System and method for improving conversion rate prediction via self-supervised pretraining in online advertising

1 . A method, comprising:

receiving bids directed to a display ad opportunity, wherein each of the bids includes a candidate advertisement and the display ad opportunity involves a user and an associated context;

obtaining auxiliary features with respect to each of the bids based on a code generated by an autoencoder for the bid, wherein the autoencoder is implemented with an artificial neural network and trained based on click attributed conversion data and non-click attributed conversion data, and the click attributed conversion data and non-click attributed conversion data are collected based on online user activities on displayed ads;

predicting, via an enriched performance prediction model implemented with an artificial neural network and trained based on the click attributed conversion data, a predicted performance metric for the candidate advertisement in each of the bids based on the auxiliary features associated with the candidate advertisement; and

selecting the winning advertisement from candidate advertisements of the bids according to a ranking determined based on the respective predicted performance metrics of the candidate advertisements.

2 . The method of claim 1 , wherein

the auxiliary features are computed based on codes generated by the autoencoder using the click-attributed performance data.

3 . The method of claim 2 , wherein the click attributed and non-click attributed conversion data include events associated with a specified performance, wherein each of the events includes:

user features characterizing a user involved in the specified performance;

ad features characterizing an advertisement through which the user achieved the specified performance; and

context features characterizing contextual information associated with the user, a delivery of the advertisement, and other relevant information.

4 . The method of claim 1 , wherein:

the candidate advertisement is characterized by ad features;

the user associated with the display ad opportunity is characterized by user features; and

the context associated with the display ad opportunity is characterized by context features.

5 . The method of claim 4 , wherein the obtaining auxiliary features comprises:

accessing the ad features, the user features, and the context features;

generating, by the autoencoder, the code with respect to the bid based on the accessed ad features, user features, and context features; and

transforming the code into the auxiliary features with respect to the bid.

6 . The method of claim 4 , wherein the predicting a predicted performance metric comprises:

generating, via embeddings learned in training the enriched performance prediction model, ad feature vectors, user feature vectors, and context feature vectors based on respectively the ad features, the user features, and the context features;

combining the ad feature vectors, the user feature vectors, and the context feature vectors to create a first vector;

creating a second vector based on the auxiliary features;

combining the first and the second vectors to generate an overall vector; and

generating a predicted performance metric based on the overall vector.

7 . The method of claim 1 , further comprising:

displaying the selected advertisement to the user;

tracking response of the user to the displayed winning advertisement;

generating new training data on tracked user activities on display ads; and

conducting periodic adaptive learning of the autoencoder and the enriched performance prediction model based on the new training data.

8 . A machine readable and non-transitory medium having information recorded thereon, wherein the information, when read by the machine, causes the machine to perform the following steps:

receiving bids directed to a display ad opportunity, wherein each of the bids includes a candidate advertisement and the display ad opportunity involves a user and an associated context;

obtaining auxiliary features with respect to each of the bids based on a code generated by an autoencoder for the bid, wherein the autoencoder is implemented with an artificial neural network and trained based on click attributed conversion data and non-click attributed conversion data, and the click attributed conversion data and non-click attributed conversion data are collected based on online user activities on displayed ads;

predicting, via an enriched performance prediction model implemented with an artificial neural network and trained based on the click attributed conversion data, a predicted performance metric for the candidate advertisement in each of the bids based on the auxiliary features associated with the candidate advertisement; and

selecting the winning advertisement from candidate advertisements of the bids according to a ranking determined based on the respective predicted performance metrics of the candidate advertisements.

9 . The medium of claim 8 , wherein

the auxiliary features are computed based on codes generated by the autoencoder using the click-attributed performance data.

10 . The medium of claim 9 , wherein the click attributed and non-click attributed conversion data include events associated with a specified performance, wherein each of the events includes:

user features characterizing a user involved in the specified performance;

ad features characterizing an advertisement through which the user achieved the specified performance; and

context features characterizing contextual information associated with the user, a delivery of the advertisement, and other relevant information.

11 . The medium of claim 8 , wherein:

the candidate advertisement is characterized by ad features;

the user associated with the display ad opportunity is characterized by user features; and

the context associated with the display ad opportunity is characterized by context features.

12 . The medium of claim 11 , wherein the obtaining auxiliary features comprises:

accessing the ad features, the user features, and the context features;

generating, by the autoencoder, the code with respect to the bid based on the accessed ad features, user features, and context features; and

transforming the code into the auxiliary features with respect to the bid.

13 . The medium of claim 11 , wherein the predicting a predicted performance metric comprises:

generating, via embeddings learned in training the enriched performance prediction model, ad feature vectors, user feature vectors, and context feature vectors based on respectively the ad features, the user features, and the context features;

combining the ad feature vectors, the user feature vectors, and the context feature vectors to create a first vector;

creating a second vector based on the auxiliary features;

combining the first and the second vectors to generate an overall vector; and

generating a predicted performance metric based on the overall vector.

14 . The medium of claim 8 , wherein the information, when read by the machine, further causes the machine to perform the following steps:

displaying the selected advertisement to the user;

tracking response of the user to the displayed winning advertisement;

generating new training data on tracked user activities on display ads; and

conducting periodic adaptive learning of the autoencoder and the enriched performance prediction model based on the new training data.

15 . A system, comprising:

a performance prediction unit implemented by a processor and configured for receiving bids directed to a display ad opportunity, wherein each of the bids includes a candidate advertisement and the display ad opportunity involves a user and an associated context, and

predicting, via an enriched performance prediction model implemented with an artificial neural network and trained based on click attributed conversion data, a predicted performance metric for the candidate advertisement in each of the bids based on auxiliary features with respect to each of the bids based on a code generated by an autoencoder for the bid associated with the candidate advertisement, wherein the autoencoder is implemented with an artificial neural network and trained based on the click attributed conversion data and non-click attributed conversion data, wherein the click attributed conversion data and non-click attributed conversion data are collected based on online user activities on displayed ads; and

a winning ad selection unit implemented by a processor and configured for selecting the winning advertisement from candidate advertisements of the bids according to a ranking determined based on the respective predicted performance metrics of the candidate advertisements.

16 . The system of claim 15 , wherein

the auxiliary features are computed based on codes generated by the autoencoder using the click-attributed performance data, wherein

the click attributed and non-click attributed conversion data include events associated with a specified performance, each of which includes

user features characterizing a user involved in the specified performance,

ad features characterizing an advertisement through which the user achieved the specified performance, and

context features characterizing contextual information associated with the user, a delivery of the advertisement, and other relevant information.

17 . The system of claim 15 , wherein:

the candidate advertisement is characterized by ad features;

the user associated with the display ad opportunity is characterized by user features; and

the context associated with the display ad opportunity is characterized by context features.

18 . The system of claim 17 , further comprising an auxiliary feature generator implemented ‘by a processor and configured for obtaining the auxiliary features by:

accessing the ad features, the user features, and the context features;

generating, by the autoencoder, the code with respect to the bid based on the accessed ad features, user features, and context features; and

transforming the code into the auxiliary features with respect to the bid.

19 . The system of claim 17 , wherein the performance prediction unit predicts the predicted performance metric for each candidate advertisement by:

generating, via embeddings learned in training the enriched performance prediction model, ad feature vectors, user feature vectors, and context feature vectors based on respectively the ad features, the user features, and the context features;

combining the ad feature vectors, the user feature vectors, and the context feature vectors to create a first vector;

creating a second vector based on the auxiliary features;

combining the first and the second vectors to generate an overall vector; and

generating a predicted performance metric based on the overall vector.

20 . The system of claim 15 , further comprising

an online ad performance tracker implemented by a processor and configured for tracking response of the user to the winning advertisement displayed to the user, and generating new training data on tracked user activities on display ads;

a batch-based autoencoder generator implemented by a processor and configured for conducting periodic adaptive learning of the autoencoder based on the new training data; and

a performance prediction model training unit implemented by a processor and configured for conducting periodic adaptive learning of the enriched performance prediction model based on the new training data.