PIIS_SimilarityFunction

This program was made for the Programming of Inteligent Information Systems course as part of the movie reccomendation system project.

The general idea behind it is to check how two movies are similar to each other. The similarity is expressed as the number from 0 do 1, where 0 is completely different and 1 is identical.

Instructions

The desired format of the .txt / .csv file is as follows: [RecordId];[MovieId];[FeatureId];[FeatureValue] where [FeatureValue] can store multiple values in the following form: [featureValue1]/[featureValue2]/[featureValue3].

The following features are hard coded in the application:

FeatureId	FeatureName	FeatureDescription	Weight
1	CollectionId	If the movie belongs to a collection, it's ID is put here. If no the value is null.	10
2	Budget	Budget in $.	4
3	GenresIds	Movie genres seperated by "/",	9
4	Popularity	Popularity of the movie.	5
5	ProductionCompaniesIds	Movie production companies separated by the "/".	5
6	ProductionCountriesIso	Movie production counries codes separated by the "/".	6
7	ReleaseDate	Movie release date.	7
8	Revenue	Revenue in $.	5
9	Runtime	Runtime in minutes.	2
10	SpokenLanguagesIso	Movie spoken languages codes separated by the "/".	8
11	VoteAverage	Movie average vote.	5
12	VoteCount	Number of votes.	4
13	CastIds	Ids of 10 people from movie cast separated by the "/". Starting from the most important characters.	9
14	Costume&Make-UpIds	Ids of people from movie Costume & Make-Up department separated by the "/".	3
15	DirectingIds	Ids of people from movie Directing department separated by the "/".	9
16	CameraIds	Ids of people from movie Camera department separated by the "/".	4
17	EditingIds	Ids of people from movie Editing department separated by the "/".	2
18	ProductionIds	Ids of people from movie Production department separated by the "/".	4
19	SoundIds	Ids of people from movie Sound department separated by the "/".	2
20	WritingIds	Ids of people from movie Writing department separated by the "/".	6
21	ArtIds	Ids of people from movie Art department separated by the "/".	2
22	CrewIds	Ids of people from movie Crew department separated by the "/".	1
23	VisualEffectsIds	Ids of people from movie Visual Effects department separated by the "/".	3

Their values have been taken from The Movie Database (TMDb) using WatTMDb API wrapper library.

Weights were assigned to each feature in order to make some of the features more influencial when calculating the general similarity rate (weighted average) of the movie. This assignment was subjective and it is possible that it will change during the project development.

Algorithm

1. The similarity rate for each separate feature is calculated.

Features can be grouped basing on the algorithms used for calculating the similarity rates for those features:

• CollectionId

  If both movies belongs to the same collection or the don't belong to any collection, the similarity rate equals 1.
  If they belong to different collections, the similarity rate equals 0.5.
  If one of them belongs to some collection and the second one doesn't belong to any, the similarity rate equals 0.

• Budget, Popularity, ReleaseDate, Revenue, Runtime, VoteAverage, VoteCount

  The following equation is used for calculating similarity rate for those features:

  similarityRate = 1 - (Math.abs(x - y)/(max - min))

  where:
  x - feature value of the first movie
  y - feature value of the second movie
  min - smallest feature value from the whole set of movies
  max - biggest feature value from the whole set of movies

  As for now, it has been assumed that if there is no information for any of the movie feature, similarity rate equals 0.5 (there is 50/50 chance).

• GenresIds, ProductionCompaniesIds, ProductionCountriesIso, SpokenLanguagesIso, CastIds, Costume&Make-UpIds, DirectingIds, CameraIds, EditingIds, ProductionIds, SoundIds, WritingIds, ArtIds, CrewIds, VisualEffectsIds

  Feature sets of the two movies are compared in order to check how many of its elements are the same. The similarity rate is calculated using the following formula:

  similarityRate = (numberOfMatchingElements * 2) / (set1.length + set2.length)

  where:
  set1.lenght - number of values for the specific feature (e.g. a movie can be assigned to many genres) of the first movie
  set2.lenght - number of values for the specific feature (e.g. a movie can be assigned to many genres) of the second movie

  As for now, it has been assumed that if there is no information for any of the movie feature, similarity rate equals 0.5 (there is 50/50 chance).

2. The general similarity rate for all of the features is calculated.

In order to find the general similarity rate, the weighted average of the separate features similarity rates is calcuated. Weights needed for that calculation can be found in the table above.

Future plans

• Improvement of the similarity rate for genre calculation. As for now, only 1:1 similarity is approved, so thriller won't be treated as similar to the action movie, which is not entirely true.
• Improvement of the similarity rate for release date calculation. As for now, only 1:1 similarity is approved, so movie from 1952 won't be treated as similar to the movie from 1955, which is not entirely true, because very often there is a big resemblance between movies from the same decades.
• Omitting features for which the data has not been provided. Assuming that similarity rate in that case is equal 0.5 can distort the results in many cases.