Implicit rating recommendations feedback loop
Examples described herein generally relate to a computer device including a memory and a processing system configured to customize a user interface based on an implicit rating of user interface elements. The computer device monitors user interactions with a user interface prior to a goal action. The computer device trains weights of a model, based on collected events over a first window of time, to generate an implicit rating that is predictive of the goal action. The computer device generates the implicit rating for a specific user of at least one element of the user interface based on the weights of the model applied to each type of event of the user over a second window of time. The computer device applies the implicit rating as an input to a collaborative filter to generate a recommendation to modify the local version of the user interface for the specific user.
1 . A computer device for customizing a user interface, comprising:
a memory; and
a processing system comprising at least one processor communicatively coupled with the memory and configured to:
collect page views and page actions between each user of a plurality of users and a local version of a user interface prior to a goal action;
train weights of a model that is configured to generate an implicit rating that is predictive of the goal action, wherein to train the weights of the model, the processor is configured to:
compute a goal-setting variable as a weighted combination of counts of a plurality of different goal actions observed in a first window; and
minimize a difference between the implicit rating and the goal-setting variable by adjusting the weights;
generate the implicit rating for a specific user of at least one element of the user interface based on the adjusted weights of the model applied to each type of page view and page action of the user over a second window of time; and
recommended modification of the local version of the user interface for the specific user by applying the implicit rating as an input to a collaborative filter.
2 . The computer device of claim 1 , wherein the first window of time includes a plurality of time periods, each time period associated with a different set of weights for the page views and page actions during a corresponding time period.
3 . The computer device of claim 1 , wherein the collaborative filter uses a user-item transaction matrix, wherein a score for an item is based on the implicit rating of the user.
4 . The computer device of claim 1 , wherein to collect page views and page actions, the user interface is configured to generate an object including user input and metadata.
5 . The computer device of claim 4 , wherein the processing system is configured to classify the user input and metadata into a key-value pair that defines an event associated with a set of weights.
6 . The computer device of claim 1 , wherein the model is represented by the weights of nodes in a neural network, wherein to train the model, the processing system is configured to calculate a gradient descent based on a difference between the rate of the goal action and the implicit rating.
7 . The computer device of claim 1 , wherein the second window of time is the same as the first window of time.
8 . The computer device of claim 1 , wherein the second window of time is a same duration as the first window of time measured back from a current time.
9 . The computer device of claim 1 , wherein the processing system is further configured to:
determine a first rate of the goal action on the user interface for a first group of users that receive a recommendation based on the implicit rating;
determine a second rate of the goal action on the user interface for a second group of users that do not receive a recommendation based on the implicit rating; and
continue use of the model for the implicit rating when the first rate is higher than the second rate.
10 . The computer device of claim 9 , wherein the processing system is further configured to retrain the model when the second rate is higher than the first rate.
11 . A method of recommending modifications to a user interface, comprising:
collecting page views and page actions between each user of a plurality of users and a local version of a user interface prior to a goal action;
training weights of a model that is configured to generate an implicit rating that is predictive of the goal action, wherein training the weights of the model comprises:
computing a goal-setting variable as a weighted combination of counts of a plurality of different goal actions observed in a first window; and
minimizing a difference between the implicit rating and the goal-setting variable by adjusting the weights;
generating the implicit rating for a specific user of at least one element of the user interface based on the adjusted weights of the model applied to each type of page view and page action of the user over a second window of time; and
generating a recommended modification of the local version of the user interface for the specific user by applying the implicit rating as an input to a collaborative filter.
12 . The method of claim 11 , wherein the first window of time includes a plurality of time periods, each time period associated with a different set of weights for the page views and page actions during a corresponding time period.
13 . The method of claim 11 , wherein the collaborative filter uses a user-item transaction matrix, wherein a score for an item is based on the implicit rating of the user.
14 . The method of claim 11 , wherein the collecting comprises generating, by instrumentation within the user interface, an object including user input and metadata.
15 . The method of claim 14 , wherein the collecting comprises classifying the user input and metadata into a key-value pair that defines an event associated with a set of weights.
16 . The method of claim 11 , wherein the model is represented by the weights of nodes in a neural network, wherein training the model comprises calculating a gradient descent based on a difference between the rate of the goal action and the implicit rating.
17 . The method of claim 11 , further comprising modifying the local version of the user interface based on the recommendation to include a recommended element.
18 . The method of claim 17 , wherein the recommended element is one of a content item within a frame of the user interface, a pop-up notification, or text within a chat interface.
19 . The method of claim 11 , further comprising:
comparing a first rate of the goal action on the user interface for a first group of users that receive a recommendation based on the implicit rating to a second rate of the goal action on the user interface for a second group of users that do not receive a recommendation based on the implicit rating; and
continuing use of the model for implicit rating when the first rate is higher than the second rate.
20 . The method of claim 19 , further comprising retraining the model when the second rate is higher than the first rate.