Intelligent Subsidy and Differential Pricing in DiDi Food: Economic Principles, System Architecture, and Experimental Insights

With the establishment of causal inference theory, intelligent marketing and subsidy solutions have achieved significant practical results in industry. Since the first half of 2020, DiDi Food has been experimenting with intelligent subsidy algorithms, achieving notable offline and online improvements.

Table of Contents

1. Economic principles of differential pricing

2. Differential pricing levers in DiDi Food

3. Definition of the subsidy problem

4. Architecture and workflow of intelligent subsidy in DiDi Food

5. Incremental estimation

6. Distribution strategies

7. Experimental results

8. Future plans

9. Q&A

1. Economic Principles of Differential Pricing

The essence of differential pricing is to create more value for consumers, users, and society, grounded in revenue‑management economics. An illustrative demand‑price curve shows how adding a higher price tier ($7) increases total profit (area A1+A2) and consumer surplus (area C1+C2), thereby delivering social value.

2. Differential Pricing Levers in DiDi Food

DiDi Food’s total price is expressed as P = Pi + Pd – Subsidyc , where Pi is the dish price, Pd the delivery fee, and Subsidyc the platform subsidy. This formulation provides three levers: merchant side (B), user side (C), and rider side (D). Direct user‑coupon subsidies are identified as the most effective differential pricing tool.

Constraints such as avoiding “big‑data price discrimination” limit the use of user‑level features in delivery‑fee pricing.

3. Definition of the Subsidy Problem

Unlike traditional recommendation, search, and advertising, the subsidy problem introduces a cost dimension. The core metric is ROI: the incremental revenue generated per unit of subsidy cost.

Two sub‑problems are identified: (a) estimating the uplift (increment) of an intervention, and (b) allocating interventions to users to achieve a global optimum.

4. Architecture and Workflow of Intelligent Subsidy

The system consists of offline and online modules. Core components are Incremental Prediction and Subsidy Allocation, which solve sub‑problems (a) and (b) respectively. Feature selection relies on a LightGBM model trained to approximate the predicted increment.

5. Incremental Estimation

Uplift modeling is used to predict the increase in purchase intent caused by a subsidy. The uplift is defined as P(Y=1|X,W=1) – P(Y=1|X,W=0) . Two families of models are discussed:

Meta‑Learners : Two‑Model, Single‑Model, X‑Learner, R‑Learner.

Tree‑Based Models : Distribution‑divergence trees, CTS (Contextual Treatment Selection) trees.

Each meta‑learner is described with its construction steps, advantages, and drawbacks. For example, the Two‑Model approach builds separate models for treatment and control groups and takes their difference as uplift.

Tree‑based uplift models directly split the feature space to maximize uplift, using criteria such as KL‑divergence, Euclidean distance, or Chi‑square divergence.

5.3 Model Evaluation

Because true uplift cannot be observed for the same user under both treatment and control, evaluation relies on indirect metrics such as Qini curves and AUUC (Area Under the Uplift Curve). The Qini curve partitions users into deciles based on predicted uplift and computes incremental gains; AUUC integrates the uplift curve and is less sensitive to imbalance between treatment and control group sizes.

6. Distribution Strategies

Two strategies are explored:

Greedy Allocation : Allocate coupons from low to high denomination, respecting budget constraints, and assign users with the highest predicted uplift first.

Integer Programming : Formulate subsidy allocation as an integer‑programming problem to maximize overall uplift under budget limits.

7. Experimental Results

AB tests in two cities compared three groups: a blank (no intervention), an operation‑controlled group, and the model‑driven strategy group. The model‑driven group achieved either the same order volume with significantly lower subsidy or the same subsidy with substantially higher order volume, demonstrating clear ROI improvements.

8. Future Plans

Four directions are outlined for the second half of 2020:

Push offline decision modules online, add real‑time features, and expand coupon scenarios.

Iterate on incremental prediction models, incorporating richer features and deep‑learning techniques.

Optimize integer‑programming solution speed and explore reinforcement‑learning approaches.

Model long‑term user value (LTV) to create a sustainable subsidy ecosystem.

9. Q&A

Q1 discusses the proper use of ROI as a business metric, emphasizing that ROI should be evaluated under constraints such as equal order volume or equal subsidy to ensure fair comparison.

Q2 explains the challenges of AUUC, noting its dependence on sample size and the difficulty of establishing a universal scale, while still being the most informative offline uplift metric.

The article concludes with contact information for further inquiries.