Browse Prior Art Database

Method and System for Cross Feature Engineering for Ad Click and Conversion Prediction

IP.com Disclosure Number: IPCOM000249744D
Publication Date: 2017-Mar-30
Document File: 5 page(s) / 183K

Publishing Venue

The IP.com Prior Art Database

Related People

Shaunak Mishra: INVENTOR [+4]

Abstract

A method and system is disclosed for finding predictive cross features in a scalable manner for the task of ad click and conversion prediction given data from the target domain and domains related to the target domain. The method and system is implemented to reverse engineer relevant cross features from a trained deep learning model (multilayer perceptrons). In finding such cross features, the method and system leverages data in domains related to the target domain using transfer learning with multilayer perceptrons. The cross features are then readily used as additional features in a logistic regression model.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 45% of the total text.

Method and System for Cross Feature Engineering for Ad Click and Conversion Prediction

Abstract

A method and system is disclosed for finding predictive cross features in a scalable manner for the task of ad click and conversion prediction given data from the target domain and domains related to the target domain.  The method and system is implemented to reverse engineer relevant cross features from a trained deep learning model (multilayer perceptrons).  In finding such cross features, the method and system leverages data in domains related to the target domain using transfer learning with multilayer perceptrons.  The cross features are then readily used as additional features in a logistic regression model.

Description

A natural way of capturing feature interactions in linear predictive models is to consider quadratic features (crosses) from the original set of features.  In a large scale setting with millions of features, considering all possible cross features in the model may not be feasible and hence there is a need to have a pruned list of cross features.  Typically, domain knowledge is used to manually design such cross features which is not a very scalable option for large scale systems.

Thus, there exists a need for a method and system for finding predictive cross features in a scalable manner for the task of ad click and conversion prediction.

Disclosed is a method and system for finding predictive cross features in a scalable manner for the task of ad click and conversion prediction given data from the target domain and domains related to the target domain.  The method and system utilizes a machine learning method for finding such cross features by leveraging transfer learning and deep learning techniques.

The method and system is based on transfer learning using a multilayer perceptron (MLP), followed by extraction of cross features from the hidden layer outputs of the MLP.  For the transfer learning step, the method and system first trains an MLP using data from the target domain as well as domains related to the target domain.  The resultant

model is then used as the initial model for training another target MLP on just the target domain data.

Multilayer perceptrons are feed forward neural networks.  MLP with a single hidden layer is described as shown below.

where  is the feature vector (dimension n), is the hidden layer weights matrix, is the hidden layer bias vector and is a non-linear function.  The outputs from this layer are further processed in the output layer as follows.

where p(c = 1|x) and p(c = 0|x) denote the probabilities of the feature vector x being in class c = 1 and c = 0 respectively and   is the softmax function.

The single hidden layer MLP as described above is shown in FIG. 1.  Additional hidden layers can be stacked to get deeper models as shown in FIG. 2 (two hidden layer MLP).

Figure 1

Figure 2

For cross feature extraction, the method and system first identifies a hidden layer in the traine...