Heng-Qing Ye

In cross-border e-commerce, effective marketing resource allocation is crucial due to the complexity introduced by diverse product categories, regional differences, and competition among category managers. Current methods either overlook these constraints or fail to enforce them efficiently due to computational challenges. We propose a two-stage optimization framework that integrates predictive models with constrained optimization. In the first stage, predictive models estimate user purchase probabilities and determine upper bounds on product-specific sending volumes. In the second stage, the resource allocation problem is formulated as a large-scale integer programming model, which is then transformed into a minimum-cost flow problem to ensure computational efficiency while preserving solution optimality. Experiments on real-world data show that our framework significantly outperforms baseline strategies, achieving a 14.48% increase in order volume and revenue improvements ranging from 0.19% to 43.91%. The minimum-cost flow algorithm consistently outperforms the greedy approach, especially in large-scale instances. The proposed framework enables scalable and constraint-compliant marketing resource allocation in cross-border e-commerce. It not only improves sales performance but also ensures strict adherence to operational constraints, making it well-suited for large-scale commercial deployment.

Routing control is an important component in many engineering and management systems consisting of multiple and possibly heterogeneous servers. Imagine that upon the arrival of each job (or customer), a controller will evaluate the available (dynamic) state information and make a decision to dispatch the job to one of the servers. The state information can be queue length, arrival history, service history, and so on, depending on the nature of the application. How will the controller use the available state information to minimize the average waiting time an arriving job may experiences? In the paper, “Optimal Routing to Parallel Servers in Heavy Traffic,” Ye carries out the heavy traffic analysis to identify the routing policies that best use the available state information. For example, when there is no state information available for routing control, the best “blind” strategy is to dispatch the incoming jobs in a weighted round-robin fashion that exhibits certain form of the square-root rule. Although in the case that the job arrival history is available, the controller should use the information by closely chasing a kind of “arrival deviation,” which can reduce up to 50% of the waiting time compared with the best blind strategy. This study sheds new insights into the value of state information for routing control and provides new tools for engineering and service system design.

Depression is a heterogeneous and complex psychological syndrome with highly variable manifestations, which poses difficulties for treatment and prognosis. Depression patients are prone to developing various comorbidities, which stem from different pathophysiological mechanisms, remaining largely understudied. The current study focused on identifying comorbidity-specific phenotypes, and whether these clustered phenotypes are associated with different treatment patterns, clinical manifestations, physiological characteristics, and prognosis. We have conducted a 10-year retrospective observational cohort study using electronic medical records (EMR) for 11,818 patients diagnosed with depression and hospitalized at a large academic medical center in Chengdu, China. K-means clustering and visualization methods were performed to identify phenotypic categories. The association between phenotypic categories and clinical outcomes was evaluated using adjusted Cox proportional hazards model. We classified patients with depression into five stable phenotypic categories, including 15 statistically driven clusters in the discovery cohort (n = 9925) and the validation cohort (n = 1893), respectively. The categories include: (Category A) the lowest incidence of comorbidity, with prominent suicide, psychotic, and somatic symptoms (n = 3493/9925); (Category B) moderate comorbidity rate, with prominent anhedonia and anxious symptoms (n = 1795/9925); (Category C) the highest incidence of comorbidity of endocrine/metabolic and digestive system diseases (n = 1702/9925); (Category D) the highest incidence of comorbidity of neurological, mental and behavioral diseases (n = 881/9925); (Category E) other diseases comorbid with depression (n = 2054/9925). Patients in Category E had the lowest risk of psychiatric rehospitalization within 60-day follow-up, followed by Category C (HR, 1.57; 95% CI, 1.07–2.30), Category B (HR, 1.61; 95% CI, 1.10–2.40), Category A (HR, 1.82; 95% CI, 1.28–2.60), and Category D (HR, 2.38; 95% CI, 1.59–3.60) with P < 0.05, after adjustment for comorbidities, medications, and age. Regarding other longer observation windows (90-day, 180-day and 365-day), patients in Category D showed the highest rehospitalization risk all the time while there were notable shifts in rankings observed for Categories A, B and C over time. The results indicate that the higher the severity of mental illness in patients with five phenotypic categories, the greater the risk of rehospitalization. These phenotypes are associated with various pathways, including the cardiometabolic system, chronic inflammation, digestive system, neurological system, and mental and behavioral disorders. These pathways play a crucial role in connecting depression with other psychiatric and somatic diseases. The identified phenotypes exhibit notable distinctions in terms of comorbidity patterns, symptomology, biological characteristics, treatment approaches, and clinical outcomes.

In the precision marketing of a new product, it is a challenge to allocate limited resources to the target customer groups with different characteristics. We presented a framework using the distance-based algorithm, K-nearest neighbors, and support vector machine to capture customers’ preferences toward promotion channels. Additionally, online learning programming was combined with machine learning strategies to fit a dynamic environment, evaluating its performance through a parsimonious model of minimum regret. A resource optimization model was proposed using classification results as input. In particular, we collected data from an institution that provides financial credit products to capital-constrained small businesses. Our sample contained 525,919 customers who will be introduced to a new product. By simulating different scenarios between resources and demand, we showed an up to 22.42% increase in the number of expected borrowers when KNN was performed with an optimal resource allocation strategy. Our results also show that KNN is the most stable method to perform classification and that the distance-based algorithm has the most efficient adoption with online learning.

The goal of this paper is to illustrate the optimality of reflection control in three different settings, to bring out their connections and to contrast their distinctions. First, we study the control of a Brownian motion with a negative drift, so as to minimize a long-run average cost objective. Weprove the optimality of the reflection control, which prevents the Brownian motion from dropping below a certain level by cancelling out from time to time part of the negative drift; and we show that the optimal reflection level can be derived as the fixed point thatequates the long-run average cost to the holding cost. Second, we establish the asymptotic optimality of the reflection control when it is applied to a discrete production-inventory system driven by (delayed) renewal processes; and we do so via identifying the limiting regime of the system under diffusion scaling. Third, in the case of controlling a birth-death model, we establish the optimality of the reflection control directly via a linear programming-based approach. In all three cases, we allow an exponentially bounded holding cost function, which appears to be more general than what's allowed in prior studies. This general cost function reveals some previously unknown technical fine points on the optimality of the reflection control, and extends significantly its domain of applications.