基于LDA的互联网广告点击率预测研究

广告点击率是互联网广告投放的重要依据,有效地预测广告的点击率,对于提高广告投放的效率有着至关重要的作用。在训练点击率预测模型的过程中,往往面临着广告及用户的数量巨大以及训练数据集稀疏的问题,从而导致点击率预测的准确度下降。针对这些问题提出了一种基于LDA (Latent Dirichlet Allocation)的点击率预测算法,即LDA-FMs,该算法对原有训练集进行基于主题的分割,利用分割后的子训练集分别建立不同主题下的点击率预测模型,在此基础上,利用广告属于不同主题的概率,有权重的结合每个预测模型的预测结果,进而计算广告的点击率。实验基于KDD Cup 2012-Track2的真实数据集,证明了算法的可行性与有效性。

Research on click-through rate prediction of Internet advertising based on LDA

Advertisement click-through rate is essential for Internet advertising. Therefore, estimating click-through rate precisely makes significant difference in the efficiency of advertising on the Internet. During the training of predicting models, many problems will arise such as the massive scale of advertisements and users, and the sparseness of training set, which usually lead to a low accuracy of the predictive click-through rate. In order to solve these problems, this paper proposes an algorithm named LDA-FMs, which is a kind of predicting click rate algorithm based on LDA (Latent Dirichlet Allocation). Specifically, LDA-FMs partitions the original training sets according to different topics, and then builds click-through rate prediction models respectively upon different topics using partitioned sub-training sets. On this basis, the advertisement click-through rate can be calculated by using the probability of advertisement belonging to different topics and prediction results combined with every prediction model. The experiment is based on real data sets from KDD Cup 2012-Track2, which can prove the feasibility and validity of this method.