Decision-focused learning (DFL) offers an end-to-end approach to the predict-then-optimize (PO) framework by training predictive models directly on decision loss (DL), enhancing decision-making performance within PO contexts. However, the implementation of DFL poses distinct challenges. Primarily, DL can result in deviation from the physical significance of the predictions under limited data. Additionally, some predictive models are non-differentiable or black-box, which cannot be adjusted using gradient-based methods. To tackle the above challenges, we propose a novel framework, Decision-Focused Fine-tuning (DFF), which embeds the DFL module into the PO pipeline via a novel bias correction module. DFF is formulated as a constrained optimization problem that maintains the proximity of the DL-enhanced model to the original predictive model within a defined trust region. We theoretically prove that DFF strictly confines prediction bias within a predetermined upper bound, even with limited datasets, thereby substantially reducing prediction shifts caused by DL under limited data. Furthermore, the bias correction module can be integrated into diverse predictive models, enhancing adaptability to a broad range of PO tasks. Extensive evaluations on synthetic and real-world datasets, including network flow, portfolio optimization, and resource allocation problems with different predictive models, demonstrate that DFF not only improves decision performance but also adheres to fine-tuning constraints, showcasing robust adaptability across various scenarios.

Estimating long-term causal effects by combining long-term observational and short-term experimental data is a crucial but challenging problem in many real-world scenarios. In existing methods, several ideal assumptions, e.g. latent unconfoundedness assumption or additive equi-confounding bias assumption, are proposed to address the latent confounder problem raised by the observational data. However, in real-world applications, these assumptions are typically violated which limits their practical effectiveness. In this paper, we tackle the problem of estimating the long-term individual causal effects without the aforementioned assumptions. Specifically, we propose to utilize the natural heterogeneity of data, such as data from multiple sources, to identify latent confounders, thereby significantly avoiding reliance on idealized assumptions. Practically, we devise a latent representation learning-based estimator of long-term causal effects. Theoretically, we establish the identifiability of latent confounders, with which we further achieve long-term effect identification. Extensive experimental studies, conducted on multiple synthetic and semi-synthetic datasets, demonstrate the effectiveness of our proposed method.

Estimating the individuals' potential response to varying treatment doses is crucial for decision-making in areas such as precision medicine and management science. Most recent studies predict counterfactual outcomes by learning a covariate representation that is independent of the treatment variable. However, such independence constraints neglect much of the covariate information that is useful for counterfactual prediction, especially when the treatment variables are continuous. To tackle the above issue, in this paper, we first theoretically demonstrate the importance of the balancing and prognostic representations for unbiased estimation of the heterogeneous dose-response curves, that is, the learned representations are constrained to satisfy the conditional independence between the covariates and both of the treatment variables and the potential responses. Based on this, we propose a novel Contrastive balancing Representation learning Network using a partial distance measure, called CRNet, for estimating the heterogeneous dose-response curves without losing the continuity of treatments. Extensive experiments are conducted on synthetic and real-world datasets demonstrating that our proposal significantly outperforms previous methods.

Achieving optimal order dispatching has been a long-standing challenge for online ride-hailing platforms. Early methods would make shortsighted matchings as they only consider order prices alone as the edge weights in the driver-order bipartite graph, thus harming the platform's revenue. To address this problem, recent works evaluate the value of the order's destination region to be the long-term income a driver could obtain in average in such region and incorporate it into the order's edge weight to influence the matching results. However, they often result in insufficient driver supplies in many regions, as the values evaluated in different regions vary greatly, mainly because the impact of one region's value on the future number of drivers and revenue in other regions is overlooked. This paper models such impact within a cooperative Markov game, which involves each value's impact over the platform's revenue with the goal to find the optimal region values for revenue maximization. To solve this game, our work proposes a novelgoal-reaching collaboration (GRC) algorithm that realizes credit assignment from a novel goal-reaching perspective, addressing the difficulty for accurate credit assignment with large-scale agents of previous methods and resolving the conflict between credit assignment and offline reinforcement learning. Specifically, during training, GRC predicts the city's future state through an environment model and utilizes a scoring model to rate the predicted states to judge their levels of profitability, where high-scoring states are regarded as the goal states. Then, the policies in the game are updated to promote the city to stay in the goal states for as long as possible. To evaluate GRC, we deploy a baseline policy online in several cities for three weeks to collect real-world dataset. Training and testing results on the collected dataset indicate that our GRC consistently outperforms the baselines in different cities and peak periods.

Recently, network embedding has been successfully used in recommendation systems. Researchers have made efforts to utilize additional auxiliary information (e.g., social relations of users) to improve performance. However, such auxiliary information lacks compatibility for all recommendation scenarios, thus it is difficult to apply in some industrial scenarios where generality is required. Moreover, the heterogeneous nature between users and items aggravates the difficulty in network information fusion. Many works tried to transform user-item heterogeneous network to two homogeneous graphs (i.e., user-user and item-item), and then fuse information separately. This may limit the representation power of learned embedding due to ignoring the adjacent relationship in the original graph. In addition, the sparsity of user-item interactions is an urgent problem need to be solved. To solve the above problems, we propose a universal and effective framework named Gemini, which only relies on the common interaction logs, avoiding the dependence on auxiliary information and ensuring a better generality. For the purpose of keeping original adjacent relationship, Gemini transforms the original user-item heterogeneous graph into two semi homogeneous graphs from the perspective of users and items respectively. The transformed graphs consist of two types of nodes: network nodes coming from homogeneous nodes and attribute nodes coming from heterogeneous node. Then, the node representation is learned in a homogeneous way, with considering edge embedding at the same time. Simultaneously, the interaction sparsity problem is solved to some extent as the transformed graphs contain the original second-order neighbors. For training efficiently, we also propose an iterative training algorithm to reduce computational complexity. Experimental results on the five datasets and online A/B tests in recommendations of DiDiChuXing show that Gemini outperforms state-of-the-art algorithms.