Semantics-aware Post-hoc Explainability on Social Recommender Systems

社交推荐系统的一个基本权衡是准确性和可解释性。为了减轻这种权衡，本研究计划的目标是开发事后解释模型，该模型将受益于知识图（如DBpedia， Freebase或YAGO）中表示的结构化和语义知识，以系统地增强社交媒体上推荐系统的可解释性，同时保持其预测准确性。事后解释模型证明了推荐系统先前给用户的推荐是正确的。基于知识图的事后解释模型将系统地研究如何利用推荐系统的领域知识来提高解释模型在近似推荐时的可信度。本研究有三个主要的新颖方面：(1)我们将开发知识图修剪技术，使推荐系统中的项目与相关的知识图子图相关联，以便高效和有效地为用户生成推荐项目的事后解释；(2)结合知识图谱子图，开发基于用户在社交媒体上的日常活动为用户生成个性化事后解释的机制；(3)我们还将探索不同的技术，通过将解释与知识图联系起来来丰富解释，提高人类对生成的事后解释的可理解性。这项研究计划的主要成果将是开发下一代的事后解释模型，这些模型可以被不断增长的社交媒体分析部门采用，以有效地响应人类理解模型为什么做出特定决定的自然需求。研究结果可用于特定领域的社交推荐系统，如工作推荐、音乐推荐和在线营销解决方案，以提高用户对推荐的接受度或满意度，增加信任，从而为社交媒体的可持续性奠定基础。该研究项目将为培养3名博士和6名硕士生提供坚实的基础，这些硕士生将在信息检索、社交媒体挖掘和语义网络技术等领域获得专业知识。预计这些高素质的人才能够立即被加拿大蓬勃发展的数据科学和分析行业所吸收。

One of the fundamental trade-offs in social recommender systems is that of accuracy and explainability. To mitigate this trade-off, the objective of this research program is to develop post-hoc explanation models that would benefit from the structured and semantic knowledge represented in knowledge graphs, such as DBpedia, Freebase or YAGO, to systematically enhance the explainability of recommender systems on social media, whist maintaining their predictive accuracy. A post-hoc explanation model justifies the recommendations previously given to a user by a recommender system. A knowledge graph-based post-hoc explanation model would systematically investigate how to employ domain knowledge of a recommender system to increase the faithfulness of the explanation model in approximating the recommendations. There are three main novel aspects of this research: (1) we will develop techniques for pruning knowledge graphs such that items in the recommender system are linked to relevant knowledge graph subgraphs in order to efficiently and effectively generate post-hoc explanations for the recommended items to the users; (2) we will develop mechanisms for generating personalized post-hoc explanations for users based on their daily activities on social media integrated with the knowledge graph subgraphs; (3) we will also explore different techniques for enriching the explanations via linking them to knowledge graphs, improving the human-understandability of the generated post-hoc explanations. The primary outcome of this research program will be development of the next generation of post-hoc explanation models that can be adopted by the growing social media analytics sector to effectively respond to the natural demand of human-beings for understanding why a model has made a specific decision. The outcomes can be utilized by domain specific social recommender systems, such as job recommendation, music recommendation and online marketing solutions, to improve the users' acceptance or satisfaction of recommendations and increase trust, which consequently, sets the stage for the sustainability of social media. The proposed research program will also provide a solid foundation for the training of highly qualified personnel including 3 PhD and 6 MASc students who will gain expertise in the fields of information retrieval, social media mining, and semantic web technologies. It is expected that these highly qualified personnel are able to be immediately absorbed in the lively and growing data science and analytics industry in Canada.