FastMMRec: Efficient Multimodal Recommendation

• FastMMRec is a multimodal recommendation framework that defers training-time graph convolution by using a static, fused item–item graph for enhanced embedding learning.
• The model fuses modality-specific embeddings through trainable weighting to preserve cross-modal coherence and optimize user–item interactions.
• It applies a LightGCN-style propagation only at test time, significantly reducing training overhead and achieving measurable improvements in Recall and NDCG metrics.

Searching arXiv for the FastMMRec paper and closely related multimodal recommendation/LightGCN references to ground the article. Searching arXiv for "The Best is Yet to Come: Graph Convolution in the Testing Phase for Multimodal Recommendation" and related baseline papers. to=arxiv_search {"query":"The Best is Yet to Come: Graph Convolution in the Testing Phase for Multimodal Recommendation", "max_results": 5} FastMMRec is a multimodal recommendation framework for user–item interaction graphs in which graph convolution is removed from training and deferred to inference. It is defined in the setting of multimodal recommendation (MMRec), where items carry multiple modalities—typically visual and textual features—and user preference over an item is modeled from user embeddings and modality-specific item embeddings. The central claim is that Graph Convolutional Networks (GCNs), although widely used in MMRec for neighbor aggregation, are better treated as a test-time representation enhancer than as a train-time component. In this formulation, FastMMRec trains multimodal embeddings with a frozen multimodal item–item graph and standard Bayesian Personalized Ranking (BPR), then applies a LightGCN-style propagation over the user–item graph only during testing (Xu et al., 24 Jul 2025).

1. Problem formulation and motivation

FastMMRec is built on a user–item interaction graph

$$\mathcal{G} = (\mathcal{U}, \mathcal{I}, \mathcal{E}),$$

where $\mathcal{U}$ is the user set, $\mathcal{I}$ is the item set, and $\mathcal{E}$ contains binary implicit interactions. For each item $i$ and modality $m \in \mathcal{M}$, the model assumes an embedding $e_{i^m}$ derived from pre-trained encoders; user embeddings $e_{u^m}$ are randomly initialized and learned. A generic MMRec predictor is abstracted as

$$s_{u,i} = f(e_u, \{ e_{i^m} \mid m \in \mathcal{M} \} \mid \Theta),$$

with $s_{u,i}$ the predicted preference score.

The motivation for FastMMRec begins from the observation that GCNs, especially LightGCN-style propagation, are standard in recommender systems because they smooth embeddings along edges and aggregate collaborative signals. In multimodal settings, this is typically instantiated per modality. A LightGCN-style update for each modality $\mathcal{U}$0 takes the form

$\mathcal{U}$1

followed by scoring and BPR-style optimization.

FastMMRec is motivated by two challenges attributed specifically to using GCNs during training. The first is efficiency and scalability: training cost grows roughly with $\mathcal{U}$2, and MMRec often requires one GCN per modality. The second is representational and optimization-related. FastMMRec argues that train-time GCNs create unhelpful or harmful induced training pairs beyond the sampled triplet $\mathcal{U}$3 and can also produce modality isolation, in which modality-specific embedding spaces drift or remain tied to their own local graph structure instead of aligning in a shared multimodal space. Empirically, a variant that uses GCN only at test time outperforms standard MMGCN and trains faster, while cosine similarity between visual and textual embeddings is higher without GCN in training (Xu et al., 24 Jul 2025).

2. GCN-free training architecture

The training phase of FastMMRec contains no user–item GCN. Instead, it uses modality-specific embeddings, a fused multimodal representation, and a fixed item–item graph constructed from multimodal similarity. This design preserves direct user–item supervision during optimization while avoiding user–item neighborhood mixing.

For each modality $\mathcal{U}$4, item embeddings are arranged as

$\mathcal{U}$5

and user embeddings as

$\mathcal{U}$6

These are pre-GCN embedding spaces. Item–item relations are then derived separately for each modality through cosine similarity:

$\mathcal{U}$7

For each item, only the top-$\mathcal{U}$8 neighbors are retained:

$\mathcal{U}$9

and the adjacency is symmetrically normalized:

$\mathcal{I}$0

The modality-specific item graphs are fused into a single unified graph:

$\mathcal{I}$1

where $\mathcal{I}$2 are trainable modality weights. This unified item–item graph is computed once offline and frozen during training.

User and item representations are fused by weighted concatenation:

$\mathcal{I}$3

If each modality embedding has dimension $\mathcal{I}$4, the fused dimension is $\mathcal{I}$5.

Item representations are then enhanced by multi-hop linear propagation on the unified item–item graph:

$\mathcal{I}$6

where $\mathcal{I}$7 is the number of hops. In the one-hop case, this corresponds informally to adding the normalized average of the top-$\mathcal{I}$8 similar items’ embeddings to the item representation. This is not a GCN on the user–item graph; it is a fixed linear enhancement over an items-only graph. A plausible implication is that FastMMRec uses graph structure selectively: semantic item proximity enters training, whereas collaborative user–item propagation is postponed until embeddings are already optimized (Xu et al., 24 Jul 2025).

3. Test-phase graph convolution

FastMMRec applies graph convolution only after training has converged. At test time, it runs a standard LightGCN-style propagation over the original user–item graph, using the trained multimodal embeddings as input features:

$\mathcal{I}$9

for layers $\mathcal{E}$0.

The initial features $\mathcal{E}$1 and $\mathcal{E}$2 are the trained embeddings obtained from the GCN-free optimization stage, including multimodal fusion and item–item enhancement. After $\mathcal{E}$3 propagation layers, FastMMRec aggregates the refined embeddings as

$\mathcal{E}$4

and ranks items with the inner product

$\mathcal{E}$5

This test-time propagation is purely forward: no parameters are updated and no backpropagation is performed. Because the convolution is absent from the training computational graph, the model avoids storing intermediate activations for gradients. The refined embeddings $\mathcal{E}$6 and $\mathcal{E}$7 can be precomputed once after training and then reused for inner-product inference.

The conceptual distinction from conventional GCN recommenders is precise. Standard graph-based recommenders apply GCN propagation in both training and testing and optimize embeddings through GCN layers. FastMMRec instead treats graph convolution as a post-hoc representation enhancer. This suggests a decoupled view of recommendation pipelines in which representation learning and graph-based smoothing need not share the same optimization path (Xu et al., 24 Jul 2025).

4. Optimization analysis and modality isolation

FastMMRec uses classic BPR with $\mathcal{E}$8 regularization:

$\mathcal{E}$9

where $i$0 is a triplet of user, positive item, and sampled negative item; $i$1 is the sigmoid; and $i$2 includes user and item embeddings together with the modality weights $i$3.

The core analytical claim is that applying BPR through a LightGCN induces additional optimization constraints that are not limited to the sampled triplet. For one-layer propagation, the paper decomposes the resulting objective into terms P1–P5:

• P1: Node $i$4 with Node $i$5
• P2: Node $i$6 with Neighbors of $i$7
• P3: Node $i$8 with Neighbors of $i$9
• P4: Neighbors of $m \in \mathcal{M}$0 with Neighbors of $m \in \mathcal{M}$1
• P5: Neighbors of $m \in \mathcal{M}$2 with Neighbors of $m \in \mathcal{M}$3

These induced terms encode assumptions such as “all users who purchased a negative item are dissimilar to the target user” or “all items co-consumed with a negative item are dissimilar to items associated with the target user.” FastMMRec identifies such assumptions as unrealistic and potentially harmful to generalization.

By contrast, the FastMMRec training objective expands only into a direct user–item term and an item-neighbor term derived from the multimodal item–item graph:

$m \in \mathcal{M}$4

Here the neighbors are semantically similar items, not other users. Consequently, FastMMRec frames the induced constraints as semantically coherent: users are pulled toward items similar to positives and away from items similar to negatives.

A second analytical target is modality isolation. FastMMRec measures alignment between visual and textual embeddings by average cosine similarity:

$m \in \mathcal{M}$5

The reported values show lower similarity for MMGCN-like training with GCN and higher similarity when GCN is moved to test time; on Baby, the reported numbers are 0.2207 for MMGCN-like training, 0.3722 for the test-time-only GCN variant, and 0.3848 for FastMMRec. Other advanced MMRec models with GCN also show lower visual–textual similarity, approximately 0.20–0.30. FastMMRec interprets this as evidence that train-time GCN aggregation can smear modality-specific patterns across neighbors and reduce the ability to learn effective modality weights, whereas GCN-free training in a shared fused space better preserves cross-modal coherence (Xu et al., 24 Jul 2025).

5. Empirical profile

The empirical study uses three Amazon review datasets in 5-core form with visual and textual features, split 8:1:1 into training, validation, and test sets.

Dataset	Users	Items
Baby	19,445	7,050
Sports	35,598	18,357
Clothing	39,387	23,033

The corresponding interaction counts are 160,792 for Baby, 296,337 for Sports, and 278,677 for Clothing; sparsity levels are 99.88%, 99.95%, and 99.97%, respectively. Visual features are 4,096-dimensional and textual features are 384-dimensional.

The baseline set includes conventional recommenders—MF-BPR, LightGCN, SimGCL, LayerGCN—and multimodal recommenders such as VBPR, MMGCN, DualGNN, LATTICE, FREEDOM, SLMRec, BM3, MMSSL, LGMRec, and DiffMM. Across Recall@10, Recall@20, NDCG@10, and NDCG@20, FastMMRec is reported to achieve the best result on all datasets and metrics. For Recall@20 and NDCG@20, the reported outcomes are:

• Baby: FastMMRec reaches R@20 = 0.1034 and N@20 = 0.0453; relative improvement over the best baseline is +4.23% and +5.35%.
• Sports: FastMMRec reaches R@20 = 0.1151 and N@20 = 0.0517; relative improvement is +5.89% and +7.48%.
• Clothing: FastMMRec reaches R@20 = 0.0992 and N@20 = 0.0447; relative improvement is +5.42% and +6.43%.

The reported improvements are statistically significant, with p-values around $m \in \mathcal{M}$6. Ablation studies further separate the contributions of the item–item enhancement and the test-time GCN. Removing the item–item graph (“w/o-item”) significantly degrades performance, and using item–item enhancement only at test time (“test-item”) underperforms the full FastMMRec configuration. This is presented as evidence that item–item enhancement is beneficial during optimization, whereas user–item graph convolution is beneficial after optimization.

The test-time GCN depth is varied over $m \in \mathcal{M}$7. Performance improves up to $m \in \mathcal{M}$8 or 4 and does not exhibit strong over-smoothing, unlike models trained with GCN. Suggested item–item graph hyperparameters are $m \in \mathcal{M}$9 for Baby and $e_{i^m}$0 for Sports and Clothing, with top-$e_{i^m}$1 set to 10 for Baby and Sports and 5 for Clothing. Under different user sparsity ranges, FastMMRec consistently outperforms strong baselines, which the study presents as robustness under highly sparse interaction regimes (Xu et al., 24 Jul 2025).

6. Efficiency, scalability, limitations, and significance

FastMMRec’s efficiency claim is grounded in the removal of graph convolution from the training stage. In the complexity analysis, train-time graph convolution costs are listed as

$e_{i^m}$2

for MMGCN and MMSSL, and

$e_{i^m}$3

for LATTICE, whereas FastMMRec has no graph convolution in training and therefore a graph-convolution cost of 0. For feature mapping, FastMMRec is assigned

$e_{i^m}$4

in contrast to more expensive alternatives such as LATTICE’s

$e_{i^m}$5

For the loss term, MMGCN, LATTICE, and FastMMRec all use

$e_{i^m}$6

while MMSSL incurs additional adversarial and contrastive costs.

The empirical efficiency measurements on an RTX 3090 with 24GB memory show consistent reductions in epoch time and memory. On Baby, FastMMRec uses 0.61 s/epoch and 1.93 GB, compared with 4.09 s/epoch and 2.69 GB for MMGCN, 3.20 s and 4.53 GB for LATTICE, 6.31 s and 3.77 GB for MMSSL, and 9.45 s and 4.23 GB for DiffMM. On Sports, FastMMRec uses 1.01 s and 2.79 GB, while LATTICE reaches 19.93 GB. On Clothing, FastMMRec uses 1.39 s and 3.11 GB, compared with 28.22 GB for LATTICE. Convergence plots are reported to show that FastMMRec reaches higher Recall@20 in fewer epochs than LATTICE and FREEDOM.

The limitations are correspondingly specific. The item–item graphs are precomputed and frozen, and the user–item graph is assumed relatively stable, so dynamic environments may require recomputation. The model has no explicit cold-start mechanism beyond multimodal item features from pre-trained visual and textual encoders. Test-time GCN still incurs a cost of $e_{i^m}$7, although the study notes that this can be amortized by precomputing refined embeddings after training. The experiments are limited to two modalities, images and text, so extension to additional modalities may require careful tuning of $e_{i^m}$8 and graph construction.

The principal significance of FastMMRec is not the abandonment of GCNs but their reassignment. A common misconception would be to interpret the framework as anti-GCN; the reported evidence instead argues that GCN’s main utility is neighbor-based representation enhancement, while its participation in optimization can distort the loss landscape and worsen multimodal alignment. On that reading, FastMMRec contributes a decoupled training–inference paradigm for MMRec: multimodal representations are learned in a simpler supervised setting, semantic item similarity is incorporated during training through a frozen unified item–item graph, and graph convolution is reserved for the testing phase as a shallow refinement step (Xu et al., 24 Jul 2025).