CineGeist: A Movie Recommender for People Who Can't Describe Their Taste

I constructed a movie recommender for individuals who struggle to articulate their own preferences, a concept I've explored in depth. People often find it challenging to describe their tastes when asked, a phenomenon driven by indecisiveness. When presented with well-received films, I've been on hold listening to them for up to thirty minutes.

People are often unreliable narrators of their own tastes, making it easier to choose between two options than to describe one's taste in abstract terms. This inconsistency is the core concept of this project, a conversational movie recommender aimed at those who can't describe their preferences. The project's architecture prevents the language model from hallucinating or false aligning.

Instead of asking open-ended questions about taste, the system prompts users to react and choose between concrete options. This approach is based on the concept of a "wildcard," a film with a low connection to the user's profile but still relevant to their taste. This approach was tested using artificial identities based on six popular films, with the wildcard feature never appearing in 3,000 simulated users.

The system's effectiveness is measured by cosine similarity, with a maximum similarity threshold set at 0.35. This approach ensures that the model only phrases responses, not decides facts, resulting in an accurate and effective recommendation system.