How Reinforcement Learning Can Supercharge New Media Marketing Strategies

Background

New media marketing, which includes platforms such as WeChat, Weibo, Douyin, Kuaishou, and Xiaohongshu, has become popular due to its high efficiency, precise targeting, large reach, low cost, and broad audience compared with traditional media.

What Is New Media Marketing

For example, a cosmetics company may create copy, engage key opinion leaders (KOLs) on Weibo, publish soft articles on WeChat, and launch live‑stream sales to promote a new lipstick, hoping the target audience sees the content and purchases the product.

Traditional New Media Marketing Process

The process is typically divided into three steps: (1) define project requirements (budget, channels, KOL types, content formats, phases, timeline, expected outcomes); (2) determine channel‑allocation strategy (distribute the budget across channels, phases, and KOL tiers); (3) select KOL segment combinations that fit the strategy matrix while respecting budget constraints.

After these steps, KOLs publish the agreed content, and the campaign’s success is measured by metrics such as reads, likes, comments, and hot‑search rankings.

Current Limitations of New Media Marketing

Existing expert‑template or experience‑driven strategies are static and lag behind the rapidly changing market, making it difficult to adjust budgets and KOL selections in real time. A digital solution that models the workflow as project demand → strategy generation → segment combination → online deployment is needed.

Reinforcement Learning (RL) for New Media Marketing

What Is Reinforcement Learning

RL is an interactive learning paradigm where an agent interacts with an environment over a sequence of time steps, receiving states, taking actions, and obtaining rewards. The goal is to maximize the expected cumulative reward.

In the marketing scenario, the agent receives business requirements (budget, channels, etc.) and outputs a budget‑allocation action for each cell of the strategy matrix; the environment predicts the resulting KOL combination performance using historical data and deep models.

RL Frameworks

OpenAI‑baselines (https://github.com/openai/baselines)

Google‑Dopamine (https://github.com/google/dopamine)

Baidu‑PARL (https://github.com/PaddlePaddle/PARL)

DeepMind‑OpenSpiel (https://arxiv.org/abs/1908.09453)

Tianshou (https://github.com/thu-ml/tianshou)

Technical Solution

Problem Modeling

The strategy system is treated as an agent, while a KOL performance prediction model serves as the environment. Each interaction corresponds to processing one cell of the strategy matrix, where the agent decides the budget proportion and the environment returns the estimated effect.

Algorithm Selection

Because the action space is low‑dimensional and continuous, Deep Deterministic Policy Gradient (DDPG) is chosen. DDPG outputs deterministic actions and adds exploration noise.

Action Design

The action a ∈ [0,1] represents the proportion of the remaining budget to spend in the current state.

Reward Design

Maximize a composite score that reflects KOL follower quality, interaction rate, price, predicted reach, and content type.

Enforce hierarchical budget ratios (head > mid > tail KOL tiers).

Penalize abnormal budget ratios.

Assign zero score when no feasible KOL combination can be formed.

Model Training

Training aims for convergence and alignment with business intent. Key aspects include hyper‑parameter tuning, monitoring actor and critic losses, and adaptive adjustment of RL hyper‑parameters throughout training.

Experimental Results

Average propagation effect increased by over 35%.

Average cost reduced by over 8%.

For a 100 k budget, the RL solution achieved the same effect with only 92 k, or a 35% higher effect with the original budget.

Conclusion and Outlook

The study demonstrates that RL can effectively optimize new media marketing budget allocation by modeling the problem as an MDP, designing a business‑aligned reward function, and employing DDPG. Future work includes model transfer across product categories, improving confidence of the performance‑prediction model, and researching adaptive hyper‑parameter strategies.