交叉科学研究院研究生分享会

2025年11月9日（周日）；

上午 9 : 00 - 12 : 00

信息管理与工程学院102室

上海财经大学（第三教学楼西侧）

上海市杨浦区武东路100号

报告 ①

Dynamic Assortment Planning under Markov Chain Choice Model: Facing Impatient Customers

Abstract

In retailing and other selling scenarios, determining the set of products to offer in order to maximize expected revenue is an important decision problem. To address the assortment optimization problem, various choice models have been proposed to characterize customers’ choice probabilities among available alternatives. Among these models, the standard Markov Chain (MC) choice model is powerful, as it generalizes or closely approximates many utility-based models. However, it assumes that the transition probability is independent of the number of transitions one customer has already occurred, which is unsuitable for modeling impatient customers’ behaviors. To overcome this limitation, we generalize the MC model and propose the Markov Chain choice model with Decayed Patience (MCDP), by allowing the transition probabilities toward tangible products to decay over time, while increasing the no-purchase probability when preferred items are unavailable. The real-world empirical results illustrates that our MCDP model provides a more realistic characterization of customers’ purchasing behaviors than the MC model. Considering that customers can purchase any product offered in previous periods, we study a multi-period assortment planning problem under the MCDP model with a nested constraint on the assortment sequence. To address this problem, we develop two iterative algorithms: an externality-adjusted algorithm and a more efficient fixed-point-based algorithm. Both algorithms demonstrate that the problem is tractable in polynomial time and reveal structural properties on this problem.

报告 ②

Abstract

Session-based Recommendation (SR) has recently garnered significant attention. However, these methods often operate as black boxes, lacking the capability to provide clear explanations. Moreover, existing explanatory methods are limited in providing true explanations (i.e., critically influential inputs) for SR due to issues such as reliance on user information and the inability to capture sequential dynamics. To address these limitations, our work, Factual and Counterfactual Reasoning for Explainable Session-Based Recommendation with Reinforcement Learning (i.e., FCESR), aims to enhance the transparency of traditionally non-explainable SR models by focusing on the sufficiency and necessity of recommendation outcomes. Our approach improves SR model explainability by generating both factual and counterfactual explanations. Specifically, we formulate the explanation generation as a combinatorial optimization problem and employ reinforcement learning to effectively capture the sequential dynamics of session data while identifying the minimal set of items that drive recommendation outcomes. Furthermore, motivated by the potential of explanations to enhance recommendation accuracy, we feed factual and counterfactual explanations as high-quality positive and negative samples into a contrastive learning framework to fine-tune the SR models for more accurate recommendations. We conduct both qualitative analyses and empirical studies to validate the effectiveness of our proposed framework across multiple datasets and SR models, demonstrating substantial improvements in accuracy and explanation quality.

报告 ③

Abstract

For Mixed-Integer Linear Programming (MILP), the Local Branching (LB) heuristic is a well-established local search technique. However, its performance is highly sensitive to the neighborhood size—a parameter known to be instance-dependent. While recent learning-based methods aim to predict this numerical parameter, they often require extensive offline training data. This work introduces a novel approach that reframes neighborhood control in LB. Instead of predicting a size parameter, we learn a policy to select a subset of variables to which the LB constraint is applied. Our framework operates in two stages: first, we model the MILP instance as a graph and apply community detection to partition variables into structurally meaningful clusters, which serve as candidate neighborhoods. Second, a reinforcement learning (RL) agent dynamically selects the number of clusters to explore per iteration. Variables within chosen clusters are subjected to the LB constraint, while others are temporarily fixed. This results in an adaptive LB scheme where neighborhoods are defined by structural properties and dynamically scoped via RL—rather than by a single numerical parameter. Computational experiments demonstrate that our method automates neighborhood design without prior data collection. Evaluations across diverse MIP problems show that the proposed framework consistently outperforms state-of-the-art learning-based LB models and the open-source solver SCIP.