From Sun Ying’s doctoral dissertation at the Institute of Computing Technology, Chinese Academy of Sciences, selected for the preliminary evaluation list of the 2023 CCF Doctoral Dissertation Incentive Program!
https://www.ccf.org.cn/Focus/2023-11-29/798503.shtml
Talent refers to individuals with certain professional knowledge or specialized skills, who engage in creative labor and contribute to society. Under the strategy of strengthening the nation with talent, efficient talent management helps to fully utilize this important resource. To achieve precise talent management, it is necessary to evaluate talent characteristics with skills at the core and formulate corresponding strategies for selection, use, training, and retention. Compared to general human resource analysis tasks, talent assessment has higher dynamic, fine-grained, and personalized requirements. Although some research on human resource analysis algorithms based on classical data mining methods such as tree classification and clustering has emerged in recent years, their flexibility and expressiveness are poor, making it difficult to model the differentiated characteristics of talent effectively. The expressive power and adaptability of neural networks provide richer possibilities for talent assessment, but their lack of interpretability makes it difficult for them to be widely applied in talent assessment scenarios. Although there has been research on explainable neural networks in recent years, which symbolically represent key information at different stages of neural networks such as input, transformation, and output, to provide intuitive evidence for users’ understanding, most existing studies are retrospective explanations of algorithms, which cannot accurately reflect the true modeling process of neural networks. To meet the higher requirements for interpretability in practical expert decision-making scenarios such as talent assessment and better support talent management decisions, there is an urgent need to improve the interpretability of neural network models themselves.
This article focuses on the talent assessment tasks of selection, retention, training, and use, conducting research on explainable neural network algorithms from four levels: input contribution, concept transformation, decision objectives, and relational impact. It proposes effective neural network algorithms that ensure model accuracy while meeting the interpretability requirements of different tasks, achieving intuitive symbolic modeling of judgment basis at different levels to support user understanding, and providing a comprehensive talent assessment solution. Specifically, first, for the talent selection stage, an explainable dominance feature value network is proposed from the perspective of global feature contribution for skill value assessment; then, for the talent attraction stage, an explainable incremental feature contribution network is proposed from the perspective of local feature contribution for talent compensation assessment; third, addressing the issue of heterogeneous variable information extraction in practical scenarios such as talent management, an explainable voting mechanism concept inference network is proposed from the perspective of concept transformation; fourth, for the talent training stage, an explainable multi-objective dual Q network is proposed from the perspective of decision objectives for skill learning assessment; fifth, for the talent utilization stage, an explainable tree-structured relational attention network is proposed from the perspective of relational impact for assessing organizational matching of talent. The main results obtained are as follows:
1. An explainable dominance feature value network and skill value assessment algorithm are proposed. For the talent selection stage, a skill value assessment solution is proposed that can model the relationship between skills and compensation in complex work scenarios, finely quantifying the global contribution of different job skills to compensation in the market, aiding in identifying talents with key skills. In this problem, skill value can be viewed as a global explanation of the contribution of input features to the overall task output. To this end, a feature element set utility modeling method under the influence of edge information is proposed for general prediction problems. Based on this, an explainable dominance feature value neural network is proposed from the perspective of global feature contribution, which fits the global independent contribution (feature value) of each feature under the influence of edge information based on a dynamic depth-width network, and fits the dominance of features in combinations using graph neural networks and attention mechanisms, combining feature values into prediction targets in an explainable and highly expressive linear manner. This model can provide intuitive explanations of the global contribution of input to output while making accurate result predictions. Experimental results on large-scale real compensation data show that the proposed model can provide meaningful value assessments for different skills in various work scenarios and achieves significantly better results than state-of-the-art models in compensation prediction tasks, proving the advantages of the proposed model in interpretability and expressiveness compared to existing models.
2. An explainable incremental feature contribution network and job compensation assessment algorithm are proposed. For the talent attraction stage, a talent compensation prediction solution is proposed that can accurately predict the compensation corresponding to different skill sets in complex work scenarios and clarify the contribution of each skill. In this problem, the compensation contribution of each skill to a talent can be viewed as a local explanation of the contribution of input features to output in the sample neighborhood. To this end, from the perspective of local feature contribution, an explainable incremental feature contribution neural network is proposed, which retains the feature element set utility modeling problem under the influence of edge information and flexibly expresses the set utility as the sum of feature marginal contributions. Notably, unlike existing set modeling methods that achieve permutation invariance through inductive bias, a sequentially sensitive incremental set utility network is proposed that models the sequentially sensitive marginal contributions and utilizes permutation invariance to achieve data augmentation, effectively improving the model’s generalization and robustness. This model can provide intuitive explanations of the local contribution of input to output while making accurate result predictions. Experimental results on large-scale real compensation data show that the proposed model achieves over 30% improvement in skill set compensation prediction tasks while effectively assessing skill contributions and discovering the impact of skills on each other’s contributions.
3. An explainable voting mechanism concept inference network is proposed. Addressing the multi-source heterogeneous variable information extraction problem in talent assessment, an explainable voting mechanism concept inference neural network is proposed from the perspective of concept transformation. Unlike existing neural network models that embed intermediate information into large-scale hidden units, this model explicitly symbolizes a small number of hidden concepts and models hierarchical transformations between concepts through univariate nonlinear voting transformations. Thus, while achieving intuitive concept transformations, the model’s expressive capability is ensured. Through theoretical analysis, this study proves the superiority of the proposed model in interpretability compared to existing neural networks and explains the roles of different units in the model, accordingly designing simple and efficient auxiliary algorithms for visualization explanations, judgment process analysis, network pruning, etc. Experimental results on multiple large-scale public datasets show that the proposed model has significant advantages in interpretability and accuracy compared to existing neural network methods. Its classification accuracy significantly outperforms state-of-the-art models, and the modeling process is completely transparent, allowing users to intuitively understand the intermediate concept transformations and discover meaningful hidden patterns.
4. An explainable multi-objective dual Q network and skill learning assessment algorithm are proposed. For the talent training stage, a skill learning planning solution is proposed that models the utility of the skill learning process from multiple perspectives, achieving personalized learning path recommendations. Since this task is a type of decision-making task focusing on the explanation of the results produced by decisions, it is formalized from the perspective of decision objectives as a multi-objective set expansion decision problem, establishing an efficient and explainable reward feedback simulation environment, and proposing a multi-objective dual Q neural network for dynamic evaluation of actions’ long-term and short-term utilities and associations across different objectives, which can provide multi-faceted decision basis explanations while learning action strategies. In particular, addressing the enormous discrete action space in the problem, a candidate pool-based training strategy is proposed to enhance the training efficiency of the multi-objective utility model. Experimental results on real talent datasets show that the proposed model has significant advantages over existing skill recommendation methods in terms of returns, costs, and sustainability, and can explain the short- and long-term learning benefits and difficulties of skills from multiple decision basis perspectives.
5. An explainable tree-structured relational attention network and organizational matching assessment algorithm are proposed. For the talent utilization stage, a talent-organization matching solution is proposed that conducts explainable modeling of the talent organizational environment, analyzing the matching of talent with the organization and its impact on talent development, and explaining the role of organizational relationships in matching. In this problem, to achieve explainable modeling of tree-structured hierarchical structures such as organizational structure, an explainable tree-structured relational attention network is proposed from the perspective of relational impact, which can identify the relative relationship impacts at different granularities on the tree structure and aggregate explainable intra- and inter-relational impacts. Notably, to enhance the modeling efficiency of relational impacts on tree structures, an efficient algorithm based on sparse computation is proposed to identify and aggregate relationships with linear time complexity. Experimental results on large-scale talent management data show that the proposed model significantly outperforms state-of-the-art models in various talent work outcome prediction tasks while explaining the key factors influencing talent organizational matching, providing insights for talent management.
Convenient access to knowledge
Convenient download, please followZhuanzhi public account (click the above blue Zhuanzhi to follow)
Zhuanzhi, a professional and reliable AI knowledge distribution platform, enabling cognitive collaboration faster and better! Welcome to register and log in to Zhuanzhi www.zhuanzhi.ai to access 100,000+ AI(AI in military, medicine, public security, etc.) themed knowledge resources!
Scan the QR code to join Zhuanzhi AI Knowledge Star Group, to get the latest AI professional knowledge tutorials and exchange consultations with experts!
