CPGRec+: Balanced Game Recommendation

• The paper introduces CPGRec+ which balances accuracy and diversity by integrating PER for edge reweighting and PRG for LLM-driven preference extraction.
• It employs two novel modules that use statistical tests and LLM semantic insights to differentiate significant player-game interaction signals.
• Experimental results on Steam datasets demonstrate superior accuracy and diversity compared to state-of-the-art GNN-based recommender models.

CPGRec+ is a balance-oriented framework for personalized video game recommendations that extends CPGRec by addressing a specific limitation in Graph Neural Network (GNN)-based recommenders: existing methods primarily prioritize accuracy over diversity, and CPGRec itself fails to account for critical disparities in player-game interactions, which carry varying significance in reflecting players' personal preferences and may exacerbate over-smoothness issues inherent in GNN-based models. The framework introduces two modules—Preference-informed Edge Reweighting (PER) and Preference-informed Representation Generation (PRG)—to distinguish significant interest and disinterest signals and to leverage LLMs for contextualized game and player descriptions. Experiments on two Steam datasets report superior accuracy and diversity over state-of-the-art models (Li et al., 16 Apr 2026).

1. Architectural position and data model

CPGRec+ operates on players $U$ and games $I$. For each player-game interaction, the inputs are dwelling time $t_{u,i}$ and implicit feedback in the form of play history. For each game, the available attributes are category metadata—genre, developer, publisher—together with popularity measured by player counts, average rating $r_i$, price, and release date.

The graph construction is explicitly multi-relational. Three item-item “strict” graphs are built: $G^{g\wedge d}$, $G^{g\wedge p}$, and $G^{d\wedge p}$, where an edge exists only if two games share at least two category attributes. In parallel, one item-item “connectivity” graph $G^{Co}$ is constructed where games share at least one category. A player-game bipartite graph $G^{bi}$ contains edges $(u,i)$ whenever player $I$0 has played game $I$1.

The overall architecture is organized into an accuracy-driven path, a diversity-driven path, and a comprehensive fusion stage. On the accuracy side, Stringency-improved Game Connection (SGC) learns $I$2 via LightGCN on the three strict graphs plus graph-wise attention; PER computes signed, information-content weights $I$3 for edges in $I$4; and PRG prompts an LLM to produce textual descriptions $I$5 for games and $I$6 for players, embeds them via a text encoder, and fuses them with $I$7 and $I$8. On the diversity side, Connectivity-enhanced Neighbor Aggregation (CNA) performs multi-layer LightGCN on $I$9 with layer-wise reweighting $t_{u,i}$0 to obtain $t_{u,i}$1, while Popularity-guided Edges and Nodes Reweighting (PENR) precomputes node and edge weights $t_{u,i}$2 and $t_{u,i}$3 on $t_{u,i}$4 and incorporates them into LightGCN message passing to obtain $t_{u,i}$5 and $t_{u,i}$6.

The final item representation is fused by learnable coefficients:

$t_{u,i}$7

This suggests that CPGRec+ treats recommendation as a joint optimization of accuracy-sensitive and diversity-sensitive signals rather than as a pure ranking problem followed by reranking.

2. Preference-informed Edge Reweighting

PER is motivated by two claims. First, in standard GCN, all edges are equally positive, so representations of $t_{u,i}$8 and all its neighbors converge, producing over-smoothing. Second, historic plays are heterogeneous: some reflect strong personal interest, such as long play versus average, while others reflect disinterest, such as short play despite a highly rated game. PER therefore assigns a positive sign to edges where a player’s dwelling time deviates significantly upward from the global rating distribution, a negative sign where it deviates downward, and zero otherwise. The magnitude is then scaled by information content.

The preprocessing stage applies a Box-Cox transform with parameter $t_{u,i}$9 to both $r_i$0 and $r_i$1, followed by Z-score normalization to obtain $r_i$2 and $r_i$3. A KS-test confirms empirical fit to standard normal. This normalization makes the subsequent significance test and information-content calculation comparable across interaction time and rating statistics.

The sign decision is Fisher-distribution-based. The statistic is defined as

$r_i$4

Under $r_i$5 independent, $r_i$6. Let $r_i$7 be the $r_i$8 upper quantile of $r_i$9. Then

$G^{g\wedge d}$0

In this formulation, the sign is not a heuristic label; it is the output of a statistical decision rule tied to normalized dwelling time and normalized rating.

The edge-weight magnitude is defined by information content:

$G^{g\wedge d}$1

The final PER edge weight is

$G^{g\wedge d}$2

Within graph convolution on $G^{g\wedge d}$3, the standard adjacency weight is replaced by the balance-oriented weight

$G^{g\wedge d}$4

The stated effect is that positive edges pull representations together, negative edges push them apart, and zero edges block over-smoothing (Li et al., 16 Apr 2026).

3. Preference-informed Representation Generation

PRG provides what the framework describes as an external knowledge and reasoning channel. Its role is to enrich raw embeddings with LLM-derived semantics of global and personal preferences. The module therefore does not replace graph-based representation learning; it refines it by injecting text-conditioned semantic information.

For rating-informed game description generation, the prompt $G^{g\wedge d}$5 contains the game title, average rating $G^{g\wedge d}$6 interpreted as global interest, price $G^{g\wedge d}$7, release date $G^{g\wedge d}$8, and the instruction: “Focus on why players like/dislike this game, given its average rating.” The LLM used is Qwen2.5. The output is a textual game description $G^{g\wedge d}$9 emphasizing aspects correlated with $G^{g\wedge p}$0. This description is embedded via a pretrained text encoder EMB, with “e.g. M3-Embedding” given as the example, to produce $G^{g\wedge p}$1, and is then aligned and fused with the game representation.

For preference-informed player description generation, the input to the LLM includes, for each historical game $G^{g\wedge p}$2, its LLM-generated description $G^{g\wedge p}$3, the normalized dwelling time $G^{g\wedge p}$4, and the normalized rating $G^{g\wedge p}$5. The prompt explicitly instructs: “Compare your own play time vs. the game’s rating; infer the player’s unique tastes.” The resulting textual player description $G^{g\wedge p}$6 is embedded to $G^{g\wedge p}$7, then aligned and fused into the player representation.

The significance of PRG lies in its use of comparative reasoning. Rather than describing games only by metadata or players only by interaction aggregates, the module compares personal play behavior against global rating signals. A plausible implication is that PRG attempts to encode preference residuals—how a player’s behavior departs from aggregate sentiment—into the final embedding space.

4. Optimization objective and balance mechanism

CPGRec+ is trained with a reweighted Bayesian Personalized Ranking objective. For each player $G^{g\wedge p}$8, positive item $G^{g\wedge p}$9, and negative item $G^{d\wedge p}$0, the scores are

$G^{d\wedge p}$1

The loss is

$G^{d\wedge p}$2

Here $G^{d\wedge p}$3 is the sigmoid, $G^{d\wedge p}$4 controls negative-sample score reweighting (NSR) intensity, and $G^{d\wedge p}$5 is the $G^{d\wedge p}$6 regularization weight.

The paper characterizes the balance between accuracy and diversity in three coordinated terms. Raising negative scores of low-similarity, often long-tail, items encourages diversity. PER’s signed weights preserve sharp, user-specific signals for accuracy. Learnable fusion weights $G^{d\wedge p}$7, $G^{d\wedge p}$8, and $G^{d\wedge p}$9 let the model trade off accuracy, represented by SGC + PER + PRG, versus diversity, represented by CNA + PENR.

The training and inference outline is fixed. The procedure first precomputes PER weights $G^{Co}$0 and PENR weights $G^{Co}$1 and $G^{Co}$2. It then builds strict graphs and $G^{Co}$3 and computes $G^{Co}$4 and $G^{Co}$5 via LightGCN plus attention. The LLM is prompted once per game to obtain $G^{Co}$6, which is embedded to $G^{Co}$7 and fused. For each player $G^{Co}$8, the LLM is prompted on the interaction history to obtain $G^{Co}$9, which is embedded to $G^{bi}$0 and fused. During training epochs, the model samples $G^{bi}$1, computes $G^{bi}$2 and the scores $G^{bi}$3, accumulates $G^{bi}$4, and backpropagates to update embeddings, MLPs, fusion weights, and the NSR parameter $G^{bi}$5. The stopping condition is when validation performance plateaus. At inference time, for a given player $G^{bi}$6, the system ranks all $G^{bi}$7 by $G^{bi}$8 and returns the top-$G^{bi}$9.

5. Experimental configuration and reported performance

The evaluation is conducted on two Steam datasets. Steam I contains 3,908,744 players, 2,675 games, and 95M interactions, with an 80/10/10 train/valid/test split. Steam II contains 334,730 players, 13,047 games, and 3.7M interactions, with a 50/5/45 split. Both datasets contain dwelling time, ratings, and genre/developer/publisher information.

The baselines are partitioned by emphasis. Accuracy-focused baselines are LightGCN, SURGE, BIGCF, MVGNN, and SCGRec. Diversity-focused baselines are MMR, EDUA, DDGraph, DGCN, and DGRec. Trade-off baselines are EXPLORE and the original CPGRec. The evaluation metrics are Accuracy@K—Recall@K, Precision@K, Hit@K, and NDCG@K for $(u,i)$0—and Diversity@K—Coverage@K by genre, developer, publisher, and total, Entropy@K, Conventional Coverage@K, Tail Coverage@K, and Tail@K as the fraction of long-tail items.

On Steam I at $(u,i)$1, the reported results are explicit. LightGCN obtains NDCG@10 $(u,i)$2, Recall@10 $(u,i)$3, Prec@10 $(u,i)$4, ConvCov@10 $(u,i)$5, and Tail@10 $(u,i)$6. SCGRec obtains $(u,i)$7, $(u,i)$8, $(u,i)$9, $I$00, and $I$01. CPGRec (acc-only) obtains $I$02, $I$03, $I$04, $I$05, and $I$06. CPGRec + (acc-only) obtains $I$07, $I$08, $I$09, $I$10, and $I$11. CPGRec (trade-off) obtains $I$12, $I$13, $I$14, $I$15, and $I$16. CPGRec + (trade-off) obtains $I$17, $I$18, $I$19, $I$20, and $I$21. The paper adds that similar gains appear on Steam II, and in diversity-only and hybrid settings (Li et al., 16 Apr 2026).

The reported hyperparameters are embedding size $I$22, GNN layers $I$23, learning rate $I$24, batch size $I$25, $I$26, $I$27, and $I$28 for PER. These values define the specific experimental instance rather than a claimed universal optimum.

6. Limitations, interpretation, and future directions

The stated strengths are threefold: CPGRec+ addresses GCN over-smoothing via signed, info-theoretic edge reweighting; it exploits LLM reasoning to inject global and personal preference semantics; and it achieves state-of-the-art on both accuracy and diversity metrics with a single unified framework. These are the terms in which the method is positioned relative to earlier GNN-based recommenders.

The limitations are also explicit. Dependence on LLM prompts introduces external cost and potential hallucination. The static PER threshold $I$29 may require per-domain tuning. Cold-start players and games are not explicitly addressed. These constraints are significant because they delimit the scope of the reported gains: the framework is validated on interaction-rich Steam data with LLM assistance, not on a cold-start or prompt-free regime.

The proposed future directions are dynamic graph updates to capture evolving tastes, uncertainty quantification for LLM-generated descriptions, cross-modal PRG using game trailers or audio, and automatic $I$30 scheduling via meta-learning. This suggests a research trajectory in which preference-aware graph weighting, diversity-aware ranking, and LLM-assisted semantic augmentation are integrated more tightly, especially under temporal, multimodal, and uncertainty-aware settings.