been deployed online and is impacting billions of traffic every day. Codes are available at https://github.com/BayesMAB/BayesMAB.

The real-time bidding (RTB) paradigm allows the advertisers to submit a bid for each impression in online display advertising. A usual demand of the advertisers is to maximize the total value of winning impressions under constraints on some key performance indicators. Unfortunately, the existing RTB research in industrial applications can hardly achieve the optimum due to the stochastic decision scenarios and complex consumer behaviors. In this study, we address the application of RTB to mobile gaming where the in-app purchase action is of high uncertainty, making it challenging to evaluate individual impression opportunities. We first formulate the bidding process into a constrained optimization problem and then propose an actor-critic reinforcement learning (ACRL) model for obtaining the optimal policy under a dynamic decision environment. To avoid feeding too many samples with zero labels to the model, we provide a new way to quantify impression opportunities by integrating the in-app actions, such as conversion and purchase, and the characteristics of the candidate ad inventories. Moreover, the proposed ACRL learns a Gaussian distribution to simulate the audience's decision in a more real bidding scenario by taking additional contextual side information about both media and the audience. We also introduce how to deploy the learned model online to help adjust the final bid. At last, we conduct comprehensive offline experiments to demonstrate the effectiveness of ACRL and carefully set an online A/B testing experiment. The online experimental results verify the efficacy of the proposed ACRL in terms of multiple critical commercial indicators. ACRL has been deployed in the Tencent online display advertising platform and impacts billions of traffic every day. We believe proposed modifications for optimal bidding problems in RTB are practically innovative and can inspire the relative works in this field.

Ordinal regression predicts the objects’ labels that exhibit a natural ordering, which is vital to decision-making problems such as credit scoring and clinical diagnosis. In these problems, the ability to explain how the individual features and their interactions affect the decisions is as critical as model performance. Unfortunately, the existing ordinal regression models in the machine learning community aim at improving prediction accuracy rather than explore explainability. To achieve high accuracy while explaining the relationships between the features and the predictions, we propose a new method for ordinal regression problems, namely the Explainable Ordinal Factorization Model (XOFM). XOFM uses piecewise linear functions to approximate the shape functions of individual features, and renders the pairwise features interaction effects as heat-maps. The proposed XOFM captures the nonlinearity in the main effects and ensures the interaction effects’ same flexibility. Therefore, the underlying model yields comparable performance while remaining explainable by explicitly describing the main and interaction effects. To address the potential sparsity problem caused by discretizing the whole feature scale into several sub-intervals, XOFM integrates the Factorization Machines (FMs) to factorize the model parameters. Comprehensive experiments with benchmark real-world and synthetic datasets demonstrate that the proposed XOFM leads to state-of-the-art prediction performance while preserving an easy-to-understand explainability.

For the task of information retrieval from massive online reviews, people may be faced to some challenges in feature extraction, and then aspects summarization from these features. In this paper, by combining two methods of word vector representing and k-means clustering, an unsupervised method for product aspects summarizing is proposed. The experimental results with real data set verify the validity of the proposed method. Moreover, in comparison with the common LDA like methods, the proposed method shows better performance on both aspect mining and aspect features clustering.