Multi-Level Conflict-Aware Network (MCAN)

• MCAN is a specialized neural architecture that explicitly models both alignment and conflict in multimodal sentiment analysis using dual-branch fusion.
• It employs a micro and macro branch design, utilizing SVD-based decomposition and conflict-aware cross-attention to segregate and integrate multimodal signals.
• Empirical results on datasets like CMU-MOSI demonstrate enhanced predictive performance, with ablation studies underscoring the role of discrepancy constraints.

A Multi-Level Conflict-Aware Network (MCAN) is a specialized neural architecture designed to address the challenges of modeling both alignment and conflict in multimodal sentiment analysis, with explicit mechanisms for disentangling and leveraging inter-modal contradictions. MCAN also refers, in the context of multi-cloud network scheduling, to a hybrid conflict-aware resource allocation paradigm based on conflict graphs and maximum-weight independent set solutions. This entry focuses primarily on its formulation and impact in multimodal machine learning, referencing recent results, but additionally contrasts with the network scheduling domain to clarify the broader conceptual underpinnings.

1. Motivation and Theoretical Foundations

MCAN is motivated by the need to move beyond traditional multimodal fusion models that emphasize only cross-modal alignment (i.e., extracting agreement across modalities such as text, audio, and vision) or generic modality-invariant representations. Prior works neglect the fact that real multimodal input often contains explicit conflicts—cases where modalities provide contradictory sentiment cues (e.g., positive text with sarcastic prosody). MCAN formalizes and operationalizes the notion that not only should agreement be modeled, but so should disagreement, at multiple semantic levels (Gao et al., 13 Feb 2025).

The core principle is the progressive segregation of alignment and conflict information during the fusion process. This is achieved via explicit architectural modules and mathematical criteria that identify, isolate, and re-integrate conflicting constituents for both representation learning and predictive modeling.

2. Network Architecture and Module Details

MCAN is organized in two coupled branches: a main fusion/alignment branch and a conflict modeling branch.

2.1 Unimodal Encoders

• Text: Encoded with a pre-trained BERT model, producing hidden states $F_t \in \mathbb{R}^{n_t \times d}$.
• Audio: Processed with a two-layer bi-directional LSTM, yielding $F_a \in \mathbb{R}^{n_a \times d}$.
• Vision: Also processed with a two-layer bi-directional LSTM, generating $F_v \in \mathbb{R}^{n_v \times d}$.

2.2 Alignment Segregation (Main Branch)

• Micro Multi-Step Interaction Network (Micro-MSIN):
  • Fuses unimodal pairs (text-audio, text-vision) through stacked cross-transformer layers.
  • After $I$ layers, the resulting features are concatenated and subjected to singular value decomposition (SVD), yielding:
  • Alignment components ($F_{t,a}^{aligned}$, $F_{t,v}^{aligned}$) from the first $k$ singular vectors.
  • Conflict components ($F_{t,a}^{conflict}$, $F_{t,v}^{conflict}$) from the remaining vectors.
• Macro Multi-Step Interaction Network (Macro-MSIN):
  • Further fuses aligned bimodal representations, again with cross-transformer layers and SVD, providing $F_c^{aligned}$ and $F_c^{conflict}$.
  • The aligned macro-level component is used for final sentiment prediction, while conflict components are routed to the conflict branch.

2.3 Conflict Modeling Branch

• Micro Conflict-Aware Cross-Attention (Micro-CACA):
  • For each modality $p \in \{t, a, v\}$, takes the two conflict components involving $p$ as query, and attends over $F_p$ (its unimodal features).
  • Outputs $F_{p}'$, each used for auxiliary sentiment prediction $\hat{y}_p'$.
  • Enforces representation orthogonality and predictive discrepancy across modalities.
• Macro Conflict-Aware Cross-Attention (Macro-CACA):
  • Takes bimodal conflict constituents as queries and attends into the “baseline” bimodal fused features.
  • Produces $F_{t,a}''$, $F_{t,v}''$ and their respective predictions, again with discrepancy constraints.

3. Key Mathematical Components

3.1 Alignment versus Conflict Decomposition

For a fused feature matrix $F$, SVD yields

$F = U \Sigma V^\top$

Selecting the top $k$ singular values produces the aligned component: $$F^{aligned} = U_{[:,1:k]}\Sigma_{1:k,1:k}V_{[:,1:k]}^\top$$ with the conflict component defined as

$F^{conflict} = F - F^{aligned}$

or, equivalently, as the sum over the remaining $(k+1):h$ singular vectors.

3.2 Discrepancy Constraints

These constraints ensure that conflict-extracted representations and their predictive outputs are decorrelated:

• Representation-level orthogonality:

$$\mathcal{L}_{micro}^{oc} = \sum_{p \neq q} \| F_p'^{\top} F_q' \|_F^2$$

$$\mathcal{L}_{macro}^{oc} = \| F_{t,v}''^{\top} F_{t,a}'' \|_F^2$$

• Prediction-level difference:

$$\mathcal{L}_{micro}^{diff} = \sum_{p \neq q} | \hat{y}_p' - \hat{y}_q' |^2$$

$$\mathcal{L}_{macro}^{diff} = | \hat{y}_{t,v}'' - \hat{y}_{t,a}'' |^2$$

• Main sentiment loss: Mean squared error over the primary output.

$$\mathcal{L}_{main} = \frac{1}{N} \sum_{i=1}^N (y_i - \hat{y}_i)^2$$

• Overall objective:

$$\mathcal{L} = \mathcal{L}_{main} + \alpha ( \mathcal{L}_{micro}^{oc} + \mathcal{L}_{macro}^{oc}) + \beta ( \mathcal{L}_{micro}^{diff} + \mathcal{L}_{macro}^{diff})$$

with hyperparameters $\alpha=0.01$, $\beta=0.001$ typically.

4. Model Training and Implementation

• Joint training of main and conflict branches is performed end-to-end.
• MCAN avoids noisy, externally-generated unimodal labels by applying discrepancy constraints directly on internal model predictions and representations.
• Optimizer: Adam with learning rates $5 \times 10^{-5}$ (for BERT) and $1 \times 10^{-4}$ (elsewhere). Standard training partitions and batch sizes are used for CMU-MOSI and MOSEI datasets.
• The SVD truncation hyperparameter $k$ is critical; optimal performance was observed around $k=44$ on validation.

5. Empirical Performance and Ablation

MCAN was evaluated on CMU-MOSI (2,199 clips) and CMU-MOSEI (23,453 clips), with sentiment in $[-3, 3]$. Metrics include Acc$_2$ (binary), Acc$_7$ (7-class), F1, Pearson correlation, and mean absolute error (MAE) (Gao et al., 13 Feb 2025).

Dataset	Metric	MCAN	Best Baseline
CMU-MOSI	Acc$_2$ (\%)	84.5	84.2
CMU-MOSI	Corr	0.811	0.805
CMU-MOSI	MAE	0.675	0.671

Ablation studies demonstrate that:

• Removing the conflict modeling branch (“w/o CMB”) reduces Acc$_2$ to 82.3%.
• Eliminating either discrepancy loss ($\mathcal{L}^{diff}$ or $\mathcal{L}^{oc}$) causes a drop of approximately 2 points in Acc$_2$.
• Model performance peaks at SVD truncation $k \approx 44$, supporting the explicit decomposition of alignment versus conflict.

MCAN outperforms all baselines, including TFN, LMF, MARN, RAVEN, MulT, MISA, Self-MM, GFML, MMIN, and MSAN, across both datasets.

6. Comparative Perspective: MCAN in Multi-Cloud Scheduling

In multi-cloud radio access networks, MCAN references a distinct but structurally related framework: the Multi-Cloud hybrid scheduling model that leverages a conflict-aware assignment using conflict graphs and maximum-weight independent set (MWIS) optimization (Douik et al., 2016). In that context, the set of feasible user-to-resource assignments is modeled as the independent set of a conflict graph, with various solution paradigms (centralized optimal, distributed optimal, heuristic) providing significant gains over scheduling-only schemes. While the underlying application differs, the foundational insight—explicit modeling and exploitation of conflict at multiple granularity levels—remains a unifying theme.

7. Limitations and Future Research

MCAN’s reliance on SVD requires dataset-specific selection of the truncation parameter $k$, which can affect performance and may not generalize out-of-the-box. All conflict signals are currently weighted equally, disregarding their semantic import or severity. The model presently eschews optimization-level conflict measures (e.g., Jacobian-based analysis), which could provide more nuanced modeling of modality interactions. Further extensions may include hierarchical weighting of conflict, adaptive SVD truncation, and integration with gradient-level conflict metrics (Gao et al., 13 Feb 2025).

A plausible implication is that MCAN’s dual-branch, explicit alignment/conflict architecture sets a new standard for principled multimodal fusion, particularly in domains where inter-modal contradiction is semantically meaningful and abundant.