How AI Optimizes E‑Commerce ‘Bundle‑Buy’ with Graph Embedding & Knapsack

Overall Solution

The goal of the "bundle‑buy" project is to recommend additional items that complement a user's existing cart under the same coupon, increasing transaction value, meeting price thresholds, and improving shopping experience.

Scenario‑Based Recommendation Method

We first identify the scenario concept of the added item (e.g., hiking shoes imply a hiking‑gear scenario) using query analysis, mutual information, and left/right entropy to discover new phrases and determine their broad intent.

Based on the identified scenario, we retrieve the most clicked categories and combine long‑term and short‑term user preferences to generate a list of candidate categories.

Examples: after adding a facial cleanser, we recommend conditioner, body wash, shampoo, and soap; after adding a tea set, we recommend tea plates, cups, towels, jars, and accessories.

Price‑Constraint Grouped Knapsack Algorithm

We model the selection of one item per candidate category as a Multiple‑Choice Knapsack Problem (MCKP), where each category is a group and each item has a price (weight) and purchase probability (value).

The objective is to maximize expected revenue while ensuring the total price exceeds the coupon threshold, with a relaxation variable \(\delta\) to adapt to different user purchasing power.

Formally, for each group \(i\) we select at most one item \(j\) with price \(p_{ij}\) and value \(v_{ij}\), subject to \(\sum_i p_{ij} \ge P + \delta\) where \(P\) is the coupon amount.

We solve the MCKP using an O(n) algorithm (Dyer & Zemel) and set the maximum number of groups \(M\) based on the first‑page limit (e.g., \(M \le 4\)).

Experience Assurance Mechanisms

Virtual category dispersion to limit the number of items from the same top‑level category on a page.

Exposure filtering to avoid over‑showing popular items.

User fatigue model to balance novelty and relevance.

Virtual categories are constructed by clustering real categories into a hierarchy that satisfies both coarse and fine granularity constraints.

Algorithm Effects

Deploying deep‑learning models, the grouped‑knapsack optimizer, and cross‑category graph embedding increased overall GMV and conversion rates compared to the previous year.

Online A/B tests showed significant lifts in exposure value and conversion metrics.

Conclusion and Outlook

We demonstrated that combinatorial optimization and scenario‑aware recommendations effectively improve bundle‑buy performance under price constraints. Future work includes dynamic determination of group count \(M\) and relaxation factor \(\delta\) via reinforcement learning, and more granular price‑preference modeling.