A General Architecture for an Emotion-awareContent-based Recommender System
Fedelucio NarducciDept. of Computer Science
University of Bari ‘Aldo Moro’Italy
Marco De GemmisDept. of Computer Science
University of Bari ‘Aldo Moro’Italy
Pasquale LopsDept. of Computer Science
University of Bari ‘Aldo Moro’Italy
Vienna, Austria, 19th September 2015
outline• background and motivations
• general architecture for an emotion-aware content-based recommender system
• emotion analysis services
• experimental evaluation
• conclusions and future work
emotions & decision making• emotions influence the decision making process
during which, brain areas related to emotions are stimulated1
in the next few years… I will have a stable economic position,
I am getting married, I can buy a house
in the next months… my postdoc will be ended,
I will be out of work, I will beg, I can’t buy a house
1G. L. Clore, N. Schwarz, and M. Conway, “Affective causes and consequences of social information processing”, Handbook of social cognition, vol. 1, pp. 323-417, 1994.A. Bechara, “Risky business: emotion, decision-making, and addiction," Journal of Gambling Studies, vol. 19, no. 1, pp. 23-51, 2003.
emotions & recommendations
• “emotions are crucial for user’s decision making in recommendation process”1
• thanks to social networks, users disseminate data related to their emotions on the Web
• on April 2013, Facebook allows users to choose an emoticon to express their mood
1 G. Gonzalez, J. L. De La Rosa, M. Montaner, and S. Delfin, “Embedding emotional context in recommender systems”, in Data Engineering Workshop, 2007 IEEE 23rd International Conference on Data Engineering, pp. 845-852.
emotional models• discrete
basic emotions identified by labels
• dimensional emotion is a point in a multidimensional space
• componential emotions elicited by a cognitive evaluation of antecedent situations
a general architecture for a EA Content-based RS
Content Analyzer
Profile Learner
Recommender
Emotion
Analyzer
Item descriptions
Processed ItemsRateditems
Suggested Items
a general architecture for an EARS
Content Analyzer
Profile Learner
Recommender
Emotion
Analyzer
Item descriptions
Processed ItemsRateditems
Suggested Items
Analyzes unstructured text and performs NLP tasks on item descriptions and text associated to user emotional state
a general architecture for an EARS
Content Analyzer
Profile Learner
Recommender
Emotion
Analyzer
Item descriptions
Processed ItemsRateditems
Suggested Items
Generates a user profile. The user profile has two dimensions: preferences, emotion
a general architecture for an EARS
Content Analyzer
Profile Learner
Recommender
Emotion
Analyzer
Item descriptions
Processed ItemsRateditems
Suggested Items
Matches user profile and item representations. Both user profile and items are p r o v i d e d w i t h a n emotional label
a general architecture for an EARS
Content Analyzer
Profile Learner
Recommender
Emotion
Analyzer
Item descriptions
Processed ItemsRateditems
Suggested Items
Implements one or more s e n t i m e n t - a n a l y s i s algorithms able to assign emotional labels to a NL text
@work - emotion analysis• text classifiersthree different classifiers are learned on two distinct labelled datasets on the Ekman emotional model
• thesaurifor each emotion of the Ekman model a thesaurus is automatically generated by exploiting the WordNet synsets
two approaches combined by Borda count
synonym set
n timesseed
seed
experimental evaluation
• domain: music recommendation
• training datasets: LiveJournal1, Aman2
• music dataset: ~40,000 music tracks from Last.fm
• 578 songs evaluated by 77 users1https://snap.stanford.edu/data/soc-LiveJournal1.html2S. Aman and S. Szpakowicz, Identifying expressions of emotion in text, in Text, Speech and Dialogue. Springer, 2007, pp. 196-205.
recommendation approaches
• favoritetwo songs were randomly chosen from the set of tracks of the favorite artists (from Facebook), labeled with the user entry emotion
• not favoritetwo songs were randomly chosen from the set of tracks labeled with the user entry emotion, but not belonging to favorite artists
• random (baseline)two songs were randomly chosen by filtering out favorite artists and tracks labeled with the user entry emotion
research questions• RQ1: Is the defined algorithm able to effectively
extract an emotion from a NL text?
• RQ2: Is the emotion detection able to improve the user rating?
• RQ3: Is our model able to effectively associate an emotion to an item provided with an unstructured text?
user study• Users were asked to
• express her emotional state by a sentence and validate the emotional label automatically assigned by the system
• allow the extraction of her musical preferences from Facebook
• receive suggestions according to her emotional state or can choose a different one
• evaluate a set of recommendations by answering to two questions
results - emotion analysisEmotion # Precision Recall F1ANGER 8 0.25 0.50 0.33
DISGUST 2 1.00 0.50 0.67FEAR 7 0.43 0.43 0.43JOY 35 0.84 0.74 0.79
SADNESS 23 0.67 0.61 0.64SURPRISE 2 1.00 0.50 0.67
0
0,25
0,5
0,75
1
ANGER (8) DISGUST (2) FEAR (7) JOY (35) SADNESS (23)SURPRISE (2)
Precision Recall F1
resultsDo you like this song?
0
0,25
0,5
0,75
1
Favorite Not Favorite Random
YES/PART. NO
Is this song suitable with your emotion?
0
0,225
0,45
0,675
0,9
Favorite Not Favorite Random
YES/PART. NO
conclusions & future work• Contributions
• designing and testing a general architecture for an emotion-aware content based recsys
• implementing sentiment analysis services freely available online1
• implementing a prototypal music recommender system that exploits the proposed architecture and services2
• Future Work • testing new sentiment analysis, recommendation
algorithms, emotion models also in other domains1http://193.204.187.192:8080/MyEmotionsRest/webresources/service/getEmotion/<text>2http://193.204.187.192:8080/eMusic/
thanksFedelucio Narducci
Dept. of Computer ScienceUniversity of Bari ‘Aldo Moro’
Italy