复现arxiv论文实现Facebook数据Big5作者分类效果差，求解论文方法含义

Hey there! I get why you're frustrated—simple pooling of GloVe vectors (mean/min/max) often loses too much nuance for personality classification, especially when compared to the paper's results. Let's unpack exactly what that "combines Word Embedding with Gaussian Process" line means, because it's not about feeding pooled vectors into a GP/SVM like you're doing now.

The Core Idea: Embeddings in the Kernel, Not Just as Input Features #

The paper doesn't treat aggregated word vectors as fixed input features for GP/SVM. Instead, it integrates GloVe embeddings directly into the Gaussian Process's kernel function—this is the key difference from your current approach. Here's how it works in practice:

• Text similarity via word embeddings: Instead of compressing each user's posts into a single vector first, the GP calculates the similarity between two users' text sequences using their word embeddings. For example, for two sets of posts, it might compute the similarity by looking at all pairs of words across the two users' texts, using their GloVe vector cosine similarity as a weight, then aggregating those weights to get a single "similarity score" (the kernel value) for the GP.
• Custom kernel design: The paper likely uses a modified string kernel (like a weighted matching kernel) that leverages GloVe to assign more weight to semantically similar word pairs, rather than just exact word matches. This lets the GP capture contextual and semantic similarities between users' writing styles that simple pooling misses.

Why Your Current Approach Falls Short #

Mean/min/max pooling throws away critical information like:

• Word order and context (e.g., "I love my job" vs "My job loves me" have the same mean vector but opposite meanings)
• Frequency of impactful words (a single strong personality indicator word gets diluted in a mean pool)
• Semantic relationships between words that aren't captured by simple aggregation

Steps to Align with the Paper's Method #

1. Dig into the paper's kernel section: Head to the Methodology section and look for the exact kernel formula. It should outline how they combine GloVe embeddings with the GP's similarity calculation—look for terms that involve word vector dot products or cosine similarities across text pairs.
2. Implement the custom kernel: Instead of feeding pooled vectors into a standard GP kernel (like RBF), build a kernel function that takes two text sequences, computes pairwise word embedding similarities, and aggregates them into a single kernel value. For example, you could sum the cosine similarities of all relevant word pairs, or use a weighted sum where frequent, low-information words get lower weights.
3. Test intermediate steps: If building the custom kernel feels overwhelming, first try using a more sophisticated text encoding method (like Sentence-BERT or Doc2Vec) to generate fixed-length user-level vectors—this won't replicate the paper's method exactly, but it will give you a better baseline than manual GloVe pooling.

内容的提问来源于stack exchange，提问作者qwazy