CIKM 2026: One Click Isn't Enough — Intent Decomposition Boosts Orders 6.9%, GMV 7.7%

Research Background and Problem

In e-commerce product detail pages, relevant recommendation uses the clicked item ("trigger item") as context to suggest subsequent items. However, a single click does not fully reveal user intent. Existing methods face three challenges: (1) static trigger-item representations cannot capture intent shifts — the same item may signal different needs for different users, and a user's focus may change while browsing; business data shows CTR and CVR of same-category candidates drop more sharply than cross-category ones as exposure position moves down, indicating a shift from targeted search to broader exploration. (2) Preferences across attributes (category, brand, shop, price) are entangled in a single coarse-grained vector, masking fine-grained differences. (3) Irrelevant long-term historical behavior interferes with current intent judgment because user history spans multiple interests.

Proposed Method: CAIM

CAIM (Context-Aware User Intent Modeling) organizes modeling into three interconnected modules: dynamic trigger-item representation construction, long-term interest retrieval, and attribute-level sub-intent extraction.

2.1 Intention Enhancement Module (IEM)

IEM fuses user profile, static trigger-item representation, and recent click sequence. The short-term sequence is average-pooled, then concatenated with mapped user profile and trigger-item features. A SENet learns channel-wise weights to re-weight the fused input, producing a context-aware dynamic trigger-item representation. Formally: let f_map denote feature mapping, pool average pooling, concat concatenation. SENet computes channel weights via squeeze-excitation: z = σ(W_2 · ReLU(W_1 · x)), then re-weights: x' = z ⊙ x. This allows the same trigger item to obtain different intent representations under different users and browsing contexts.

2.2 Intention Retrieval Module (IRM)

IRM uses the dynamic trigger-item representation as query, and long-term behavior sequence representations as keys and values, applying multi-head attention to aggregate historically relevant information. The core computation: MultiHead(Q, K, V) = Concat(head_1, ..., head_h)W^O where head_i = Attention(QW_i^Q, KW_i^K, VW_i^V) and Attention(Q, K, V) = softmax(QK^T / √d_k)V. The model weights historical representations by their match with current intent; e.g., when browsing down jackets, past apparel brand preferences receive higher weight while unrelated appliance clicks are down-weighted.

2.3 Sub-intention Extraction Module (SIEM)

SIEM models sub-intents on four attribute dimensions — category, brand, shop, price — plus a general interest representation. It employs an Intention Weighted Fusion Network (IWFN) inspired by PLE's shared/expert design: a gating network fuses outputs of shared experts (learning reusable cross-granularity information) and task-specific experts (capturing granularity-specific preferences). For each granularity g, input x_g produces output o_g = Σ_k g_k · E_k(x_g) where g_k are gate weights and E_k expert networks. Subsequently, an Intention-level Token-wise Feed-forward Network (ITF²N) applies sparse expert routing: for each sub-intent token, a gate selects top- k experts, computes y = x + Σ_{i∈S} g_i · FFN_i(x), preserving input via residual connection and controlling compute by activating only a subset of experts.

2.4 Decision Layer

The final prediction concatenates user profile, dynamic trigger-item representation, candidate item representation, and all refined sub-intent representations, feeding them into an MLP with sigmoid output to predict click probability.

Experimental Results

3.1 Offline Experiments

Experiments on an industrial dataset and the public Alimama advertising dataset (where the most recent click before impression serves as trigger item) compare against sequential recommendation models, trigger-item-aware models, and large-scale unified architectures. CAIM achieves AUC of 0.6797 (industrial) and 0.5898 (Alimama), improving over the strongest baselines by 0.20 pt and 0.21 pt respectively (1 pt = 0.01 absolute AUC difference).

3.2 Ablation Study

On the industrial dataset, removing IEM, IRM, or SIEM individually from full CAIM reduces AUC by 0.56 pt , 0.47 pt , and 1.04 pt respectively. All three modules contribute positively; SIEM removal causes the largest drop, indicating its critical role in this setting.

3.3 Online A/B Test

A 14-day online A/B test in the real relevant-recommendation scenario, using the deployed sequential recommendation model as control, shows full CAIM improves order volume by 6.9% and GMV by 7.7% relative to control. Module combination experiments confirm that the full three-module CAIM outperforms any two-module subset on both metrics, demonstrating complementary effects consistent with offline ablation.