Constructing Recommender Systems for Effective Health Messages: Smoking Cessation

描述(由申请人提供)：成功的公共卫生运动在很大程度上取决于信息传播的有效性。为公众健康设计有效的信息既是一门艺术，也是一门科学，艺术占主导地位，因为科学产生的知识积累太慢，理论指导不足。这里提出的研究放弃了消息设计的标准实验方法，放弃了理论开发，转而支持以商业系统为模型的“推荐机”的开发。这种方法将允许基于内容相似性、偏好相似性或它们的组合的算法，根据个人偏好定制消息推荐。推荐系统本质上是在密集数据上运行的派生算法，这些数据既涉及消息的偏好(吸烟者的评级)，也涉及客观的消息特征(内容)。他们的目标是预测用户对以前未见过的消息的评分。传统的消息研究方法推进消息设计的科学进展太慢，受不充分的理论驱动，并且需要受众特征、消息特征和目标行为之间非常复杂的因素交互作用。健康信息推荐机的开发将基于大量信息档案、吸烟者的密集偏好数据以及对信息客观特征的广泛(且大多是自动化的)评估。结果将提供一个从大型档案中选择有效邮件的程序，该档案将为特定目标人员量身定做。与定制研究不同，不会对需要确定哪些受众特征来限制消息选择做出先验假设。推荐系统有可能改变关于有效消息的研究。其成果将包括(1)有效(戒烟)信息的偏好算法；(2)信息设计方法的飞跃，避免单调乏味的逐个功能实验；(3)采用任何人都熟悉的方法，在亚马逊上购买书籍或通过Netflix选择电影；(4)为自动用户友好型推荐系统奠定基础。增加健康和降低风险的行为的信息选择过程将发生根本性的变化。还将启用使用移动技术和个性化保健网站等新媒体的应用程序。这项研究建议：(1)准备现有数据，用于对推荐系统进行预测试；(2)使用商业领域的最先进程序，开发协作和内容方法的混合推荐算法；(3)在吸烟者样本中测试混合算法，将推荐信息的偏好与两种比较条件进行比较；(4)跟进，以确定那些通过推荐算法收到建议的信息与收到随机选择或“最偏好”集合的信息之间的戒烟意图是否存在差异。 公共卫生相关性：对公共卫生的影响寻求改变行为的公共卫生运动之所以成功，在很大程度上是因为所发布的信息本身是有效的，因为它们与目标受众和--如果可能的话--目标个人协调一致。在健康行为改变中设计有效信息的传统方法--主要是实验方法和因素方法--在经验和理论上进展太慢，使得信息设计没有原则和直觉。这项研究从根本上改变了信息设计和选择的方法，开发了基于商业产品推荐模型的经验性“推荐算法”，并在大量戒烟信息档案上进行了应用和测试。

DESCRIPTION (provided by applicant): Successful public health campaigns depend in large measure on how effectively information is communicated. Designing effective messages for the public's health is both an art and a science with the art dominating because knowledge generated by the science is accumulating too slowly and with insufficient theoretical guidance. The research proposed here abandons standard experimental approaches to message design and abandons theory development in favor of the development of a "recommendation machine" modeled after commercial systems. This approach will allow message recommendations tailored to individual preferences based on algorithms for content similarity, preference similarity or their combination. Recommendation systems are essentially derived algorithms operating on dense data involving both preferences for messages (ratings by smokers) and objective message features (content). Their goal is to predict a user's ratings for messages not previously seen by the user. Conventional approaches to message research advance the science of message design too slowly, are driven by inadequate theory, and require very complex factorial interactions among audience characteristics, message features and the target behavior. The development of a recommendation machine for health messages will operate on a large archive of messages, dense preference data from smokers, and extensive (and mostly automated) assessment of the objective features of messages. The results will provide a procedure for the selection of effective messages from a large archive that will be tailored to a specific target person. Unlike tailoring research, no a priori assumptions will be made about which audience characteristics would need to be identified to constrain message selection. Recommendation systems have the potential to transform research about effective messages. The outcomes would include (1) an algorithm for preferences for effective (smoking cessation) messages; (2) a leap beyond approaches to message design side-stepping the tedious work in one-feature-at-a-time experiments; (3) an approach employing methods familiar to anyone ever having bought a book on Amazon or selected a movie via Netflix; (4) setting the stage for automatic user friendly recommender systems. Message selection processes for behaviors to increase health and lower risk would change radically. Applications using new media such as mobile technologies and personalized health web sites would be enabled as well. The research proposed: (1) prepares existing data to use in pretesting recommendation systems; (2) develops recommendation algorithms that are hybrids of collaborative and content approaches using state-of- the-art procedures from the commercial arena; (3) tests hybrid algorithms in a sample of smokers comparing the preferences for recommended messages to two comparison conditions; (4) follows up to determine whether differences in smoking cessation intentions differ between those receiving messages suggested via the recommender algorithms vs. those receiving a random selection or a "most preferred" set. PUBLIC HEALTH RELEVANCE: Impact on Public Health Public health campaigns that seek to change behavior are successful in large part when the messages deployed are themselves effective by being attuned to the target audience and -- when possible -- the target individual. Conventional approaches to the design of effective messages in health behavior change -- primarily experimental and factorial in approach -- have moved too slowly both empirically and theoretically leaving message design unprincipled and intuitive. The research proposed radically alters the approach to message design and selection by developing empirically based "recommendation algorithms" based on a model of commercial product recommendation and applied and tested on a large archive of smoking cessation messages.