A New Dual‑Tower Multi‑Objective Recall Framework (CSMF) Boosts Efficiency Without Extra Parameters

Background

Large‑scale e‑commerce recommendation platforms must retrieve a tiny set of highly relevant items from billions of candidates under strict latency constraints. Embedding‑Based Retrieval (EBR) uses a dual‑tower architecture where user and item vectors are pre‑computed and matched via inner product. Business objectives such as relevance, exposure, click, and conversion create a multi‑objective retrieval problem.

Existing multi‑objective EBR approaches fall into two categories. Multi‑model solutions train a separate model per objective, which cannot capture inter‑objective relationships and suffers from data sparsity for downstream tasks. Single‑model solutions share parameters across objectives, either by merging training data (causing gradient conflicts) or by using Mixture‑of‑Experts subnetworks (increasing parameter count, latency, and storage). These methods also ignore the sequential relationship exposure → click → conversion.

CSMF Method

CSMF (Cascaded Selective Mask Fine‑Tuning) adapts parameter‑efficient fine‑tuning (PEFT) for large language models to the EBR setting. Training is divided into three cascaded stages—exposure, click, conversion—mirroring the business funnel. Each stage creates a relatively independent parameter subspace via selective masking, preserving upstream knowledge while allocating capacity for downstream objectives.

Training stages

Pre‑training (exposure) : Train the dual‑tower backbone on massive exposure data (positive samples = exposed items, negatives = BNS‑generated or non‑exposed items). After pre‑training, apply a selective mask to prune low‑information neurons, producing a retained‑parameter set and a redundant‑parameter set.

Cascaded fine‑tuning (click) : Freeze the retained exposure parameters, update only the redundant set with click data, then prune again to obtain retained click parameters and a remaining redundant set. A small click‑data subset is used for a second‑stage fine‑tune to recover any lost accuracy before freezing.

Fine‑tuning (conversion) : Freeze both exposure and click retained parameters; update the remaining redundant parameters with conversion data to learn conversion probabilities.

At inference the three disjoint parameter subsets produce exposure, click, and conversion probabilities, which are linearly combined with configurable weights, enabling flexible multi‑objective retrieval without increasing vector dimensionality.

Cumulative Percentile Pruning (CPP)

CPP replaces a fixed pruning ratio per layer (as in PackNet) with a percentile‑based threshold computed from the cumulative distribution of absolute weight values in each layer. Layers with dense information retain more neurons, while sparse layers are pruned aggressively, preserving critical information while reducing parameter space.

Cross‑Stage Adaptive Margin Loss (AML)

AML adjusts the margin between positive and negative samples based on consistency between upstream and downstream tasks. If exposure ranking aligns with click targets, the margin remains small; if they conflict, AML enlarges the margin for the downstream loss, forcing stronger correction. The same principle applies to conversion, using cascaded probabilities as references.

Flexible online multi‑objective retrieval

During deployment, CSMF assigns weights k_d, k_o, k_r to exposure, click, and conversion probabilities. Because the three probability vectors are concatenated, the weighted sum can still be expressed as a single inner product, preserving ANN index structures. Operators can adjust the weight triple to prioritize different business goals without changing vector dimensions or incurring extra latency.

Offline experiments

Experiments on public and industrial datasets compare CSMF against multiple baselines. Results show consistent improvements across core metrics. In Lazada’s live ad system, CSMF deployment yielded +0.42 % ad revenue and +0.57 % click‑through‑rate gains, after which the model was rolled out to full traffic.

Paper

CSMF: Cascaded Selective Mask Fine‑Tuning for Multi‑Objective Embedding‑Based Retrieval<br>Authors: Hao Deng, Haibo Xing, Kanefumi Matsuyama, Moyu Zhang, Jinxin Hu, Hong Wen, Yu Zhang, Xiaoyi Zeng, Jing Zhang<br>Link: https://dl.acm.org/doi/10.1145/3726302.3729939