Websites like Netflix and HBO Go, provide lists of movies based on genres, this makes it easier for user to select the movie that interests him/her based on the genre he/she is more inclined towards. Tagging of movies is a complex process and usually involves a manual process where the movies are assigned to one or more Genres based on the suggestions sent by the users and consumers. If we can automate this process of movie tagging, not only will it be fast , save human effort but it will be more accurate than an untrained human as well.

Predicting movie Genres is a multilabel classification problem where a a movie can be tagged to more than one Genres.For the sake of simplicity, we will be predict whether a movie belongs to a particular Genre or not (one-vs-rest approach) by the end of this project using various data mining and machine learning processes. We will collect data using one of many available apis on internet and compile a data set which will be primarily based on IMDB data. We will rely on text analysis of the Plot/Summary of the movie data collected and train our classifier using text analysis techniques.

There is paper published by Ka-Wing Ho called “Movies’ Genres Classification by Synopsis” which investigated four different approaches for automated movies’ genres classification based on their synopsis. The methods used in the paper are SVM, Multi-label K-nearest neighbor, Parametric Mixture Model and Neural network. I have based some of my ideas from this paper whereas the dataset I have used is completely created by me and is fetched from the internet using a web-api.

I used a web api to build my data set. The name of the web api that I used is OMDB api as shown in Figure 1. This api is capable of returning all the the data about a movie/series in a json/xml format given the imdb id of the movie. I started with a seed title ID and kept on increamenting it until I reached the target number of movies.

Although the api returns a lot of data about the movies/series, I just selected some features out of the whole set which are as follows :

• Movie ID
• Plot : Full (choices were Short or Full)
• Genre

There are movies which are associated with more than one Genre and there can be as much as 4 Genres. I selected the first one as the main Genre and based my findings on that.

The process to cache all the data into a CSV file had a logic where movies were pruned and only those titles were selected which :

• Were of ‘Movie’ type
• Length of plot string was more than 50
• Genre is not equal to ‘N/A’ (there were some cases with this value)
• Language is ‘English’

Here are some numbers for my data:

• Number of titles collected : 4456
• Genres Collected : 19

The distibution of movies towards different genres is shown in figure 2. As you can see that the number of comedy records is way more than any other genre, hence we will use comedy as our main genre in our one-vs-all approach. Note that comedy movies constitute for about 25 percent of the movies in the total set, therefore , if we just have a naive classifier which classifies all the movies as comedy movies, we will still have an accuracy of 25 percent of the total genres. We have to do a lot better than that in order to produce useful trends.

After all the data was collected, some preprocessing was done on it:

• Any/all use of HTML tags were removed using BeautifulSoup library.
• Punctuation, Numbers and special characters were removed using a regex find and replace.
• All the stop words which were irrelevant to the context were removed using stopwords from Natural Language Toolkit (NLTK)

I used a 90-10 train-test split ratio for test and train sets respectively. After the split, the number of testing records were found to be : 446. And its distribution is shown in figure 3.

Note that the distribution is shown just for comedy records vs other records since we are using one-vs-rest approach where we will test comedy records against all the other records.

All the preprocessed plot strings were then converted to bag of Words(bow) form and various Machine learning methods were applied.

I used scikit-learn’s pyhton bindings to apply these methods, the results are as shown below:

Given a set of training examples, each marked as belonging to one or the other of two categories, an SVM training algorithm builds a model that assigns new examples to one category or the other, making it a non-probabilistic binary linear classifier. An SVM model is a representation of the examples as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible

Results:

Accuracy 0.675 Precision 0.5 Recall 0.489 F Measure 0.495

Logistic Regression is a type of regression model where the dependent variable (target) has just two values

Results:

Accuracy 0.657 Precision 0.471 Recall 0.462 F Measure 0.467

K nearest neighbors is a simple algorithm that stores all available cases and classifies new cases based on a similarity measure (e.g., distance functions)

Results:

Accuracy 0.639 Precision 0.431 Recall 0.344 F Measure 0.383

Although the results were fairly good for the earlier methods but still the accuracy was not jumping the 70 percent mark and hence I decided to introduce another preprocessing step.

To further improve the results, the BOW representation was enhanced by computing TFIDF (Term frequency inverse document frequency). TFIDF not only takes into account the frequency of each word in the vocabulary but also the weight associated with each word in the vocabulary to further improve the results. It is given by the formula :

where, wki is the frequncy of the k-th word in the i-th movie’s synopsis (di) and m is the number of training/test sets.

After applying tfidf, there was a significant jump in the efficiency of the algorithms, I could see not only increase in terms of accuracy but also there was an improvement in the precision, recall and f-measure values. Here are the results :

Accuracy 0.758 Precision 0.694 Recall 0.455 F Measure 0.55

There was also an increase in the performance for logistic regression with tfidf,

Results:

Accuracy 0.733 Precision 0.614 Recall 0.482 F Measure 0.54

KNN almost perfomed the same even with tfidf and there was no improvement in the results.

Results:

Accuracy 0.639 Precision 0.42 Recall 0.324 F Measure 0.368

Here is a summary of all the results:

As you can see from this table, the best accuracy was achieved by using SVM with tfidf and an accuracy of 75.8 was achieved.

Since right now, the data set is not that large because I just created this as a prototype for this approach, therefore this data set can be increased to have more variation in genres which can improve the results. Also more data preprocessing steps such as data stemming and word repetitions can be implemented to increase accuracy of the model. We can also take into consideration multiple genres associated with a single movie as an enchancement to the current method of just using the main genre and also expand the current methods from (one-vs-rest, here ‘Comedy’ vs ‘Others’) to include more genres in the future.