Large Language Models (LLMs) exhibit strong potential in mathematical reasoning, yet their effectiveness is often limited by a shortage of high-quality queries. This limitation necessitates scaling up computational responses through self-generated data, yet current methods struggle due to spurious correlated data caused by ineffective exploration across all reasoning stages. To address such challenge, we introduce **MARGE**: Improving **Ma**th **R**easoning with **G**uided **E**xploration, a novel method that enhances mathematical reasoning through hit-guided exploration. MARGE systematically explores intermediate reasoning states derived from self-generated solutions, enabling adequate exploration and improved credit assignment throughout the reasoning process. Notably, MARGE improves both single-shot accuracy and exploration diversity, mitigating a common trade-off in alignment methods. These results demonstrate MARGE's effectiveness in enhancing mathematical reasoning capabilities and unlocking the potential of scaling self-generated training data.

Click-Through Rate (CTR) prediction serves as a fundamental component in online advertising. A common practice is to train a CTR model on advertisement (ad) impressions with user feedback. Since ad impressions are purposely selected by the model itself, their distribution differs from the inference distribution and thus exhibits sample selection bias (SSB) that affects model performance. Existing studies on SSB mainly employ sample re-weighting techniques which suffer from high variance and poor model calibration. Another line of work relies on costly uniform data that is inadequate to train industrial models. Thus mitigating SSB in industrial models with a uniform-data-free framework is worth exploring. Fortunately, many platforms display mixed results of organic items (i.e., recommendations) and sponsored items (i.e., ads) to users, where impressions of ads and recommendations are selected by different systems but share the same user decision rationales. Based on the above characteristics, we propose to leverage recommendations samples as a free lunch to mitigate SSB for ads CTR model (Rec4Ad). After elaborating data augmentation, Rec4Ad learns disentangled representations with alignment and decorrelation modules for enhancement. When deployed in Taobao display advertising system, Rec4Ad achieves substantial gains in key business metrics, with a lift of up to +6.6% CTR and +2.9% RPM.

Cascading architecture has been widely adopted in large-scale advertising systems to balance efficiency and effectiveness. In this architecture, the pre-ranking model is expected to be a lightweight approximation of the ranking model, which handles more candidates with strict latency requirements. Due to the gap in model capacity, the pre-ranking and ranking models usually generate inconsistent ranked results, thus hurting the overall system effectiveness. The paradigm of score alignment is proposed to regularize their raw scores to be consistent. However, it suffers from inevitable alignment errors and error amplification by bids when applied in online advertising. To this end, we introduce a consistency-oriented pre-ranking framework for online advertising, which employs a chunk-based sampling module and a plug-and-play rank alignment module to explicitly optimize consistency of ECPM-ranked results. A ΔNDCG-based weighting mechanism is adopted to better distinguish the importance of inter-chunk samples in optimization. Both online and offline experiments have validated the superiority of our framework. When deployed in Taobao display advertising system, it achieves an improvement of up to +12.3% CTR and +5.6% RPM.

Despite the development of ranking optimization techniques, pointwise loss remains the dominating approach for click-through rate prediction. It can be attributed to the calibration ability of the pointwise loss since the prediction can be viewed as the click probability. In practice, a CTR prediction model is also commonly assessed with the ranking ability. To optimize the ranking ability, ranking loss (e.g., pairwise or listwise loss) can be adopted as they usually achieve better rankings than pointwise loss. Previous studies have experimented with a direct combination of the two losses to obtain the benefit from both losses and observed an improved performance. However, previous studies break the meaning of output logit as the click-through rate, which may lead to sub-optimal solutions. To address this issue, we propose an approach that can Jointly optimize the Ranking and Calibration abilities (JRC for short). JRC improves the ranking ability by contrasting the logit value for the sample with different labels and constrains the predicted probability to be a function of the logit subtraction. We further show that JRC consolidates the interpretation of logits, where the logits model the joint distribution. With such an interpretation, we prove that JRC approximately optimizes the contextualized hybrid discriminative-generative objective. Experiments on public and industrial datasets and online A/B testing show that our approach improves both ranking and calibration abilities. Since May 2022, JRC has been deployed on the display advertising platform of Alibaba and has obtained significant performance improvements.

Being able to discover diverse useful skills without external reward functions is beneficial in reinforcement learning research. Previous unsupervised skill discovery approaches mainly train different skills in parallel. Although impressive results have been provided, we found that parallel training procedure can sometimes block exploration when the state visited by different skills overlap, which leads to poor state coverage and restricts the diversity of learned skills. In this paper, we take a deeper look into this phenomenon and propose a novel framework to address this issue, which we call Recurrent Skill Training (ReST). Instead of training all the skills in parallel, ReST trains different skills one after another recurrently, along with a state coverage based intrinsic reward. We conduct experiments on a number of challenging 2D navigation environments and robotic locomotion environments. Evaluation results show that our proposed approach outperforms previous parallel training approaches in terms of state coverage and skill diversity. Videos of the discovered skills are available at https://sites.google.com/view/neurips22-rest.

Learning user representation is a critical task for recommendation systems as it can encode user preference for personalized services. User representation is generally learned from behavior data, such as clicking interactions and review comments. However, for less popular domains, the behavior data is insufficient to learn precise user representations. To deal with this problem, a natural thought is to leverage content-rich domains to complement user representations. Inspired by the recent success of BERT in NLP, we propose a novel pre-training and fine-tuning based approach U-BERT. Different from typical BERT applications, U-BERT is customized for recommendation and utilizes different frameworks in pre-training and fine-tuning. In pre-training, U-BERT focuses on content-rich domains and introduces a user encoder and a review encoder to model users' behaviors. Two pre-training strategies are proposed to learn the general user representations; In fine-tuning, U-BERT focuses on the target content-insufficient domains. In addition to the user and review encoders inherited from the pre-training stage, U-BERT further introduces an item encoder to model item representations. Besides, a review co-matching layer is proposed to capture more semantic interactions between the reviews of the user and item. Finally, U-BERT combines user representations, item representations and review interaction information to improve recommendation performance. Experiments on six benchmark datasets from different domains demonstrate the state-of-the-art performance of U-BERT.

We study two user demands that are important during the exploitation of explanations in practice: 1) understanding the overall model behavior faithfully with limited cognitive load and 2) predicting the model behavior accurately on unseen instances. We illustrate that the two user demands correspond to two major sub-processes in the human cognitive process and propose a unified framework to fulfill them simultaneously. Given a local explanation method, our framework jointly 1) learns a limited number of groupwise explanations that interpret the model behavior on most instances with high fidelity and 2) specifies the region where each explanation applies. Experiments on six datasets demonstrate the effectiveness of our method.

Existing review-based recommendation models mainly learn long- term user and item representations from a set of reviews. Due to the ignorance of rich side information of reviews, these models suffer from two drawbacks: 1) they fail to capture short-term changes of user preferences and item features reflected in reviews and 2) they cannot accurately model high-order user-item collaborative signals from reviews. To overcome these limitations, we propose a multi-view approach named Set-Sequence-Graph (SSG), to augment existing single-view (i.e., view of set) methods by introducing two additional views of exploiting reviews: sequence and graph. In particular, with reviews organized in forms of set, sequence, and graph respectively, we design a three-way encoder architecture that jointly captures long-term (set), short-term (sequence), and collaborative (graph) features of users and items for recommendation. For the sequence encoder, we propose a short-term priority attention network that explicitly takes the order and personalized time intervals of reviews into consideration. For the graph encoder, we design a novel review-aware graph attention network to model high-order multi-aspect relations in the user-item graph. To combat the potential redundancy in captured features, our fusion module employs a cross-view decorrelation mechanism to encourage diverse representations from multiple views for integration. Experiments on public datasets demonstrate that SSG significantly outperforms state-of-the-art methods.

Recommender systems have been playing an increasingly important role in our daily life due to the explosive growth of information. Accuracy and explainability are two core aspects when we evaluate a recommendation model and have become one of the fundamental trade-offs in machine learning. In this paper, we propose to alleviate the trade-off between accuracy and explainability by developing an explainable deep model that combines the advantages of deep learning-based models and existing explainable methods. The basic idea is to build an initial network based on an explainable deep hierarchy (e.g., Microsoft Concept Graph) and improve the model accuracy by optimizing key variables in the hierarchy (e.g., node importance and relevance). To ensure accurate rating prediction, we propose an attentive multi-view learning framework. The framework enables us to handle sparse and noisy data by co-regularizing among different feature levels and combining predictions attentively. To mine readable explanations from the hierarchy, we formulate personalized explanation generation as a constrained tree node selection problem and propose a dynamic programming algorithm to solve it. Experimental results show that our model outperforms state-of-the-art methods in terms of both accuracy and explainability.

In modern cities, complaining has become an important way for citizens to report emerging urban issues to governments for quick response. For ease of retrieval and handling, government officials usually organize citizen complaints by manually assigning tags to them, which is inefficient and cannot always guarantee the quality of assigned tags. This work attempts to solve this problem by recommending tags for citizen complaints. Although there exist many studies on tag recommendation for textual content, few of them consider two characteristics of citizen complaints, i.e., the spatio-temporal correlations and the taxonomy of candidate tags. In this paper, we propose a novel Spatio-Temporal Taxonomy-Aware Recommendation model (STAR), to recommend tags for citizen complaints by jointly incorporating spatio-temporal information of complaints and the taxonomy of candidate tags. Specifically, STAR first exploits two parallel channels to learn representations for textual and spatio-temporal information. To effectively leverage the taxonomy of tags, we design chained neural networks that gradually refine the representations and perform hierarchical recommendation under a novel taxonomy constraint. A fusion module is further proposed to adaptively integrate contributions of textual and spatio-temporal information in a tag-specific manner. We conduct extensive experiments on a real-world dataset and demonstrate that STAR significantly performs better than state-of-the-art methods. The effectiveness of key components in our model is also verified through ablation studies.

Diagnosis prediction, which aims to predict future health information of patients from historical electronic health records (EHRs), is a core research task in personalized healthcare. Although some RNN-based methods have been proposed to model sequential EHR data, these methods have two major issues. First, they cannot capture fine-grained progression patterns of patient health conditions. Second, they do not consider the mutual effect between important context (e.g., patient demographics) and historical diagnosis. To tackle these challenges, we propose a model called Co-Attention Memory networks for diagnosis Prediction (CAMP), which tightly integrates historical records, fine-grained patient conditions, and demographics with a three-way interaction architecture built on co-attention. Our model augments RNNs with a memory network to enrich the representation capacity. The memory network enables analysis of fine-grained patient conditions by explicitly incorporating a taxonomy of diseases into an array of memory slots. We instantiate the READ/WRITE operations of the memory network so that the memory cooperates effectively with the patient demographics through co-attention mechanism. Experiments on real-world datasets demonstrate that CAMP consistently performs better than state-of-the-art methods.

Drug-drug interactions (DDIs) are a major cause of preventable hospitalizations and deaths. Recently, researchers in the AI community try to improve DDI prediction in two directions, incorporating multiple drug features to better model the pharmacodynamics and adopting multi-task learning to exploit associations among DDI types. However, these two directions are challenging to reconcile due to the sparse nature of the DDI labels which inflates the risk of overfitting of multi-task learning models when incorporating multiple drug features. In this paper, we propose a multi-task semi-supervised learning framework MLRDA for DDI prediction. MLRDA effectively exploits information that is beneficial for DDI prediction in unlabeled drug data by leveraging a novel unsupervised disentangling loss CuXCov. The CuXCov loss cooperates with the classification loss to disentangle the DDI prediction relevant part from the irrelevant part in a representation learnt by an autoencoder, which helps to ease the difficulty in mining useful information for DDI prediction in both labeled and unlabeled drug data. Moreover, MLRDA adopts a multi-task learning framework to exploit associations among DDI types. Experimental results on real-world datasets demonstrate that MLRDA significantly outperforms state-of-the-art DDI prediction methods by up to 10.3% in AUPR.