Convolutional Neural Networks based Click-Through Rate Prediction with Multiple Feature Sequences

Abstract

Convolutional Neural Network (CNN) achieved satisfying performance in click-through rate (CTR) prediction in recent studies. Since features used in CTR prediction have no meaningful sequence in nature, the features can be arranged in any order. As CNN learns the local information of a sample, the feature sequence may influence its performance significantly. However, this problem has not been fully investigated. This paper firstly investigates whether and how the feature sequence affects the performance of the CNN-based CTR prediction method. As the data distribution of CTR prediction changes with time, the best current sequence may not be suitable for future data. Two multi-sequence models are proposed to learn the information provided by different sequences. The first model learns all sequences using a single feature learning module, while each sequence is learnt individually by a feature learning module in the second one. Moreover, a method of generating a set of embedding sequences which aims to consider the combined influence of all feature pairs on feature learning is also introduced. The experiments are conducted to demonstrate the effectiveness and stability of our proposed models in the offline and online environment on both the benchmark Avazu dataset and a real commercial dataset.

An Ensemble Model for Click Through Rate Prediction

“…Most of the machine learning algorithms are comparable to each other in the way they preprocess the data. The key difference between the existing researches is in the feature engineering of the dataset [1]. Some of the early researches consider the entire high dimensional data set for the training while certain studies use the machine learning techniques to select the set of features that will help in fine tuning the model in turn improving the accuracy of the predictions.…”

“…This study will help the business owners to target their potential customers with the right advertisement structure on a quality website. On the other hand, this study will as well be useful for the search engines and the website owners in deciding the category of advertisements that are suitable to facilitate the online advertisers with the bidding rate for an ad, space and orientation of the ad based on the traffic and rank of their website [1]. Hence, this is an interesting area of research as it will help both the advertisement platform owners as well as the business providers with the intuitive insights in choosing the advertisements that are profitable in terms of the monetary conversion.…”

“…It is imperative to study and consider these feature interactions when designing the models to predict the probability. As the features are implicit and nonlinear [1] like the traditional beer and diaper correlation, this research also concentrates on the feature selection and analysis using different machine learning models.…”

“…The DeepFM uses two components namely the Factorization Machine and the Deep Neural Network model that take the same input. A scalar weight is used to measure the importance of a feature and a vector is generated to learn its interaction with other features [1].…”

“…The Avazu is a leading advertising platform that has provided the researchers with the dataset. The data set contains rich information about the advertisement that was displayed over a period of 10 days sorted according to increasing time [1]. This can be used for the training purpose.…”