MGDIN: Deferred Multi-Granularity CTR Model

• The paper introduces MGDIN, which defers interaction of sparse features by leveraging a hierarchical masking schedule to prioritize high-information signals.
• It groups heterogeneous CTR features at multiple granularities to homogenize information density and produce diverse interaction views.
• Deferred activation of low-information features mitigates noise and representation collapse, resulting in improved prediction accuracy and efficiency.

Searching arXiv for the cited MGDIN paper and closely related material. Multi-Granularity Information-Aware Deferred Interaction Network (MGDIN) is a click-through rate (CTR) prediction framework built around the principle that heterogeneous features should not all enter the interaction process simultaneously. It addresses a setting in which feature sparsity, cardinality, and information density vary substantially across user, item, and context signals. Rather than applying a uniform interaction strategy to all features at every layer, MGDIN partitions raw features into multiple grouping granularities and then uses a hierarchical masking schedule to defer low-information groups until deeper layers. The framework is introduced in "Deferred is Better: A Framework for Multi-Granularity Deferred Interaction of Heterogeneous Features" (Xu et al., 13 Mar 2026).

1. Motivation and problem formulation

MGDIN is motivated by a specific failure mode in CTR modeling: prevailing models typically input all features into interaction layers simultaneously even though the feature space is highly heterogeneous. The paper emphasizes three axes of heterogeneity—sparsity, cardinality, and information density—and argues that this mismatch between feature heterogeneity and uniform interaction causes noisy low-information features to be introduced too early (Xu et al., 13 Mar 2026).

The central claim is that early participation of sparse or low-information features can inject noise, overwhelm high-information features, trigger representation collapse, and ultimately hinder the learning of robust representations. This concern is not limited to extreme item-ID sparsity. The paper further states that even after replacing raw item IDs with dense semantic IDs generated from multimodal content and behavior, substantial heterogeneity remains among the other user, item, and context features.

MGDIN is therefore designed to address two coupled problems: uneven information distribution across features and premature interaction collapse. Its organizing thesis is summarized by the slogan “deferred is better”: the model should establish a stable interaction basis from high-information signals before progressively incorporating sparser information.

2. Multi-granularity grouping of heterogeneous features

The first component of MGDIN is Multi-Granularity Feature Grouping. Starting from the raw feature vector $\mathcal{X} = [x_1,\dots,x_n]$, the model forms grouped inputs under multiple window granularities. For the $h$-th branch, the grouping granularity is denoted by $g_h$, and the grouped feature units are contiguous concatenations, with $f_1 = x_1 \oplus x_2 \oplus \cdots \oplus x_{g_h}$. Rather than committing to a single grouping scale, the model uses $K$ different windows in parallel, producing $$\left\{ \mathcal{X}^{1}, \dots, \mathcal{X}^{h}, \dots, \mathcal{X}^{K} \right\}$$ (Xu et al., 13 Mar 2026).

This grouping scheme serves two stated purposes. First, it makes information density within each group more homogeneous, mitigating the effects of extreme individual feature sparsity. Second, it yields diverse interaction views because different window sizes expose different feature-combination scales. The computations for the $K$ windows are performed in parallel.

Within the paper’s formulation, multi-granularity is not merely a data-preprocessing convenience. It is part of the interaction design itself. The grouped units become the entities over which attention scores, masking decisions, and subsequent interaction layers are defined. This implies that MGDIN treats granularity as a structural prior on feature interaction rather than as a separate feature engineering step.

3. Information-aware deferred interaction

The second component, Information-Aware Deferred Interaction, is the defining mechanism of MGDIN. For the $h$-th window granularity, the model first computes an initial attention score matrix from the first-layer grouped representation:

$$\mathbf{A}^{h}_0=(\mathcal{X}^{h}_0\mathbf{Q}^h_0) (\mathcal{X}^{h}_0\mathbf{K}^{h}_0)^{\top}.$$

These initial scores are then used to determine the order in which feature groups are activated across depth (Xu et al., 13 Mar 2026).

The paper’s rationale is that deeper-layer attention becomes less reliable as an indicator of intrinsic information content, because feature representations at that point already include information from other features. Accordingly, MGDIN bases its activation order on the first-layer grouped features. Sparse features are described as often having low attention scores because they are trained fewer times and contain less information.

At interaction layer $l$, only a subset of pairwise interactions is active:

$h$0

with

$h$1

This hierarchical top-$h$2 schedule defines the deferred interaction process. Early layers retain only the most informative interactions, deeper layers progressively unmask more interactions, and final layers approach full activation. The resulting dynamic is explicitly coarse-to-fine: robust structure is learned first from high-information groups, and sparser groups are incorporated later once the representation space is more stable.

After masked interaction, the model applies residual and normalization updates followed by a per-token feed-forward network:

$h$3

$h$4

The intended structure is residual connection, LayerNorm, PFFN, another residual addition, and another normalization step.

4. End-to-end pipeline, prediction, and computational characteristics

The full pipeline begins with raw user profile features, item attributes, and context features, with original item IDs replaced by dense semantic IDs before model input. For each granularity $h$5, the model forms grouped features $h$6, computes the initial attention matrix $h$7, and then runs $h$8 deferred interaction layers under the hierarchical masking rule. After $h$9 layers, the branches produce $g_h$0, which are pooled and passed to an MLP:

$g_h$1

The output is the predicted click probability (Xu et al., 13 Mar 2026).

Training uses standard binary cross-entropy:

$g_h$2

where $g_h$3 is the training dataset, $g_h$4 is the label, and $g_h$5 is the predicted CTR.

The paper also characterizes the model’s efficiency. Its space complexity is

$g_h$6

which is stated to be typically less than the $g_h$7 complexity of standard attention-based methods. Its time complexity is

$g_h$8

because the branches operate in parallel. This makes the architecture notable not only for its deferred interaction policy but also for preserving practical computational tractability.

5. Empirical results, baselines, and ablation evidence

MGDIN is evaluated on Amazon Sports and an industrial dataset containing 7 billion user interaction records from an international e-commerce advertising system. The reported metrics are AUC, GAUC, and LogLoss. Baselines include FM, DNN, Wide & Deep, DeepFM, DCN, AutoInt, GDCN, MaskNet, PEPNet, RankMixer, and OneTrans (Xu et al., 13 Mar 2026).

The main offline results are summarized below.

Dataset	MGDIN	Best baseline
Amazon Sports	0.6789 AUC / 0.6544 GAUC / 0.3539 LogLoss	OneTrans: 0.6775 / 0.6517 / 0.3540
Industrial dataset	0.6994 AUC / 0.5936 GAUC / 0.1116 LogLoss	RankMixer and OneTrans around 0.6956 / 0.5921 / 0.1118

These correspond to reported relative improvements of +0.21% AUC, +0.41% GAUC, and -0.01% LogLoss on Amazon Sports, and +0.54% AUC, +0.25% GAUC, and -0.02% LogLoss on the industrial dataset. The interpretation given in the paper is that methods such as RankMixer and OneTrans already benefit from grouping features, but still retain a fixed, synchronous interaction pattern; they therefore do not address early collapse from low-information features and do not model multiple granularities as effectively.

Ablation studies isolate the two principal components.

Variant	Metrics
MGDIN	0.6994 AUC / 0.5936 GAUC / 0.1116 LogLoss
w/o MG	0.6964 AUC / 0.5910 GAUC / 0.1118 LogLoss
w/o DI	0.6971 AUC / 0.5917 GAUC / 0.1118 LogLoss

The variant w/o MG removes multi-granularity by using a single window, while w/o DI preserves multi-granularity grouping but replaces deferred interaction with standard attention. The reported degradation in both cases supports the paper’s claim that homogenizing information density through grouping is useful but not sufficient: residual heterogeneity remains, so sparse groups still benefit from being deferred.

6. Hyperparameters, online deployment, and interpretation

Two hyperparameter observations are emphasized. First, performance generally improves as the number of granularities $g_h$9 increases, indicating that multiple grouping scales provide useful complementary views of feature interaction. Second, the deferred ratio $f_1 = x_1 \oplus x_2 \oplus \cdots \oplus x_{g_h}$0 has relatively little effect as long as the introduction order remains progressive from dense to sparse (Xu et al., 13 Mar 2026).

This suggests that the critical design choice is the order of activation rather than a finely tuned masking ratio. A plausible implication is that MGDIN’s benefit derives less from a particular sparsification schedule than from enforcing a curriculum over heterogeneous feature groups.

The framework was also deployed in a 10-day online A/B test on a major e-commerce platform. The reported result was a 3.04% relative lift in CTR with no additional inference latency. In the context of industrial recommender systems, this is significant because it indicates that the deferred interaction policy does not require an online serving penalty to realize measurable gains.

The paper’s practical implications are correspondingly direct: not all features should be interacted with immediately; sparse and noisy features should be introduced gradually; modeling heterogeneous feature spaces benefits from multiple grouping scales; and deferred interaction can improve both robustness and business metrics without additional serving cost.

7. Conceptual scope and relation to other multi-granularity models

The term multi-granularity in MGDIN refers specifically to parallel grouping views over heterogeneous CTR features and should not be conflated with all other multi-scale architectures. A useful contrast appears in "MIND-EEG: Multi-granularity Integration Network with Discrete Codebook for EEG-based Emotion Recognition" (Zhang et al., 27 Jan 2025). MIND-EEG is described as conceptually close to an MGDIN-style design because it combines information from multiple spatial scales and uses learned sample-specific network structures, but it is not a deferred interaction network in the strict sense.

MIND-EEG organizes information at three levels—global, intra-regional, and inter-regional—and quantizes learned graph structures through discrete codebooks before fusing the resulting features for classification. The connection to MGDIN is therefore strongest at the level of multi-granularity integration and dynamic interaction structure learning, and weaker at the level of exact architectural mechanism. MGDIN’s distinctive mechanism is not merely multi-scale processing; it is the explicit deferral of low-information feature groups through hierarchical masking.

This distinction helps clarify a common misconception. Multi-granularity alone does not define MGDIN. Its identity rests on the conjunction of two ideas: grouping heterogeneous features into multiple granularities and using an information-aware schedule that delays when those groups are permitted to interact. In that sense, MGDIN is best understood as a deferred interaction framework for heterogeneous feature spaces rather than as a generic multi-branch interaction model.