[WI 2017] Affective Prediction By Collaborative Chains In Movie Recommendation

Affective Prediction By Collaborative Chains In Movie Recommendation

Yong ZhengSchool of Applied Technology

Illinois Institute of TechnologyChicago, IL, 60616, USA

The 2017 IEEE/WIC/ACM Conference on Web Intelligence (WI)August 23-26, 2017, Leipzig, Germany

Agenda

• Background and Introduction

– Context-aware Recommender Systems

– Emotions In Recommender Systems

• Research Problems

– Emotion Acquisition

– Affective Predictions

• Methodologies and Results

• Conclusions and Future Work

2

Agenda

• Background and Introduction

– Context-aware Recommender Systems

– Emotions In Recommender Systems

• Research Problems

– Emotion Acquisition

– Affective Predictions

• Methodologies and Results

• Conclusions and Future Work

3

Recommender System (RS)

• RS: item recommendations tailored to user tastes

4

Context-Aware Recommendation

5

Companion

User’s decision may vary from contexts to contexts

• Examples:➢ Travel destination: in winter vs in summer

➢ Movie watching: with children vs with partner

➢ Restaurant: quick lunch vs business dinner

➢ Music: for workout vs for study

Terminology in CARS

6

• Example of Multi-dimensional Context-aware Data set

➢Context Dimension: time, location, companion

➢Context Condition: Weekend/Weekday, Home/Cinema

➢Context Situation: {Weekend, Home, Kids}

User Item Rating Time Location Companion

U1 T1 3 Weekend Home Kids

U1 T2 5 Weekday Home Partner

U2 T2 2 Weekend Cinema Partner

U2 T3 3 Weekday Cinema Family

U1 T3 ? Weekend Cinema Kids

What is Context?

7

The most common contextual variables:

➢Time and Location

➢User intent or purpose

➢User emotional states

➢Devices

➢Topics of interests, e.g., apple vs. Apple

➢Others: companion, weather, budget, etc

Usually, the selection/definition of contexts is a domain-specific problem

Emotions and Emotional Effects

8

Incorporate Emotional Effects into RecSys

9

• Marko Tkalcic, Andrej Kosir, and Jurij Tasic. 2011. Affective recommender systems: the role of emotions in recommender systems. In Proc. The RecSys 2011 Workshop on Human Decision Making in Recommender Systems. ACM, 9–13

• Ante Odic, Marko Tkalcic, Jurij F Tasic, and Andrej Košir. 2012. Relevant context in a movie recommender system: Users' opinion vs. statistical detection. ACM RecSys 12 (2012)

• Yue Shi, Martha Larson, and Alan Hanjalic. 2013. Mining contextual movie similarity with matrix factorization for context-aware recommendation. ACM Transactions on Intelligent Systems and Technology (TIST) 4, 1 (2013), 16.

• Yong Zheng, Bamshad Mobasher, and Robin Burke. 2016. Emotions in context-aware recommender systems. In Emotions and Personality in Personalized Services. Springer, 311–326

• Yong Zheng. 2016. Adapt to Emotional Reactions In Context-aware Personalization. In 4th Workshop on Emotions and Personality in Personalized Systems (EMPIRE) 2016 co-located with ACM RecSys 2016

Agenda

• Background and Introduction

– Context-aware Recommender Systems

– Emotions In Recommender Systems

• Research Problems

– Emotion Acquisition

– Affective Predictions

• Methodologies and Results

• Conclusions and Future Work

10

Emotion Acquisition

11

We can collect emotions

➢By user surveys

➢By special user interactions, such as emoji

➢By Emotion Recognition or Extraction, e.g., from texts, voice, facial expressions, etc

➢By Affective Prediction – a learning process to predict emotional states from limited knowledge at hand

Challenges in Affective Prediction

12

Emotional expression may happen in different stages

Challenges in Affective Prediction

13

There are correlations between emotions in two stages. For example, a user may feel sad before watching a movie. He may be dissatisfied with the movie and leave a negative reaction after the movie watching

Research Problems

14

We focus on the following problems:

➢How to better predict affective states

➢How to take emotion correlations into account

Agenda

• Background and Introduction

– Context-aware Recommender Systems

– Emotions In Recommender Systems

• Research Problems

– Emotion Acquisition

– Affective Predictions

• Methodologies and Results

• Conclusions and Future Work

15

LDOS-CoMoDa Movie Data Set

16

There are 2291 ratings given by 121 users on 1232 movies. There are 12 contextual dimensions

1. Independent Emotion Classification (IEC)

17

The problem is viewed as a classification problem

➢Features: user info and item features

➢Label(s): emotional variables

We use a binary classification algorithm to predict the binary value for each emotional variable independently.

2. Dependent Emotion Classification (DEC)

18

For example, Classification Chains

➢Features: user info and item features

➢Label(s): emotional variables

3. Independent Collaborative Prediction (ICP)

19

We choose collaborative filtering as the predictive model, since it may work better on personalization than the classification.

We select one-class matrix factorization with side information as the model in our experiments.• Yi Fang and Luo Si. 2011. Matrix co-factorization for

recommendation with rich side information and implicit feedback. In Proceedings of the 2nd Workshop on Information Heterogeneity and Fusion in Recommender Systems. ACM, 65–69

4. Dependent Collaborative Chains (DCC)

20

We select one-class matrix factorization with side information as the model in our experiments.

Experimental Settings

21

➢We use the LDOS-CoMoDa movie rating data

➢5-fold cross validation is applied

➢We predict the emotions for the test set first, and examine the accuracy of the predictions

➢The predicted emotions will be incorporated into one context-aware recommendation models to examine the quality of context-aware recommendations.

Quality of the Affective Predictions

22

Quality of the Context-aware Recommendations

23

• Yong Zheng. 2016. Adapt to Emotional Reactions In Context-aware Personalization. In 4th Workshop on Emotions and Personality in Personalized Systems (EMPIRE) 2016 co-located with ACM RecSys 2016 [ the recommendation model used in the paper]

• Actual the performance when we use the actual emotions• Predicted the performance when we use the predicted emotions

Agenda

• Background and Introduction

– Context-aware Recommender Systems

– Emotions In Recommender Systems

• Research Problems

– Emotion Acquisition

– Affective Predictions

• Methodologies and Results

• Conclusions and Future Work

24

Conclusions

25

➢We explore the affective predictions

➢We predict the emotions by classification and collaborative filtering respectively

➢For each solution, we figure out a way to incorporate correlations among emotions

➢Collaborative predictions can help improve the quality of personalizations

➢The dependent collaborative chains is demonstrated as the best predictive model

➢The predicted emotional states can also help obtain good context-aware recommendations.

Future Work

26

➢We plan to evaluate the proposed models in other domains rather than the movie domain only

➢The problem of affective prediction is closely related to a novel research topic – context suggestion, where we predict or recommend appropriate contexts to the end users.

➢In our future work, we will try to utilize the context suggestion as solutions to help predict the emotional states

Affective Prediction By Collaborative Chains In Movie Recommendation

Yong ZhengSchool of Applied Technology

Illinois Institute of TechnologyChicago, IL, 60616, USA

The 2017 IEEE/WIC/ACM Conference on Web Intelligence (WI)August 23-26, 2017, Leipzig, Germany