Text Similarity Estimation for Targeted Marketing with Outlier Robust Centroids of GloVe Word Embeddings

摘要

Customer segmentation is an important task for marketeers. It is a prerequisite for precise and successful marketing campaigns. The traditional way of conducting it is by clustering based on demographic, geographic and psychographic variables like sex, age, city, or profession. Such an approach has several drawbacks. First, some of these variables might be hard to obtain in practice. Second, deducing from them actual interests for certain products is very hard in practice. In this paper, we present a different approach, in which we use short text snippets provided by users in an online contest to come up with a much more precise user interest profile. In particular, these text snippets are matched to keyword lists representing several marketing target groups like Freestyle Action Sportsmen, Young Performer, etc. For that, we employed the cosine measure on outlier robust centroids of GloVe word embeddings. These centroids are determined in an iterative fashion that gives most focus on non-outlier vectors and tends to disregard vectors, which are far off from the others. The evaluation showed that we obtained superior results with our method than several baseline approaches including one alternative method of noise reduction based on tf-idf weights.