Peningkatan Profit Melalui Strategi Diskon Dinamis Menggunakan Reinforcement Learning pada Data Transaksi Ritel Daring

Abstract

Discount strategies play a critical role in e-commerce profitability, yet most platforms continue to rely on fixed discount rates that fail to account for heterogeneous customer behavior. This study proposes a dynamic discount optimization framework based on the Q-Learning algorithm, in which discount decisions are modeled as a Markov Decision Process (MDP) with a state space constructed from two customer behavioral indicators: transaction revenue and purchase frequency. Each customer transaction is mapped to one of six discrete states formed by combining three revenue categories (Low, Medium, High) and two frequency categories (Low, High), and the agent selects from five discount levels (0%, 5%, 10%, 15%, 20%) to maximize cumulative profit. A price elasticity mechanism is incorporated into the reward function to simulate demand responsiveness to price reductions, enabling the agent to learn non-trivial discount preferences. The framework is evaluated on the Online Retail II dataset comprising 397,885 cleaned transactions. Experimental results show that the proposed method achieves a total profit of 10,586,828.77, outperforming four of five fixed discount baselines with profit improvements of +18.80%, +11.16%, +5.60%, and +1.65% over the 0%, 5%, 10%, and 15% uniform discount scenarios, respectively. The learned policy consistently assigns 15% discounts to Low-frequency customers and 20% discounts to High-frequency customers, demonstrating that purchase frequency is the primary driver of optimal discount differentiation. These findings confirm that RL offers a more targeted and profit-competitive alternative to static promotional strategies