Machine-learned neural network architectures for incremental lift predictions using embeddings
An online system trains a machine-learned lift prediction model configured as a neural network. The machine-learned lift prediction model can be used during the inference process to determine lift predictions for users and items associated with the online system. By configuring the lift prediction model as a neural network, the lift prediction model can capture and process information from users and items in various formats and more flexibly model users and items compared to existing methods. Moreover, the lift prediction model includes at least a first portion for generating control predictions and a second portion for generating treatment predictions, where the first portion and the second portion share a subset of parameters. The shared subset of parameters can capture information important for generating both control and treatment predictions even when the training data for a control group of users might be significantly smaller than that of the treatment group.
1. A method, comprising:
obtaining a plurality of training instances, a training instance including a set of features describing a respective user and a respective product associated with an online system, a treatment identifier indicating whether the user was presented with a content item of the product, and a label indicating whether the user performed a desired action for the product;
accessing a machine-learned neural network model with an initial set of parameters, the neural network model including a user portion and a product portion, the product portion including shared item layers, a control embedding branch, and a treatment embedding branch;
repeatedly performing, for one or more iterations:
selecting a subset of the plurality of training instances for the iteration,
for each training instance in the selected subset, generating an estimated likelihood for the training instance, comprising:
applying the user portion to a first set of features for the training instance to generate a user embedding that represents the user in a latent space;
responsive to the treatment identifier for the training instance indicating that the content item was presented to the user:
applying the shared item layers to the second set of features for the training instance to generate intermediate outputs;
applying the control embedding branch to the intermediate outputs to generate a control embedding;
applying the treatment embedding branch to the intermediate outputs to generate a treatment embedding;
combining the control embedding and the treatment embedding to generate a summed output;
generating a combined output by combing the user embedding with the summed embedding; and
generating the estimated likelihood by applying a mapping function to the combined output; and
updating the parameters of the neural network model based on a loss function that indicates a difference between the labels and the estimated likelihoods for the subset of training instances; and
storing the trained parameters of the neural network model on a computer-readable medium.
2. The method of claim 1 , wherein each of the control embedding branch and the treatment embedding branch are configured as neural network layers, and wherein parameters of the control embedding branch are separate from parameters of the treatment embedding branch.
3. The method of claim 1 , wherein the desired action is the user making a purchase of the item.
4. The method of claim 1 , wherein the set of features includes a first set of features characterizing the user and a second set of features characterizing the item.
5. The method of claim 4 , wherein the first set of features include at least one or a combination of age of the user, geographic location associated with the user, or account information of the user, wherein the item is a product, and wherein the second set of features include at least one or a combination of a size of the product, color of the product, weight of the product, stock keeping unit (SKU) of the product, or serial number of the product.
6. The method of claim 1 , wherein the treatment identifier for the respective training instance is zero if the user was not presented with the respective content item and a non-zero value if the user was presented with the respective content item.
7. A non-transitory computer readable storage medium storing parameters of a machine-learned neural network model, manufactured by a process comprising:
obtaining a plurality of training instances, a training instance including a set of features describing a respective user and a respective product associated with an online system, a treatment identifier indicating whether the user was presented with a content item of the product, and a label indicating whether the user performed a desired action for the product;
accessing the machine-learned neural network model with an initial set of parameters, the neural network model including a user portion and a product portion, the product portion including shared item layers, a control embedding branch, and a treatment embedding branch;
repeatedly performing, for one or more iterations:
selecting a subset of the plurality of training instances for the iteration,
for each training instance in the selected subset, generating an estimated likelihood for the training instance, comprising:
applying the user portion to a first set of features for the training instance to generate a user embedding that represents the user in a latent space;
responsive to the treatment identifier for the training instance indicating that the content item was presented to the user:
applying the shared item layers to the second set of features for the training instance to generate intermediate outputs;
applying the control embedding branch to the intermediate outputs to generate a control embedding;
applying the treatment embedding branch to the intermediate outputs to generate a treatment embedding;
combining the control embedding and the treatment embedding to generate a summed output;
generating a combined output by combining the user embedding with the summed embedding; and
generating the estimated likelihood by applying a mapping function to the combined output; and
updating the parameters of the neural network model based on a loss function that indicates a difference between the labels and the estimated likelihoods for the subset of training instances; and
storing the trained parameters of the neural network model on the computer-readable medium.
8. The non-transitory computer readable storage medium of claim 7 , wherein each of the control embedding branch and the treatment embedding branch are configured as neural network layers, and wherein parameters of the control embedding branch are separate from parameters of the treatment embedding branch.
9. The non-transitory computer readable storage medium of claim 7 , wherein the desired action is the user making a purchase of the item.
10. The non-transitory computer readable storage medium of claim 7 , wherein the set of features includes a first set of features characterizing the user and a second set of features characterizing the item.
11. The non-transitory computer readable storage medium of claim 10 , wherein the first set of features include at least one or a combination of age of the user, geographic location associated with the user, or account information of the user, wherein the item is a product, and wherein the second set of features include at least one or a combination of a size of the product, color of the product, weight of the product, stock keeping unit (SKU) of the product, or serial number of the product.
12. The non-transitory computer readable storage medium of claim 7 , wherein the treatment identifier for the respective training instance is zero if the user was not presented with the respective content item and a non-zero value if the user was presented with the respective content item.
13. The method of claim 1 , wherein the control embedding branch and the treatment embedding branch do not share parameters with each other.
14. The method of claim 1 , wherein the control embedding branch and the treatment embedding branch do not share parameters with each other.