Exploratory factor analysis is a widely used framework in the social and behavioral sciences. Since measurement errors are always present in human behavior data, latent factors, generating the observed data, are important to identify. While most factor analysis methods rely on linear relationships in the data-generating process, deep learning models can provide more flexible modeling approaches. However, two problems need to be addressed. First, for interpretation, scaling assumptions are required, which can be (at least) cumbersome for deep generative models. Second, deep generative models are typically not identifiable, which is required in order to identify the underlying latent constructs. We developed a model that uses a variational autoencoder as an estimator for a complex factor analysis model based on importance-weighted variational inference. In order to receive interpretable results and an identified model, we use a linear factor model with identification constraints in the measurement model. To maintain the flexibility of the model, we use normalizing flow latent priors. Within the evaluation of performance measures in a talent development program in soccer, we found more clarity in the separation of the identified underlying latent dimensions with our models compared to traditional PCA analyses.

中文翻译：

---

基于重要性加权变分推理和归一化流先验的深度学习因子分析模型：在青年精英足球运动员的一组多维表现评估中进行评估

探索性因素分析是社会和行为科学中广泛使用的框架。由于人类行为数据中始终存在测量误差，因此识别生成观察数据的潜在因素非常重要。虽然大多数因子分析方法依赖于数据生成过程中的线性关系，但深度学习模型可以提供更灵活的建模方法。然而，有两个问题需要解决。首先，为了解释，需要缩放假设，这对于深度生成模型来说（至少）可能很麻烦。其次，深层生成模型通常是不可识别的，这是识别潜在潜在结构所必需的。我们开发了一个模型，该模型使用变分自动编码器作为基于重要性加权变分推理的复杂因子分析模型的估计器。为了获得可解释的结果和识别的模型，我们在测量模型中使用带有识别约束的线性因子模型。为了保持模型的灵活性，我们使用归一化流潜在先验。在评估足球人才发展计划的绩效指标时，我们发现与传统的 PCA 分析相比，我们的模型可以更清晰地分离已识别的潜在维度。我们使用归一化流潜在先验。在评估足球人才发展计划的绩效指标时，我们发现与传统的 PCA 分析相比，我们的模型可以更清晰地分离已识别的潜在维度。我们使用归一化流潜在先验。在评估足球人才发展计划的绩效指标时，我们发现与传统的 PCA 分析相比，我们的模型可以更清晰地分离已识别的潜在维度。