Bi-Intentional Modeling and Knowledge Graph Diffusion for Rice Variety Selection and Breeding Recommendation

[Objective] Selection of rice varieties requires consideration of several factors, such as yield, fertility, disease resistance and resistance to downfall. In order to meet the user's rice variety selection needs, help users quickly access to the rice varieties they need, improve efficiency, and further promote the informatization and intelligence of rice breeding work, the bi-intentional modeling and knowledge graph diffusion model, an advanced method was proposed. [Methods] The research work was mainly carried out at two levels: Data and methodology. At the data level, considering the current lack of relevant data support for rice variety selection and breeding recommendation, a certain amount of recommendation dataset was constructed. The rice variety selection recommendation dataset consisted of two parts: Interaction data and knowledge graph. For the interaction data, the rice varieties that had been planted in the region were collected on a region-by-region basis, and then a batch of users was simulated and generated from the region. The corresponding rice varieties were assigned to the generated users according to the random sampling method to construct the user-item interaction data. For the knowledge graph, detailed text descriptions of rice varieties were first collected, and then information was extracted from them to construct data in ternary format from multiple varietal characteristics, such as selection unit, varietal category, disease resistance, and cold tolerance. At the methodological level, a model of bi-intentional modeling and knowledge graph diffusion (BMKGD) was proposed. The intent factor in the interaction behavior and the denoising process of the knowledge graph were both taken into account by the BMKGD model. Intentions were usually considered from two perspectives: individual independence and conformity. A dual intent space was chosen to be built by the model to represent both perspectives. For the problem of noisy data in the knowledge graph, denoising was carried out by combining the idea of the diffusion model. Random noise was introduced to destroy the original structure when the knowledge graph was initialized, and the original structure was restored through iterative learning. The denoising was completed in this process. After that, cross-view contrastive learning was carried out in both views. [Results and Discussions] The method proposed achieved optimal performance in the rice variety selection dataset, with recall and normalized discounted cumulative gain (NDCG) values improved by 2.9% and 3.7% compared to the suboptimal model. The performance improvement validated the effectiveness of the method to some extent, indicating that the BMKGD model was more suitable for rice variety recommendation. The Recall value of the BMKGD model on the rice variety selection dataset was 0.327 6, meeting the basic requirements of the recommendation system. The analysis revealed that the collaborative signals in the interaction data played a major role, while the quality of the constructed knowledge graph still had some room for improvement. The module variants with key components removed all exhibited a decrease in performance compared to the original model, which validated the effectiveness of the modules. The performance degradation of the model variants with each component removed varied, indicating that different components played different roles. The performance drop of the model variant with the cross-view contrastive learning module removed was small, indicating that there was some room for improvement in the module to fully utilize the collaborative relationship between the two views. Conclusions The BMKGD model proposed in this paper achieves good performance on the rice variety selection dataset and accomplishes the recommendation task well. It shows that the model can be used to support the rice variety selection and breeding work and help users to select suitable rice varieties.