Two-Model Click Intention Prediction Reducing False Positives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learned models are unreliable in predicting user intention behind clicks on content, as they are trained with false positives, where users may not intend to apply for jobs despite clicking on job listings.
Innovation Solution
A two-machine learned model approach, where the first model assesses the strength of positive signals and labels training data with a measure of user intention, and the second model uses these labels to predict user interest levels by calculating entity and activity similarity scores in a D-dimensional space.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learned models are trained using click data as positive signals, then the models can predict user engagement, but the predictions become unreliable due to false positives where users click without genuine intention
Solution Approach 1:
The patent segments the single machine learned model into two distinct models: a first model that processes user interaction data and generates interest scores, and a second model that uses these scores along with additional features to predict user engagement. This segmentation allows the system to separately handle the detection of genuine interest from the prediction of engagement, thereby improving prediction reliability by filtering out false positives through the intermediate interest scoring step.
Solution Approach 2:
The patent introduces an intermediary mechanism (the first machine learned model generating interest scores) between the raw click data and the final engagement prediction. This intermediary processes the raw interaction data to distinguish genuine interest from incidental clicks before passing refined signals to the second model, thereby preserving user intention information that would otherwise be lost in traditional direct training approaches.
2Quantity of substance
If machine learned models use all user clicks as training data, then more training data is available, but the model accuracy decreases due to inclusion of non-genuine clicks
Solution Approach 1:
The patent changes the parameter representation of user interactions by transforming raw click data into refined interest scores through the first machine learned model. This parameter transformation allows the system to maintain a large quantity of training data while improving measurement precision, as the interest scores encode the genuine intent behind each interaction rather than treating all clicks equally.
Solution Approach 2:
The patent applies preliminary action by having the first machine learned model pre-process user interaction data to generate interest scores before the second model performs engagement prediction. This preliminary filtering and scoring step prepares the training data by distinguishing genuine interest from incidental clicks, ensuring that the second model receives high-quality labeled data without sacrificing the overall quantity of available training examples.
3Productivity
If the system serves more content to users, then user engagement opportunities increase, but the relevance of content decreases leading to lower user interest
Solution Approach 1:
The patent implements feedback mechanisms where the two-machine learned model system continuously learns from user interactions. The first model generates interest scores that feed into the second model's predictions, and both models are trained on labeled data that reflects actual user engagement outcomes. This feedback loop enables the system to maintain high content relevance even when delivering large volumes of content, as the models adapt to user preferences and distinguish relevant from irrelevant content based on learned patterns of genuine interest.
Data Source
AI summary
In an example embodiment, a first machine learned model is trained to produce output, and a second machine learned model is then trained using training data that has been labeled, at least partially, using the output of the first machine learned model. The first machine learned model is trained to output a measure of how strong a positive signal in the training data really is. Specifically, this measure indicates the level of intention of a user who has engaged in a first user interface action with respect to a piece of content to engage in a subsequent second user interface action with the same piece of content.