Made with Humane Club
Handling Local news in personalization models can be challenging.
CopiedMedium will de-amplify Generative AI generated content from its platform and will actively evolve its personalization system to spot and de-amplify AI written content.
CopiedDriving factors: Differentiation
Differentiation typically comes from building authenticity on your core product, i.e., editorial product. This involves owning more of the value chain:
- Are you a trusted brand or voice for a well-defined topic?
- Are you able to set the agenda for a well-defined audience segment?
- How strong is your recall among that well-defined audience segment?
- Have you stitched the most active members of that audience segment into a tight knit community?
- Is the community co-building with you or via you?
- How are you ensuring that you are not only building but also retaining 1:1 relationship within that community?
- What’s your information architecture and content strategy?
- How deep are you serving that vertical?
- Is your website (yes! a website even after GenerativeAI) representative of who you are?
Amplification factors
While the above gets your driving factors in place, the next aspects define how much you can amplify. This includes:
- Distribution
- Ability to sell advertisements, subscriptions, or events
- Ability to market and develop audience
CopiedHow to think about AI for news?
Model performance metrics
- accuracy
- precision
- recall
Copied
Sometimes the model can overfit the training data and hence underperforms with real data. To reduce complexity, we use regularization.
- Add a penalty factor on complexity i.e. number of features and values of weights.
- Lasso: Force coefficients to zero if not relevant. Ends up performing feature selection.
- Ridge: Reduces coefficients of irrelevant features but does not drop them to zero.
CopiedDecision trees are versatile tools that can be employed for both Regression and Classification tasks.
When developing a decision tree, it’s crucial to ask a series of targeted questions in order to zero in on the most accurate label. The objective is to construct an efficient tree that minimizes the number of splits necessary to divide the data. This process continues until no further splits can be made.
It’s important to find the right balance, as shallow trees tend to under-fit the data, while deep trees can lead to overfitting, with each example ending up as its own leaf node.
CopiedPredict numeric values
Linear Regression builds relationship between inputs (features) and outputs (targets). It is a simple yet effective model to start with to get a benchmark.
- Using training data, the goal is to minimize the loss.
- You can measure loss function using SSE
Polynomial Regression: Linear Regression can also be used to model a non-linear relationship. To achieve this, we transform an input feature by a non-linear transformation function and create a new feature that can be used as input in the model.
CopiedChinese proverb: One chopstick is fragile, but a group of chopsticks is robust.
A big challenge with ML models is overfitting to training data. And hence, we can use ensemble models — average multiple models into a meta model, which in turn reduces variance.
CopiedChinese proverb: Make a mistake and learn a lesson
Similar to unsupervised learning, it does not need labeled data. The model learns from interactions (trial and error)and hence takes time to perfect. Initially the algorithm acts randomly, exploring the environment, but it learns with time by being rewarded when it makes the right choices.
It is best suited for use cases that require achieving a goal, for example, autonomous vehicles, headline testing or personalizing a website.
CopiedGoogle Chat now summarizes threads.
CopiedContent filtering focuses one user’s interest. In this method, users get what they like. However, this also means users have lower chance of discovering something fresh serendipitously. Content filtering models can be retrained at a drastically lower frequency.
Copied