Source-linked AI summary
Movie Recommendation System using Sentiment Analysis from Microblogging Data
Sudhanshu Kumar, Shirsendu Sukanta Halder, Kanjar De, Partha Pratim Roy
TL;DR
Movie recommendation systems traditionally rely on collaborative or content-based filtering and may require prior user history and habits. This paper combines both approaches with sentiment analysis of movie tweets to capture current audience response, achieving higher average precision than comparison models in Top-5 and Top-10 recommendations. The evaluation uses a modified MovieTweetings database selected because other public databases lacked suitable current microblogging data.
Problem
Traditional movie recommendation approaches can depend on users’ prior history and habits, motivating methods that reduce these dependencies.
Method
The paper combines collaborative filtering and content-based filtering with sentiment analysis of movie tweets, using movie metadata and a social graph.
Results
The proposed model achieves average precision values of 2.54 for Top-5 and 4.97 for Top-10, exceeding the comparison models.
Takeaways & Limitations
Combining sentiment information with hybrid recommendation produces more precise movie recommendations than the other evaluated models.
Takeaways & Limitations
The evaluation uses MovieTweetings after other public databases were judged unsuitable because appropriate microblogging data was unavailable, and movies with almost no TMDb metadata were discarded.
Abstract
from arXiv · showhide
Recommendation systems are important intelligent systems that play a vital role in providing selective information to users. Traditional approaches in recommendation systems include collaborative filtering and content-based filtering. However, these approaches have certain limitations like the necessity of prior user history and habits for performing the task of recommendation. In order to reduce the effect of such dependencies, this paper proposes a hybrid recommendation system which combines the collaborative filtering, content-based filtering with sentiment analysis of movie tweets. The movie tweets have been collected from microblogging websites to understand the current trends and user response of the movie. Experiments conducted on public database produce promising results.
1 Introduction
The paper motivates movie recommendation as a way to filter excessive information and satisfy user preferences. It proposes combining collaborative filtering, content-based filtering, and sentiment analysis of tweets, followed by comparative experiments.
- Recommendation systems help users find preferred movies and reduce the time required to search through abundant online information.
- Collaborative filtering uses ratings from like-minded users, whereas content-based filtering recommends items through similarity among their content information.
- Twitter provides limited-character user-generated information that can reveal people’s thoughts and responses to movies.
- The paper proposes a hybrid movie recommendation system that combines collaborative filtering and content-based filtering, with sentiment analysis used to boost recommendations.
- The study evaluates the proposed system through extensive experiments and qualitative and quantitative comparisons with baseline models.
2 Related work
Prior work covers collaborative, content-based, hybrid, and sentiment-based recommendation approaches. The paper positions its method within research showing that hybrid systems combine techniques to mitigate individual drawbacks, while sentiment analysis captures opinions.
- Recommendation research includes collaborative filtering, content-based filtering, hybrid systems, and sentiment analysis for recommending items.
- Prior collaborative-filtering studies address issues including cold start, scalability, clustering, and prediction accuracy through implicit ratings or optimization methods.
- Content-based filtering recommends items using item descriptions and user-preference profiles.
- Hybrid systems combine multiple recommendation techniques to mitigate the drawbacks of individual techniques.
- Sentiment analysis is used to acquire people’s opinions, including through VADER-based classification of tweets into positive, neutral, and negative classes.
- The proposed model is a hybrid recommender whose results are boosted with sentiment-analysis scores and evaluated quantitatively and qualitatively.
3 Proposed system
The proposed system is a sentiment-based movie recommendation framework. Its methodology is presented as a sequence of steps and components illustrated in the proposed framework figure.
- The proposed sentiment-based recommendation system is presented as a movie recommendation framework in Fig. 1.
- The methodology section describes the different steps and components of the proposed recommender system.
3.1 Dataset description
The system combines a user-rated movie database with Twitter data, selecting recent movies and enriching MovieTweetings metadata for hybrid recommendation. The modified dataset retains 2014–2017 movies and integrates ratings, genres, and movie attributes.
- The proposed system uses a user-rated movies database together with users’ movie tweets from Twitter.
- 3.1.1 Public database: MovieTweetings was selected because older public databases lacked suitable microblogging data, while MovieTweetings provides more up-to-date social-media ratings.
- 3.1.2 Modified MovieTweetings database: Only movies released after 2014 were retained because older movies had too few tweets for sentiment analysis.
- 3.1.2 Modified MovieTweetings database: The modified database contains 292863 ratings from 51081 users across 6209 movies, with the study using movies released from 2014–2017.
- 3.1.2 Modified MovieTweetings database: Because MovieTweetings provides ratings and genres but limited metadata, the system combines collaborative filtering with movie-attribute similarity obtained through the TMDb API.
- 3.1.2 Modified MovieTweetings database: Movies with almost no TMDb metadata were discarded, leaving a final database of around 4500 movies.
3.2 Analysis of user tweets
The tweet-analysis pipeline fetches movie tweets, removes noisy content, and applies VADER sentiment analysis. VADER produces sentiment components and converts its normalized compound score into a 1–10 movie-rating scale.
- 3.2.1 Preprocessing of tweets: Tweets were fetched through the Twitter API and cleaned by removing hashtags, emojis, repetitive words, and other irrelevant data.
- 3.2.1 Preprocessing of tweets: Preprocessing removes stop words, punctuation, web links, special characters, and repetitive words before sentiment analysis and recommendation construction.
- 3.2.2 Sentiment analysis of user tweets: VADER is a computationally efficient lexicon-based, rule-based sentiment method that evaluates words, phrases, and tweets.
- 3.2.2 Sentiment analysis of user tweets: VADER outputs positive, neutral, negative, and compound sentiment components, with the compound score normalized from -1 to 1.
- 3.2.2 Sentiment analysis of user tweets: The normalized compound score is scaled to a 1–10 movie-rating range using Equation (2), where x denotes the compound score.
3.3 Hybrid recommendation
The hybrid recommendation model combines content-based similarity with collaborative social filtering. It learns metadata-feature weights from item similarity and a sparse User-Item social graph.
- The model combines content-based similarity features with collaborative social filtering to generate recommendations.
- Feature closeness is computed from metadata similarity and combined into a feature matrix using learned weight vectors.
- Collaborative filtering uses neighboring users' ratings and a tweaked User-Item matrix to construct an item-based social graph.
- The weight vector is optimized from item and user similarity information using the Moore-Penrose pseudoinverse.
3.4 Weighted score fusion
The system fuses hybrid-model similarity with sentiment similarity derived from movie tweets. The two components are combined through tunable weights.
- Tweet-derived sentiment ratings are used to compute sentiment similarity between movies.The sentiment ratings are calculated for all movies from retrieved user tweets.
- The final combined similarity is a weighted sum of hybrid similarity and sentiment similarity.
- ω1 weights the hybrid-model similarity score, while ω2 weights the sentiment similarity score.
4 Experimental results and analysis
Experiments evaluate sentiment–rating correlations, fusion-weight selection, baseline comparisons, and qualitative recommendation overlap. The proposed model achieves higher reported precision than the two baselines and shows overlap with IMDb and TMDb recommendations.
- 4.1 Correlation between sentiment and IMDb movie ratings: The study reports correlation analysis between sentiment ratings and IMDb movie ratings using SROCC, KRCC, and PLCC.
- 4.2 Evaluation metric: Precision@5 and Precision@10 are used to evaluate Top-N recommendations rather than RMSE or MSE.
- 4.3 Weight selection for weighted fusion: Maximum precision occurs when the fusion weights are between 0.5 and 0.6, leading to ω1 = ω2 = 0.5.
- 4.4 Comparative analysis: 2.54 for Top-5 and 4.97 for Top-10 exceed the Pure Hybrid Model and Sentiment Similarity model precision values.The baseline values are 1.86 and 3.31 for the Pure Hybrid Model, and 0.54 and 1.04 for the Sentiment Similarity model.
- 4.5 Qualitative analysis: Qualitative results show intersecting recommendations with IMDb and TMDb for Hollywood and Bollywood movies.
5 Conclusion and future works
The paper proposes movie recommendation using Twitter sentiment alongside movie metadata and a social graph. Its reported precision is higher than the sentiment-only and hybrid alternatives, while future work targets richer emotion signals and dynamic weights.
- The proposed system combines Twitter sentiment, movie metadata, and a social graph for movie recommendation.
- 2.54 Top-5 and 4.97 Top-10 average precision exceed the corresponding sentiment-similarity and hybrid-model results.
- Future work includes extracting emotional tone from additional social-media platforms and updating social-graph weights in real time.