Online Internet platforms require sophisticated marketing strategies to optimize user retention and platform revenue — a classical resource allocation problem. Traditional solutions adopt a two-stage pipeline: machine learning (ML) for predicting individual treatment effects to marketing actions, followed by operations research (OR) optimization for decision-making. This paradigm presents two fundamental technical challenges. First, the prediction-decision misalignment: Conventional ML methods focus solely on prediction accuracy without considering downstream optimization objectives, leading to improved predictive metrics that fail to translate to better decisions. Second, the bias-variance dilemma: Observational data suffers from multiple biases (e.g., selection bias, position bias), while experimental data (e.g., randomized controlled trials), though unbiased, is typically scarce and costly --- resulting in high-variance estimates. We propose **Bi**-level **D**ecision-**F**ocused **C**ausal **L**earning (**Bi-DFCL**) that systematically addresses these challenges. First, we develop an unbiased estimator of OR decision quality using experimental data, which guides ML model training through surrogate loss functions that bridge discrete optimization gradients. Second, we establish a bi-level optimization framework that jointly leverages observational and experimental data, solved via implicit differentiation. This novel formulation enables our unbiased OR estimator to correct learning directions from biased observational data, achieving optimal bias-variance tradeoff. Extensive evaluations on public benchmarks, industrial marketing datasets, and large-scale online A/B tests demonstrate the effectiveness of Bi-DFCL, showing statistically significant improvements over state-of-the-art. Currently, Bi-DFCL has been deployed across several marketing scenarios at Meituan, one of the largest online food delivery platforms in the world.

In this paper, we propose TCP-FNC (TCP with fast network coding), which is designed to reduce the decoding delay for TCP with network coding. TCP-FNC retains the framework of TCP/NC and includes two schemes to meet the demand of online network coding. First, we develop a feedback based scheme named FCWL, which can efficiently reduce the waiting delay and work well with both RTT-based and loss-based TCP variants. Second, we propose an optimized progressive decoding algorithm named EFU for Gaussian-Jordan elimination, which can further reduce the computation delay from O(n2) to O(n). Evaluation results indicate that our TCP-FNC can achieve shorter decoding delay than TCP/NC and VON without trading off its goodput performance. We verify that TCP-FNC can reduce the decoding delay by 33% on average.

TCP Incast refers to the phenomenon of goodput collapse when multiple synchronized servers send data to the same client in parallel. In this paper, we propose a novel application-level scheduling approach named OSDT (Optimal Staggering Data Transfers) for TCP Incast problem. OSDT limits the number of concurrent TCP flows as well as servers' sending rate to optimal values so that the utilization of link capacity can be maximized without any packet losses. To achieve this, we build an optimization model with the usage of network and application information. Based on this model we can get the optimal values for the parameters in OSDT. Simulation results indicate that OSDT can achieve the highest goodput among all existing application-level scheduling approaches in a wide range of network and application parameters, and its performance is stable. So OSDT can be seen as an effective and scalable solution for TCP Incast problem.

TCP suffers from low throughput and high latency because of its expensive timeout based loss recovery mechanism in data center networks (DCNs). In this paper, we propose TCP with Adaptive Complementary Coding (TCP-ACC) to effectively address these problems. Without revising existing TCP congestion control, we first design a light-weight complementary coding scheme to avoid TCP timeout which will result in higher throughput and lower latency. In our scheme, the redundancy setting is adaptive to the real-time packet loss rate. Then we introduce Lyapunov optimization framework to find the optimal number of redundant coding packets for TCP-ACC, and we also prove that TCP-ACC can reduce the flow timeout probability close to that of the optimal complementary coding solution. Extensive NS2 simulations show that, compared with other three solutions for TCP's problems in DCNs, TCP-ACC can reduce the flow completion time by 45% and improve the flow throughput by 40% on average.

In this paper, we propose CGM-PS, a partial-decentralized, un-starving, work-conservative, and preemptive coflow scheduling scheme to shorten the Coflow Completion Time (CCT) for TCP flows in data center networks (DCNs). In CGMPS, we propose both inter- and intra- coflow scheduling policies. In inter-coflow scheduling, we present P-SEBF, which adopts a connected-graph model based novel concept Partialcoflow, to achieve approximate SEBF scheduling in a partial-decentralized manner. In intra-coflow scheduling, we present FP-MDFS to give flow-level priorities and appropriate rates to TCP flows for finishing the coflows as quick as possible without wasting network capacities in a decentralized manner. Trace-based simulation results show that, among existing coflow scheduling schemes, CGM-PS can achieve the minimal CCTs both on average and in the 90th percentile. In brief, CGM-PS only brings about similar scheduling overhead with the decentralized schemes, while it can achieve the CCT performance even better than the near optimal centralized scheme.

Due to the special topologies and communication pattern, in today's data center networks it is common that a large set of TCP flows and a small set of TCP flows get into different ingress ports of a switch and compete for a same egress port. However, in this case the throughput share of flows in the two sets will not be fair even though all flows have the same RTT. In this paper, we study this problem and find that TCP's fairness in data center networks is related with not only the network capacity but also the number of flows in the two sets. We propose a mathematical model of the average throughput ratio of the large set of flows to the small set of flows. This model can reveal the variation of TCP's fairness along with the change of network parameters (including buffer size, bandwidth, and propagation delay) as well as the number of flows in the two sets. We validate our model by comparing its numerical results with simulation results, finding that they match well.