Content Based Recommender System

Content-Based Recommender Systems

Content-based recommendation is a decision-support tool. Rather than having a salesperson help a shopper choose products, the system uses information about the user and items to suggest relevant options. It recommends items similar to those the user has previously engaged with, ideally accounting for how their interests change over time.

Main Ideas

• User history: Searches, views, purchases, and other interactions.
• Item descriptions and features: Information used to determine how similar items are.
• User preference model: A profile built from the user’s past interactions with items.

Content-based methods are especially useful for recommending documents, such as news articles. Documents can be compared using their words, keywords, and other text-derived features.

Item Features

• Extrinsic features are readily available and apparent. For a movie, these might include its category, ratings, length, origin, or awards.
• Intrinsic features are derived from the item itself. For text, these might include nouns and noun phrases. Text-mining measures such as TF-IDF can also help represent documents.

Choosing which features to use—and how many to retain—is an iterative process. Text can produce thousands of possible features, so statistical or machine-learning techniques can help identify the most informative ones.

Typical Process

1. Extract and select item features. Identify relevant attributes from available data or derive them from the item’s content.
2. Represent the items. Encode each item using the selected feature values.
3. Prepare user-specific training data. Use item features as input variables and the user’s preference as the output. Preferences may be recorded explicitly, such as through ratings, or inferred implicitly, such as from whether the user viewed an item.
4. Learn a user profile. Train a model to relate item features to the user’s preferences. Historical data can be divided into training and testing or validation sets.
5. Generate recommendations. Use the learned profile and item details to estimate the user’s preferences for items they have not yet encountered.

Decision Trees and Information Gain

The lecture demonstrates decision-tree induction as one way to learn a preference model. A computer-purchase example uses attributes such as age, income, student status, and credit rating to predict whether a customer will buy a computer. The example is illustrative; the same general modeling approach can be applied to item features and a user’s history.

To build the tree:

1. Start with the available labeled examples.
2. Consider splitting the data on each input attribute.
3. Measure the uncertainty in the full dataset using entropy.
4. For each candidate attribute, calculate the weighted uncertainty remaining after splitting the data into groups.
5. Subtract that result from the original uncertainty to obtain the attribute’s information gain.
6. Choose the attribute with the greatest information gain as the next split.
7. Repeat within the resulting groups, stopping when the examples lead to the same decision.

In the computer-purchase example, age is selected as the root because it has the highest information gain. Further splits use attributes such as student status or credit rating where needed.

Suggested Assignment

The instructor asks students to build a decision tree for customers of an online song store and decide whether a particular customer should be targeted with:

• A new English-language album of classical songs.
• A new rock song from a Telugu movie.

This is framed as a top-M-user targeting problem: for a new item, use each customer’s browsing history to build or apply a preference model, then identify the customers most likely to be interested.

Speakers and Sources

• Instructor/narrator: Unnamed; delivers the lecture.
• Background music: Heard briefly; no performer or source is identified.