[Paper Review] A multinomial probabilistic model for movie genre predictions

Eric Makita, Artem Lenskiy|arXiv (Cornell University)|Mar 25, 2016

Recommender Systems and Techniques7 references3 citations

TL;DR

This paper proposes a multinomial probabilistic model using Naive Bayes to predict movie genres from user ratings, treating genre prediction as a classification problem. It achieves a 70% genre prediction rate using only 15% of the MovieLens dataset for training, demonstrating that user rating patterns alone can effectively infer movie genres with high accuracy even from small training sets.

ABSTRACT

This paper proposes a movie genre-prediction based on multinomial probability model. To the best of our knowledge, this problem has not been addressed yet in the field of recommender system. The prediction of a movie genre has many practical applications including complementing the items categories given by experts and providing a surprise effect in the recommendations given to a user. We employ mulitnomial event model to estimate a likelihood of a movie given genre and the Bayes rule to evaluate the posterior probability of a genre given a movie. Experiments with the MovieLens dataset validate our approach. We achieved 70% prediction rate using only 15% of the whole set for training.

Motivation & Objective

• To address the gap in recommender systems that do not predict item genres beyond expert-labeled categories.
• To improve recommendation quality by enabling genre prediction that supports diversity, novelty, and serendipity.
• To explore whether user rating patterns alone can reliably predict movie genres without relying on metadata.
• To validate the effectiveness of a multinomial probabilistic model in genre prediction using real-world user rating data.

Proposed method

• The model uses a multinomial event model to estimate the likelihood of a movie given a genre, based on user rating patterns.
• Bayesian reasoning is applied to compute the posterior probability of a genre given a movie’s rating profile.
• The Naive Bayes classifier is employed, assuming conditional independence of rating values across users.
• Prior probabilities are estimated from the frequency of genres in the dataset, and conditional probabilities are derived from rating distributions per genre.
• Prediction is performed per rating value (1 to 5), and success is defined as matching any of the true genres if multiple are assigned.
• The model uses logarithmic probabilities to prevent numerical underflow and improve computational stability.

Experimental results

Research questions

• RQ1 Can user rating patterns alone be used to predict movie genres with high accuracy?
• RQ2 How does the size of the training set affect the genre prediction performance of the multinomial model?
• RQ3 Does the model perform better on certain rating levels (e.g., neutral or slightly positive) than others?
• RQ4 Can the model improve recommendation diversity by identifying unexpected but relevant genres?

Key findings

• The model achieves a 70% genre prediction rate when trained on just 15% of the MovieLens dataset.
• Even with only 5% of the data used for training, the model achieves a 62% prediction rate, significantly outperforming random guessing (30% accuracy).
• Prediction accuracy improves with larger training sets, but the rate of improvement diminishes beyond the initial 5% of data.
• The highest prediction accuracy is observed for ratings of 3 (neutral) and 4 (slightly like), which are the most frequent rating categories in the dataset.
• The lowest prediction accuracy is observed for ratings of 1 (strong dislike) and 2 (slightly dislike), likely due to lower data frequency and higher ambiguity.
• The model demonstrates that user rating profiles alone contain sufficient information to infer movie genres effectively, even with minimal training data.