Content‑Based Filtering: Concepts, Implementation, and Pros/Cons

Basic Concepts

Content‑based filtering recommends items similar to those a user likes by comparing item attributes such as title, year, or description, rather than using collaborative usage patterns. Each item is represented by a feature vector, and user preference models can be built using decision trees, neural networks, or vector‑based methods.

Features

1. User profiles require historical data. 2. Profiles may evolve as user preferences change.

Implementation Principle

Assume a rating matrix where rows are users and columns are items (e.g., books). Ratings range from 1 to 5; empty cells indicate no rating. The first step is to compute item similarity based on content, here simplified to title keywords. After stop‑word removal, each title is represented as a binary vector (1 if the word appears, otherwise 0).

Cosine similarity is then applied to compare vectors, producing a similarity matrix (Figure 4) and a full pairwise similarity heatmap (Figure 5). Based on these similarities, the system recommends the most similar unseen items to the user (Figure 6).

Advantages and Disadvantages

Advantages

(1) No need for extensive usage data. (2) Can recommend niche items for users with special interests. (3) Provides explainable recommendations via content features. (4) Works with a single user. (5) Avoids popularity bias, mitigating the “new‑item” problem.

Disadvantages

(1) Item content must be machine‑readable and meaningful. (2) Prone to over‑specialization. (3) Limited novelty; recommendations may be too similar. (4) Difficult to combine multiple item attributes. (5) May suffer from shallow content analysis.

Supplementary Algorithms

1. k‑Nearest Neighbor (kNN)

For a new item, find the K most similar items the user has already rated, then infer the user's preference from those ratings. Similarity can be computed with Euclidean distance for structured data or cosine for vector‑space models.

2. Rocchio Algorithm

Originally from information retrieval, Rocchio updates a query vector based on relevant feedback. In recommendation, it can adjust a user profile vector using weighted sums of liked and disliked item vectors.

3. Decision Tree (DT)

Effective when items have few, structured attributes, providing interpretable rules. Performance degrades with many unstructured attributes.

4. Linear Classifier (LC)

Finds a hyperplane that separates liked from disliked items in high‑dimensional space, typically trained with gradient descent.

5. Naive Bayes (NB)

Assumes conditional independence of item attributes given a class; simple to implement and often yields strong baseline performance for text‑based recommendation.