AlphaFree: User-Free, ID-Free, GNN-Free Recommender

• AlphaFree is a recommendation framework that removes stored user embeddings, raw ID embeddings, and GNNs, dynamically inferring preferences from current interactions.
• It leverages pretrained language models to generate frozen language representations for items, enhancing generalization and reducing memory usage in sparse settings.
• By using similar-item augmentation and symmetric contrastive learning, AlphaFree achieves significant improvements in Recall@20 and NDCG@20 over conventional and LR-based methods.

Searching arXiv for the primary paper and closely related recommendation work to ground the article in current literature. AlphaFree is a top-$K$ recommendation framework that removes three standard design dependencies in recommender systems at once: stored user embeddings, raw ID embeddings, and graph neural networks (GNNs). It formulates recommendation as scoring items $i\in\mathcal V$ for a query interaction set $I_q$ via $y_\theta(i\mid I_q)$, and returns

$$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$

Its central claim is that accurate recommendation can be obtained by inferring user preference on the fly from interacted items, representing items with language representations (LRs) from pretrained LLMs, and injecting collaborative filtering (CF) structure through similar-item augmentation and contrastive alignment rather than graph propagation (Jeon et al., 3 Mar 2026).

1. Problem formulation and motivating departures

AlphaFree studies top-$K$ item recommendation from user-item interaction sets $\mathcal D=\{I\mid I\subseteq\mathcal V\}$, where $I$ denotes a user’s historical interaction set. The method is explicitly framed against three conventions that the paper describes as entrenched: user-dependent recommenders store a trainable embedding for every user, ID-dependent recommenders initialize entities from raw discrete IDs, and GNN-dependent recommenders rely on graph message passing over the user-item graph (Jeon et al., 3 Mar 2026).

The motivation for removing these conventions is structural rather than stylistic. User embeddings scale memory linearly with the number of users, are unreliable under sparse or noisy histories, and generalize poorly to unseen users. Raw IDs are semantically empty and do not naturally support generalization to new or sparsely observed items. GNNs can model high-order CF structure, but they increase memory use, require graph storage and user embeddings, and can suffer from over-smoothing, especially for high-degree or heavy-interaction users (Jeon et al., 3 Mar 2026).

AlphaFree therefore defines three “free” properties. “User-free” means that no per-user embedding table is stored. “ID-free” means that raw item IDs are not used as learned embedding keys. “GNN-free” means that collaborative signals are introduced without graph propagation. A common misconception is that removing users, IDs, and GNNs implies removing personalization or collaborative structure. The method does not do that. It remains personalized by computing preference from the current interaction set, and it remains collaborative by augmenting interactions with behaviorally and semantically similar items and aligning original and augmented views during training (Jeon et al., 3 Mar 2026).

2. Representation model and collaborative augmentation

AlphaFree represents each item $i$ by text content $t_i$, mainly the item title in the experiments, and computes a frozen language representation

$i\in\mathcal V$0

These LRs replace raw ID embeddings. In the reported experiments, Movie and Book use preprocessed text-embedding-3-large embeddings of dimension $i\in\mathcal V$1, while the other datasets use LLaMA-3.1-8B embeddings of dimension $i\in\mathcal V$2 (Jeon et al., 3 Mar 2026).

User preference is inferred on the fly from the interacted items: $i\in\mathcal V$3 This $i\in\mathcal V$4 functions as the latent preference or user-intent representation. The method is therefore not non-personalized; rather, it relocates personalization from a stored user vector to a runtime function of the current history (Jeon et al., 3 Mar 2026).

Collaborative structure is introduced through a preprocessing stage that constructs similar-item sets. For each pair of items $i\in\mathcal V$5, behavioral similarity is the co-occurrence count

$i\in\mathcal V$6

and semantic similarity is the LR dot product

$i\in\mathcal V$7

For each item, AlphaFree first forms a candidate set $i\in\mathcal V$8 of the top-$i\in\mathcal V$9 items by behavioral similarity, then filters those candidates semantically: $I_q$0 This two-stage rule is central to the method’s GNN-free CF mechanism: behavioral similarity supplies collaborative candidates, while semantic filtering suppresses noisy co-occurrence matches (Jeon et al., 3 Mar 2026).

The similar-item sets drive two augmentations. Interaction augmentation expands a history to

$I_q$1

Representation augmentation constructs a collaborative item view

$I_q$2

The paper’s ablations report that removing semantic filtering degrades performance, and the appendix reports that the mean semantic threshold $I_q$3 gives the best Recall@20 among the tested thresholds (Jeon et al., 3 Mar 2026).

3. Dual-view training objective and serving path

AlphaFree trains two two-layer MLP encoders, $I_q$4 and $I_q$5, both mapping from LR dimension $I_q$6 to a learned embedding dimension $I_q$7, fixed to $I_q$8 in the experiments. The original view uses

$I_q$9

and the augmented view uses

$y_\theta(i\mid I_q)$0

The paper’s conceptual interpretation is that $y_\theta(i\mid I_q)$1 captures current preference from observed interactions, whereas $y_\theta(i\mid I_q)$2 captures a CF-enriched version of that preference (Jeon et al., 3 Mar 2026).

Recommendation training uses InfoNCE rather than BPR. For each interaction set $y_\theta(i\mid I_q)$3 in a batch $y_\theta(i\mid I_q)$4, one positive item $y_\theta(i\mid I_q)$5 is sampled and negatives $y_\theta(i\mid I_q)$6 are drawn. With cosine similarity $y_\theta(i\mid I_q)$7, the original-view loss is

$y_\theta(i\mid I_q)$8

with $y_\theta(i\mid I_q)$9, and an analogous loss is defined for the augmented branch (Jeon et al., 3 Mar 2026).

The distinctive component is cross-view alignment. AlphaFree applies symmetric CLIP-style contrastive losses both at the interaction level and at the positive-item level. The purpose is asymmetric distillation without asymmetric inference cost: the augmented branch contains CF-enriched structure, but the original branch is the one used at serving time. The full objective is

$$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$0

where $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$1 sums the original and augmented InfoNCE terms, and $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$2 sums the interaction-level and item-level alignment terms (Jeon et al., 3 Mar 2026).

At inference, only the original-view encoder is retained. For a query interaction set $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$3,

$$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$4

and items are ranked by cosine similarity

$$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$5

This serving path is lightweight by construction: one averaged history vector, one MLP, and cosine retrieval. Another common misconception is that AlphaFree fine-tunes the LLM end to end. It does not. LRs are computed once in preprocessing and the trainable parameters are only the two MLPs, $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$6 (Jeon et al., 3 Mar 2026).

4. Training methodology and reported empirical behavior

The experimental study uses seven real-world datasets: Amazon Movie, Amazon Book, Amazon Video, Amazon Baby, Steam, Amazon Beauty, and Amazon Health. Dataset statistics range up to $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$7 users, $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$8 items, and $$\mathcal{R}_{I_q}:=\operatorname{argtop}_{K}\Bigl(\{y_\theta(i\mid I_q)\mid i\in\mathcal{V}\}\Bigr).$$9 interactions. The main evaluation protocol is user-wise holdout, splitting each user history into train/validation/test with ratio $K$0. Metrics are Recall@20 and NDCG@20, each result is averaged over 5 random seeds, and hyperparameters are selected on validation Recall@20 (Jeon et al., 3 Mar 2026).

The optimization setup is explicit: Adam optimizer, learning rate $K$1, weight decay $K$2, batch size $K$3, embedding dimension $K$4, maximum epochs $K$5, early stopping patience $K$6, and $K$7 negatives. The hyperparameter grid is

$K$8

The paper notes that larger $K$9 tends to help sparse datasets such as Baby, while denser datasets such as Movie tend to prefer smaller $\mathcal D=\{I\mid I\subseteq\mathcal V\}$0 (Jeon et al., 3 Mar 2026).

Baselines are divided into non-LR-based methods—MF-BPR, FISM-BPR, LightGCN, XSimGCL—and LR-based methods—RLMRec and AlphaRec. In the main table, AlphaFree achieves the best Recall@20 and NDCG@20 on all reported datasets. The reported gains are up to 5.37% Recall@20 and 5.71% NDCG@20 over the best LR-based method, and up to roughly 42% improvement over the best non-LR-based method. Examples reported in the table include Movie with Recall@20 $\mathcal D=\{I\mid I\subseteq\mathcal V\}$1 and NDCG@20 $\mathcal D=\{I\mid I\subseteq\mathcal V\}$2, Book with $\mathcal D=\{I\mid I\subseteq\mathcal V\}$3 and $\mathcal D=\{I\mid I\subseteq\mathcal V\}$4, Steam with $\mathcal D=\{I\mid I\subseteq\mathcal V\}$5 and $\mathcal D=\{I\mid I\subseteq\mathcal V\}$6, Beauty with $\mathcal D=\{I\mid I\subseteq\mathcal V\}$7 and $\mathcal D=\{I\mid I\subseteq\mathcal V\}$8, and Health with $\mathcal D=\{I\mid I\subseteq\mathcal V\}$9 and $I$0 (Jeon et al., 3 Mar 2026).

The cold-start protocol is stricter: users are split $I$1 into train/validation/test, and only $I$2 of each validation/test user’s interactions, with at least one interaction, is used as the query $I$3. Under this setting, AlphaFree reports gains up to 38.65% Recall@20 and 35.95% NDCG@20 over the tested baselines. The paper interprets this as support for user-free inference on unseen users (Jeon et al., 3 Mar 2026).

The heavy-user analysis groups users by history length and reports that AlphaFree consistently outperforms AlphaRec, with the margin increasing for users with 61+ interactions. The paper interprets that pattern as evidence that AlphaRec’s GNN layers suffer from over-smoothing, whereas AlphaFree avoids that specific failure mode (Jeon et al., 3 Mar 2026).

Operational robustness also appears in the runtime results. RLMRec runs out of time on many datasets, AlphaRec runs out of memory on Beauty and Health, whereas AlphaFree runs on all datasets. During training, AlphaFree uses up to 69% less GPU memory than AlphaRec, and inference can use up to 41% less memory than AlphaFree’s own training mode because only the original-view encoder is used (Jeon et al., 3 Mar 2026).

5. Complexity profile, ablations, and limitations

The method’s main computational caveat is preprocessing. Let $I$4, $I$5, and $I$6. The paper states the time complexity as

$I$7

for preprocessing,

$I$8

per training epoch, and

$I$9

for inference. The dominant caveat is preprocessing’s $i$0 item-similarity computation, although it is run once (Jeon et al., 3 Mar 2026).

The stated space complexity is

$i$1

for preprocessing,

$i$2

for training, and

$i$3

for inference. The paper emphasizes that AlphaFree removes the $i$4 user-embedding memory term carried by methods such as LightGCN and AlphaRec (Jeon et al., 3 Mar 2026).

The ablation study removes semantic filtering, interaction-level alignment, item-level alignment, and all alignment losses. All such variants degrade performance. The reported interpretation is consistent across datasets: semantic filtering improves similar-item quality, and both interaction-level and item-level alignment contribute to the final model (Jeon et al., 3 Mar 2026).

The limitations are explicit. AlphaFree requires textual metadata and is therefore not directly applicable to settings without item titles or descriptions. Its effectiveness is bounded by the quality of the chosen pretrained LLM; the appendix reports that with weaker LMs, explicit GNN structural bias can still help. LR generation dominates preprocessing time—98–99% of preprocessing in the runtime tables—and similar-item construction is $i$5. The paper also restricts scope to static top-$i$6 recommendation, not sequential, session-based, temporal, or multi-behavior recommendation (Jeon et al., 3 Mar 2026).

A plausible implication is that AlphaFree is best read as a particular decomposition of recommendation rather than as a universal theorem about recommender design. Its empirical profile is strongest when item text is available, high-dimensional semantic representations are affordable in preprocessing, and serving-time memory pressure or unseen-user generalization is important.

6. Conceptual position and disambiguation

AlphaFree’s novelty is not any single ingredient in isolation. Language representations, contrastive learning, and set-based preference aggregation all have adjacent literatures. The specific claim of the paper is that prior methods remove at most one or two of the three dependencies—users, IDs, and GNNs—whereas AlphaFree removes all three simultaneously (Jeon et al., 3 Mar 2026).

This also clarifies what AlphaFree is not. It is not a purely content-only recommender, because co-occurrence-based behavioral similarity and cross-view alignment are used to transfer collaborative signals into the serving encoder. It is not a user-agnostic recommender, because personalization is inferred from the current interaction set. It is not a language-model fine-tuning framework, because the LM is frozen and only lightweight MLPs are trained (Jeon et al., 3 Mar 2026).

The name should also not be conflated with unrelated uses of “alpha” in nearby literatures. “AFL: A Single-Round Analytic Approach for Federated Learning with Pre-trained Models” (He et al., 2024) concerns analytic federated adaptation of frozen backbones rather than recommendation. “Beyond Size and Class Balance: Alpha as a New Dataset Quality Metric for Deep Learning” (Couch et al., 2024) studies big-$i$7 dataset diversity measures rather than a recommender architecture. Alpha-fairness formulations in wireless networking, such as (Singh et al., 2015) and (Deng et al., 16 May 2026), use $i$8 as a fairness parameter in resource allocation, again with no connection to the user-free, ID-free, GNN-free recommender defined by AlphaFree.

Within recommender-systems research, AlphaFree is therefore most precisely characterized as a user-free, ID-free, GNN-free top-$i$9 recommendation framework built from frozen language representations, behaviorally and semantically filtered similar-item augmentation, dual-branch contrastive training, and a single-branch inference path. Its reported contribution is to show that this combination can outperform both conventional and LR-based baselines while reducing memory usage and improving generalization to unseen users (Jeon et al., 3 Mar 2026).