Abstract
Online advertising has expanded to a hundred-dollar billion industry in recent years, with sales growing at faster rate in every year. Prediction of the click-through rate (CTR) is an important role in recommended systems and online ads. Click through rating (CTR) is the newest evolution in the advertising and marketing digital world. It is essential for any online advertising company in real time to display the appropriate ads to the right users in the correct context. A huge amount of research work proposed considers each ad separately and does not takes in the relationship with other ads that may have an impact on Click Through Rate. A Factorization machine, a more generalized predictor like support vector machines (SVM) is not able to estimate reliable parameters under sparsity. The main drawback is that the primary features and existing algorithms considers the large weighted parameters. KGCN (Knowledge graph-based convolution network) overcomes the drawback and works on alternating graphs which creates additional clustering and node comparison with high latency and performance. A new framework DeepLight Weight is proposed to resolve the high server latency and high usage of memory issues in online advertising. This work presents a framework to improve the CTR predictions with an objective to accelerate the model inference, prune redundant parameters and the dense embedding vectors. Field Weighed Factorization machine helps to organize the data features with high structure to improve the accuracy. For clearing latency issues, structural pruning makes the algorithm work with dense matrices by combining and executing the individual matrix values or neural nodes.