社会与工程心理学研究室知识库

    配偶决策行为是一种典型的多属性决策行为，人们在进行多属性决策时遵循不同的决策策略。尽管不同的决策策略对决策过程有着不同的假设，大部分决策策略都认为特征权重在决策中起着至关重要的作用。因此，预测决策行为的前提是测量决策权重。
    决策者在决策时对不同重要程度的特征给予的注意分配是不同的，而注视的多少恰恰反映了决策者对不同特征的注意分配，因此可以利用决策者对各个特征注视的多少测量特征的决策权重。基于此，研究1发展了一种新的权重测量方法一一眼动赋权法，通过三个实验检验了眼动赋权法的有效性。结果发现:C1)眼动赋权法测得的权重与几种主观赋权法测得的权重高相关，具有良好的相容效度;C2)眼动赋权法应用于预测配偶决策行为具有较高的预测效度，它的预测准确性优于或等于几种主观赋权法的预测准确性，也优于用选项的注视时长预测行为的准确性。(3)眼动赋权法具有实时性的优势，能够测量决策者对各个特征的实时决策权重，依据眼动赋权法测得的实时决策权重还可以对择偶行为进行实时性、前瞻性的预测。
    研究1对决策行为的预测建立在“所有决策者都采用同一种决策策略”的前提下，但决策者在实际决策中会采用不同的决策策略。因此，只有区分出决策者使用了什么决策策略才能够使用相应的决策策略对决策行为进行预测。
    决策者使用不同的决策策略时的信息搜索与加工模式是不同的，而眼跳、注视顺序、扫视轨迹等指标恰恰反映了决策者的信息搜索与加工模式，因此可以利用这些指标区分决策者采用了怎样的决策策略。研究2包含两个实验，实验设计类似:要求被试完成一个自由选择任务和两个强迫规则任务(齐当别强迫规则任务、算总分强迫规则任务)。基于两个强迫规则任务的眼动模式，我们探索了两种区分决策策略的分类方法，结果发现:C1)将SM值和注视点平均注视时长作为分类指标的逻辑回归混合模型能够有效区分算总分策略和齐当别策略，其分类准确性在2特征择偶任务(实验2a)和3特征择偶任务(实验2b)中分别高达89.9%和98.0%。通过这种策略区分方法对自由选择任务进行区分，发现决策者大部分时候会采用类似齐当别规则的决策策略。C2)比较策者在决策时的扫视轨迹与两种强迫规则任务的扫视轨迹相似性程度能够有效区分决策者采用了哪种策略区分，其分类准确性在两个实验中分别高达88.0%和91.5%，使用这种策略区分方法同样发现，决策者在进行配偶决策时会更多采用类似齐当别规则的决策策略。
    研究1实现了用眼动数据实时捕捉决策者对特征的权重变化情况，研究2实现了用眼动数据区分决策者策略改变的情况，基于前两个研究的结果，研究3探索了哪些因素能够有效预测决策者在两次重复选择时是否发生选择反转，结果发现，C1)决策者进行自由选择时，决策策略的变化能够显著预测选择反转的发生;C2)决策者采用齐当别策略时，特征权重的变化能够显著预测选择反转的发生;(3不论决策者进行自由选择，还是进行强迫规则的选择，两个选项的总价值差异或维度差异效用都能显著预测选择反转的发生。
    以上三个研究从决策过程的视角出发，创新性地提出了眼动赋权法这种实时权重测量新技术，并发展了两种眼动区分决策策略的分类方法。通过眼动技术追踪到“决策是如何做出的”这一决策过程，实现了实时预测决策行为的目的。

  Mate choice is one of the most important choices for all individuals. It is a typical multi-attribute choice, in which people will use various kinds of choice strategies. These different multi-attribute strategies describe the decision process with different hypotheses.
    Eye-movement data can reveal how attention are distributed and how information is processed during decision making process. On one hand, decision makers give different attention distribution to different attributes according to their attribute importance. And gaze distribution just reflects attention distribution.Therefore, we can use the gaze distribution data to measure the importance of different attribute, that is attribute weight. On the other hand, different decision strategies have different hypothese on how information are searched and processed.And some eye-movement data, like transitions, the order of fixation, or scanpath just reflect how decision maker search and process information directly. Therefore, we can use eye-movement data to classify what strategies decision makers are using.
    In Study 1，we developed a new weighting method一the eye-movement weighting method. Three experiments were used to examine the validity of the eye-movement weighting method and found that: (1) the eye-movement weighting methods has a high congruent validity. The attribute weights measured by eye-movement weighting method are highly correlated with the attribute weights measured by traditional subjective weighting method; (2) the eye-movement weighting methods has a high predictive validity. When using the eye-movement weighting methods to predict mate choice, the prediction accuracy is higher than or equal to the prediction accuracy using traditional weighting methods. (3) the biggest advantage of eye-movement weighting method is that it can measure real-time decision weight. Therefore, we can using the real-time weighting methods to realize the real-time prediction.
    Study 2 aim to using eye-movement data to classify what strategies decision maker are using. In this study, participants firstly finished a free mate choice task.And then they finished two imposed rule tasks: one imposed rule task asked participants to follow a classic "rational" strategies (weighted additive rule) to make mate choice, and the other one imposed rule task asked them to follow a heuristic strategies (Equate-to-differentiate rule) to make mate choice. We explored two different kinds of methods to classify decision strategy: (1) based on knowledge of Machine Learning, we firstly trained the eye-movement data recorded from the two imposed tasks, and then used the trained classifier to classify what strategies were used in the free choice task. Results showed that a Mixed Logistic Regression model with SM index and mean fixation duration as features inputs performed the best, the classifier reached 89.95% (Study 2a) and 98.4% (Study 2b) accuracy in distinguishing whether people are using the heuristic strategies or the classic "rational" strategies to make mate choice. Finally, we used this Mixed Logistic Regression model to classify trails in the free mate choice task, and found that most of the time, decision maker were using a non-compensatory, attribute-based heuristic strategies, similar to equate-to-differentiate rule; (2) we classify decision strategy based on the similarity scores of scanpath between trials in different task and found this classification methods can reach 88% (Study 2a) and 91.53% (Study 2b) accuracy in distinguish whether people are using the heuristic strategies or the classic "rational" strategies to make mate choice.
    Study 3 explored why people make choice reversed in repeated mate choice questions. We found that if decision makers use different strategies in repeated choice questions, they are more likely to choose differently; if decision makers change the decision attribute weight, they are more likely to choose differently; and if decision makers judge two options are similar or two attributes' differences are equal, they are more likely to make choice randomly and reversed.
    This study is innovative in methods in three aspects: Firstly, we developed a new weighting method which can measure real-time decision weight. This weighting method will be useful in human-computer interaction, artificial intelligence, and virtual reality areas. Secondly, the methods which use eye-movement data to classify decision strategies are helpful in future studies on decision making. Most importantly,this research provides a process-based prediction method to predict mate choice. This process-based predition may help researchers achieve the "mind reading" goal eventually.