How a Semi‑Supervised Unified Framework Boosts E‑commerce Query Intent Classification

Query intent prediction—identifying categories, brands, and product terms from user searches—is critical for e‑commerce but faces challenges such as short queries lacking context, insufficient prior information, and a reliance on post‑click labels that creates a Matthew‑effect feedback loop. Existing industrial methods depend on user click behavior to generate training data, leading to poor coverage of long‑tail categories and fragmented sub‑tasks.

Motivation and Challenges

Key problems include:

Insufficient prior information: short queries, missing context, and inability to model inter‑category relationships.

Matthew‑effect loop: dominant products receive most clicks, starving long‑tail categories of data.

Lack of a unified framework for various sub‑tasks, resulting in low algorithmic efficiency.

Proposed Solution: Semi‑Supervised Extensible Unified Framework (SSUF)

SSUF consists of three plug‑and‑play enhancement modules:

Label Enhancement : Uses label text, descriptions, and side information (including LLM‑generated world knowledge) to create semantic label embeddings via BERT, breaking dependence on posterior click labels.

Knowledge Enhancement : Incorporates factual knowledge from large language models and user‑behavior knowledge to enrich query representations. Queries and labels share a common text encoder, and a knowledge encoder produces embeddings \(\mathbf{K}\) that are fused with query vectors through an attention mechanism.

Structural Enhancement : Builds multiple label relation graphs (semantic similarity, co‑occurrence, hierarchical) and fuses them using graph convolutional networks (GCN) to obtain robust label embeddings that propagate gradients from head to tail categories.

The unified framework allows each module to be added or removed per sub‑task, supporting multi‑task joint training.

Model Details

Label encoding concatenates the label name \(n\) with side information \(m\) and feeds the character sequence into BERT, producing a CLS vector \(\mathbf{C}_j\). Queries are encoded similarly to obtain \(\mathbf{Q}_i\). Knowledge embeddings are generated from LLM‑derived descriptions and user behavior data, then fused with queries via attention: Attention(Q', K) → q_i' Similarity scores between fused query vectors and label embeddings yield semi‑supervised labels, filtered by a threshold \(\tau\). Gradients from the semi‑supervised branch are blocked to avoid collapse.

Structural graphs are constructed from co‑occurrence frequencies, cosine similarity of label semantics, and hierarchical parent‑child relations. These adjacency matrices are merged, normalized, and processed by a GCN to produce final label node representations \(\mathbf{H}\). Leaf‑label embeddings \(\mathbf{H}_l\) are extracted and projected onto the label space for prediction: ŷ_i = softmax(H_l · q_i + b) The loss combines binary cross‑entropy on click‑derived multi‑hot labels \(\mathbf{y}^{click}_i\) and the semi‑supervised labels.

Experiments

Offline evaluations compare SSUF against academic multi‑label classifiers (XML‑CNN, LEAM, LSAN) and industrial query intent models (DPHA, MMAN, HCL4QC, SMGCN, HQC). Ablation studies remove each enhancement module (w/o KE, w/o LE, w/o SE, etc.) showing that every component contributes to performance; removing all three drops F1 by ~8%.

Online A/B tests deploy SSUF with a four‑layer BERT encoder and cached GCN label vectors, incurring no extra latency. Results demonstrate significant lifts in impressions and clicks, confirming the framework’s effectiveness and scalability.

Conclusion

SSUF provides a modular, extensible solution for e‑commerce query intent classification, addressing data sparsity, label bias, and fragmented sub‑tasks while delivering strong offline and online gains.