Multi-level model architecture for controlling the size of user behavior models
Embodiments of a user behavior inference system are disclosed that uses a multi-level machine learning (MLML) model to make inferences about user behaviors. The MLML model is organized in multiple levels, where an upper level processes part of the model input to generate control parameters for a lower level, and the lower level processes another part of the model input according to the control parameters to generate inference results. In embodiments, the upper level receives a user feature vector of a user and the lower level receives a behavior feature vector of the user. In embodiments, the model is trained to output a propensity score of the user for acquiring an item, and an item feature vector is included in the model input. Advantageously, the multi-level model architecture reduces the overall size of the model and the amount of training time and data needed to achieve a desired model performance.
1 . A system, comprising:
one or more computer devices that implement a user behavior inference system, configured to:
execute a multi-level machine learning (MLML) model trained using one or more ML techniques to generate inferences about user behaviors, wherein the MLML model is organized in multiple levels according to a multi-level model architecture that includes an upper level and a lower level;
receive (a) a user feature vector encoding user features of a user and (b) a behavior feature vector encoding a behavior sequence of the user;
input the user feature vector to the upper level of the MLML model, wherein the upper level processes the user feature vector to output control parameters of the lower level without processing the behavior feature vector;
input the behavior feature vector to the lower level of the MLML model, wherein the lower level of the MLML model processes the behavior feature vector according to the control parameters produced by the upper level of the MLML model based on the user feature vector, to generate a behavior inference about the user; and
wherein the multi-level model architecture (a) limits the MLML model to a specified model size and (b) reduces an amount of training time or training data used to train the MLML model to a specified performance level.
2 . The system of claim 1 , wherein:
the user behavior inference system is implemented as part of an ad network configured to deliver ads about different items to user devices;
the MLML model is trained to output propensity scores of different users for acquiring different items;
input to the upper level of the MLML model includes an item feature vector that encodes features of an item; and
the MLML model outputs a propensity score of the user for acquiring the item, wherein the propensity score is used by the ad network to select an audience for an ad about the item.
3 . The system of claim 2 , wherein:
the input to the upper level of the MLML model includes a context feature vector that encodes features of an ad impression context associated with an impression associated with the ad; and
the ad impression context includes information that indicates one or more of:
a time of the ad impression,
a type of a user device associated with the ad impression,
a location of the user device at the time of the ad impression,
a content category or webpage used to deliver the ad impression,
a placement of the ad impression within a content or webpage,
a media type of the ad impression, or
a user action related to a previous ad impression.
4 . The system of claim 2 , wherein:
the ad network is configured to deliver ads for a plurality of advertiser clients;
the MLML model is created for a particular advertiser client according to configuration input specified by the particular advertiser client;
the ad network is managed by resources of a multi-tenant infrastructure provider network; and
the MLML model is managed by a machine learning service of the multi-tenant infrastructure provider network.
5 . The system of claim 1 , wherein:
the MLML model is a neural network;
the upper level is a one-hidden layer neural network; and
the lower level is a two-hidden layer neural network.
6 . A method, comprising:
performing, by a user behavior inference system implemented by one or more computer devices:
executing a multi-level machine learning (MLML) model trained using one or more ML techniques to generate inferences about user behaviors, wherein the MLML model is organized in multiple levels according to a multi-level model architecture that includes an upper level and a lower level;
receiving (a) a user feature vector encoding user features of a user and (b) a behavior feature vector encoding a behavior sequence of the user;
inputting the user feature vector to the upper level of the MLML model, wherein the upper level processes the user feature vector to output operating parameters of the lower level without processing the behavior feature vector; and
inputting the behavior feature vector to the lower level of the MLML model, wherein the lower level of the MLML model processes the behavior feature vector according to the control parameters produced by the upper level of the MLML model based on the user feature vector, to generate a behavior inference about the user.
7 . The method of claim 6 , wherein:
the behavior inference includes a propensity score of the user for acquiring an item; and
the method further comprises selecting an audience for an ad about the item based on a plurality of propensity scores generated by the MLML model.
8 . The method of claim 7 , further comprising:
encoding features of the item in an item feature vector, wherein the item feature vector is input to the upper level of the MLML model.
9 . The method of claim 8 , wherein:
the behavior feature vector encodes a plurality of user actions of the user with respect to different items; and
the behavior feature vector includes item feature vectors of the different items.
10 . The method of claim 6 , wherein:
the MLML model is a neural network;
the upper level is a one-hidden layer neural network; and
the lower level is a two-hidden layer neural network.
11 . The method of claim 6 , wherein:
the lower level is a neural network of neuron units connected by a plurality of unit-to-unit connections; and
the control parameters produced by the upper level include weights that are applied to individual ones of the unit-to-unit connections.
12 . The method of claim 6 , wherein:
the behavior sequence of the user includes user actions associated with a plurality of items; and
the method further comprises performing a distillation process on the behavior sequence to reduce the behavior sequence to user actions associated with a reduced set of the items, wherein the distillation process reduces a size of the behavior feature vector used to encode the behavior sequence.
13 . The method of claim 12 , wherein:
the behavior feature vector includes item encodings of items associated with individual user actions; and
the distillation process reduces a size of the item encodings in the behavior feature vector.
14 . The method of claim 6 , further comprising:
training the upper and lower levels of the MLML model together for one category of users or items; and
performing a transfer learning process to train the MLML model for a second category of users or items, wherein the transfer learning process trains only one of the upper or lower level.
15 . The method of claim 6 , further comprising:
receiving model configuration input for a new model via configuration interface of the user behavior inference system, wherein the configuration input specifies to create the new model using the multi-level model architecture; and
creating the MLML model according to the configuration input.
16 . The method of claim 15 , wherein the configuration input specifies a number of levels in the MLML model or an assignment of input feature vectors to different ones of the levels.
17 . The method of claim 15 , wherein the configuration interface indicates:
an estimated model cost associated with training or using the MLML model using the multi-level model architecture; and
an estimated savings in the model cost achieved based on use of the multi-level model architecture.
18 . One or more non-transitory computer-accessible storage media storing program instructions that when executed on one or more processors of user behavior inference system, cause the user behavior inference system to:
execute a multi-level machine learning (MLML) model trained using one or more ML techniques to generate inferences about user behaviors, wherein the MLML model is organized in multiple levels according to a multi-level model architecture that includes an upper level and a lower level;
receive (a) a user feature vector encoding user features of a user and (b) a behavior feature vector encoding a behavior sequence of the user;
input the user feature vector to the upper level of the MLML model, wherein the upper level processes the user feature vector to output control parameters of the lower level without processing the behavior feature vector; and
input the behavior feature vector to the lower level of the MLML model, wherein the lower level of the MLML model processes the behavior feature vector according to the control parameters produced by the upper level of the MLML model based on the user feature vector, to generate a behavior inference about the user.
19 . The non-transitory computer-accessible storage media of claim 18 , wherein:
the behavior inference includes a propensity score of the user for acquiring an item; and
a plurality of propensity scores generated by the MLML model are used to select an audience for an ad about the item.
20 . The non-transitory computer-accessible storage media of claim 19 , wherein the program instructions when executed on the one or more processors cause the user behavior inference system to input an item feature vector to the MLML model that encodes features of the item.