Dual-Triangular Architecture DTRec Boosts Multi-Task Recommendation Accuracy Across Sparse Platforms

Computer scientists have introduced a recommendation framework known as DTRec to address dual-triangular recommendation environments, demonstrating that uniting pre-trained large language models and structured knowledge graph encoders within a shared multi-task learning architecture overcomes severe data sparsity across reciprocal matching networks. Traditional recommendation algorithms predominantly focus on bipartite user-item interactions, projecting consumer histories and catalog entries into shared vector spaces to predict engagement with single products, videos, or musical tracks. However, modern operational environments often involve reciprocal exchanges governed by four distinct participant categories: primary demand-side actors, primary supply-side actors, demanded items, and supplied items. In domains such as technology commercialization, patent licensing, and talent recruitment, recommendations must operate bidirectionally among all four entities simultaneously. By formally defining the scientific parameters of dual-triangular recommendations and deploying an integrated neural architecture evaluated across three real-world domain datasets, the research team demonstrated that shared parameter learning across six interdependent interaction matrices yields significant performance advantages over traditional single-task and graph-based collaborative baselines.

The structural complexity of dual-triangular recommendation arises from the dense web of interdependencies that connect four separate entity categories within a heterogeneous information network. In the formalization established by the researchers, the four vertices represent suppliers, demanders, supplied items, and demanded items, which in an academic technology commercialization platform correspond directly to researchers, enterprises, technological achievements, and corporate technological demands. Unlike conventional non-triangular setups that solely connect consumers to single items, or triangular systems that connect users, creators, and works, dual-triangular networks incorporate six concurrent interaction matrices. These matrices encompass researchers exploring enterprises, enterprises posting technological demands, researchers reviewing corporate demands, enterprises identifying qualified researchers, researchers publishing research achievements, and enterprises evaluating those achievements. Furthermore, the correspondence across these groups is fundamentally many-to-many, because a single academic achievement may involve multiple collaborating co-authors, an enterprise can articulate several technical problems, and broad industry needs published by trade associations frequently apply to multiple corporate entities, generating complex structural asymmetries.

To model these intricate relationships, the researchers designed DTRec as a four-part system comprising a knowledge graph encoder, a large language model-augmented encoder, a linear projection fusion layer, and a multi-task learning engine. Raw source information describing researchers, corporate entities, technological demands, and academic achievements is inherently verbose, unstructured, and specialized, presenting severe challenges for standard embedding mechanisms. The knowledge graph component processes the descriptive text of achievements and demands by first extracting key semantic tokens through the Jieba segmentation toolkit. These segmented linguistic keywords are organized into entity triples and assembled into a unified knowledge graph. The framework subsequently applies the Structure-Aware Convolutional Network, or SACN, to capture graph topology and semantic dependencies among connected entities. To enrich these representations with contextual attributes, DTRec executes attribute aggregation inspired by AGNN, concatenating the embeddings of adjacent first-order neighbor nodes and mapping them through an interaction operator to produce attribute-enhanced knowledge graph representations constrained by binary cross-entropy and Kullback-Leibler divergence objectives.

Simultaneously, DTRec incorporates a parallel large language model-augmented pathway to capture nuanced contextual semantics that graph structures alone cannot adequately represent. Drawing on structured prompt templates tailored to the attribute configurations of all four entity categories, the model formats extracted entity triples into natural language prompts. These prompts are processed by pre-trained large language models, including commercial models such as ChatGPT 3.5 and open-source models such as Qwen 3, which leverage vast pre-trained linguistic knowledge to generate standardized, semantically enriched entity descriptions without requiring task-specific parameter fine-tuning. The regenerated narratives for researchers, enterprises, achievements, and demands are subsequently encoded into dense vector embeddings using a pre-trained RoBERTa model. Because the knowledge graph encoder and the language model encoder output representations of varying characteristics, DTRec channels all generated vectors through a learned linear transformation layer, projecting them into a shared sixty-four-dimensional latent space where element-wise vector addition combines the knowledge graph and textual embeddings for achievements and demands.

The consolidated entity representations are optimized within a unified multi-task learning architecture designed to overcome the severe data sparsity that characterizes multi-entity commercialization networks. In real-world technology platforms, corporate enterprises interact with academic researchers or inspect complex inventions relatively infrequently, typically seeking external collaboration only when confronting critical, cross-domain engineering obstacles. Consequently, the observed interaction records across the six operational matrices are extraordinarily sparse, with interaction sparsity consistently exceeding 99.7 percent across empirical benchmarks. To extract maximum collaborative signal from these limited observations, the multi-task module shares underlying model parameters across all six recommendation tasks. The global loss function aggregates cross-entropy losses for each matrix, assigning unit weight to the first triangular task set governing researcher-enterprise-demand interactions and applying a scaling coefficient lambda of 0.3 to the second triangular task set governing enterprise-researcher-achievement interactions, enabling reciprocal reinforcement across all four entity types during backpropagation.

The research team evaluated DTRec across three newly curated and expanded multi-entity datasets representing diverse real-world operational environments. The first dataset, SCTCC, originates from the South China Technology Commercialization Center platform and was expanded to include 1,481 technological achievements alongside corresponding interaction logs with researchers and enterprises, encompassing academic papers, patents, prototype systems, and technical documentation. To evaluate model generalizability beyond academic commercialization, the authors constructed the PatTrans dataset by extracting verified patent transfer records within International Patent Classification Section H, covering electricity technologies, from the GBAIPR intellectual property platform across twelve collaborating universities. In this setting, inventors and patents represent the supply side, while enterprises and candidate technical demands represent the demand side. The third benchmark, the JJSE dataset, expanded the talent recruitment domain by incorporating computer science job seekers, resumes, enterprises, and position postings, with interactions generated through simulated candidate applications and recruiter evaluations.

Model performance was systematically measured using two standard validation metrics: the Area Under the Receiver Operating Characteristic Curve, or AUC, reflecting classification discrimination across all operational decision thresholds, and overall Accuracy, or ACC, measuring classification correctness at a single operating point. All experiments were executed on a single hardware workstation equipped with an NVIDIA Tesla A100 graphics processing unit. The researchers benchmarked DTRec against five prominent collaborative filtering and graph neural network baselines, including LightGCN, Light Graph Contrastive Learning, Neural Collaborative Filtering, Self-supervised Graph Learning, and Adaptive Kernel Balancing Doubly Robust based on Gaussian kernels. To ensure rigorous and equitable comparison within the dual-triangular formulation, all baseline models were adapted to operate both under isolated single-task settings based on direct user-item pairings and under the shared multi-task parameter framework developed for the dual-triangular architecture.

The comparative experimental findings demonstrated that DTRec consistently achieved superior recommendation quality across the six interrelated prediction tasks on the SCTCC technology commercialization dataset. While certain specialized baselines demonstrated competitive performance on isolated tasks, their predictive efficacy exhibited noticeable instability across the complete dual-triangular network. For instance, Self-supervised Graph Learning performed admirably on the primary triangular recommendation tasks involving researcher and enterprise matching but suffered observable performance degradation on the subsequent three tasks involving specific achievement interactions. Conversely, Neural Collaborative Filtering exhibited competent performance on secondary triangular interactions while struggling on initial enterprise-researcher pairings. Baseline architectures including LightGCN, LightGCL, and AKBDR-Gau encountered difficulties in maintaining accuracy across the ultra-sparse interaction regimes, confirming that standard collaborative filtering architectures cannot adequately resolve asymmetric multi-entity matching without specialized semantic augmentation and reciprocal multi-task parameter sharing.

Parallel comparative evaluations conducted on the PatTrans and JJSE benchmarks further reinforced the generalizability and robustness of the dual-triangular architecture across differing domain dynamics. On the PatTrans electricity patent dataset, DTRec maintained leading performance across the majority of recommendation relations, although LightGCN attained the highest score on inventor-to-demand matching and AKBDR-Gau led on enterprise-to-inventor recommendations. Similarly, across the simulated JJSE recruitment platform, DTRec demonstrated top-tier predictive capability in five out of six tasks, trailing slightly only in direct job-seeker-to-resume associations where Neural Collaborative Filtering and LightGCL remained highly competitive. The authors noted that the adaptive kernel balancing strategy of AKBDR-Gau allowed it to secure consistent second-place rankings on several sub-tasks by mitigating selection bias, yet it remained surpassed overall by DTRec, which capitalizes on the complementary integration of graph topology, dense text modeling, and joint multi-task optimization.

Comprehensive ablation studies conducted on the SCTCC dataset underscored the relative contribution of each architectural component within DTRec, highlighting multi-task parameter sharing as the single most critical structural element. When the multi-task learning framework was removed in the single-task variant DTRec-S, performance dropped precipitously across all six tasks, with the maximum observed performance improvement of full DTRec reaching 96.64 percent over DTRec-S. Variants lacking the large language model encoder or the knowledge graph encoder also underperformed the complete system, confirming that semantic text expansion and graph attribute aggregation provide essential signals. Parameter sensitivity analyses revealed that model accuracy was highly sensitive to the learning rate, moderately responsive to batch size, and relatively stable across varied loss weighting coefficients. The authors concluded that future research may build upon this dual-triangular formulation by investigating meta-path analysis, attribute-level aggregation, graph clustering, and contrastive learning mechanisms.

Subject of Research: Dual-triangular recommender systems utilizing large language models and knowledge graphs for multi-entity matching

Article Title: Dual-triangular Recommender System

Article References: Lai, P.-Y., Dai, Q.-Y., & Wang, C.-D. (2025). Dual-triangular Recommender System. Data Science and Engineering, 10(4), 800-818. https://doi.org/10.1007/s41019-025-00310-0

Image Credits: AI Generated

DOI: 10.1007/s41019-025-00310-0

Keywords: recommender systems, dual-triangular recommendation, DTRec, large language models, knowledge graph encoder, multi-task learning, heterogeneous information network, technology commercialization, patent transfer, talent recruitment, SCTCC dataset, PatTrans dataset, JJSE dataset, structure-aware convolutional network, RoBERTa, collaborative filtering, data sparsity