Photo by Brooke Cagle on Unsplash

An introduction to group recommender systems

A recommender system is a decision support system that provides personalized recommendations by predicting the interests of the user. However, there are many instances where we socialize with people and may require recommendations for the whole group. In a few circumstances, the group members are unable to gather for face-to-face negotiation, or their preferences are not clear despite meeting each other. In such scenarios, the concept of a group recommender system is introduced.

A group recommender system analyses the interests of every member in the group and provides a final decision which would be accepted by all the members. Recommending to groups is a tedious task as we have to keep in mind each person’s likes and dislikes in the group before giving any recommendations. Here we take into consideration the user profiles of each member and then perform an aggregation over those profiles to arrive at a final decision.

There are numerous techniques for preference aggregation that is generally divided into three categories:

Majority-based strategies which use the most popular items for aggregation (e.g., Plurality Voting)

Consensus-based strategies where preferences of all group members are considered (e.g., Average, Average without Misery, Fairness)

Borderline strategies which only considers a subset (e.g., Dictatorship, Least Misery, Most Pleasure)

For understanding how these aggregation strategies are computed, let’s consider a small group of friends and their ratings for 10 items

Ratings for 10 items by Emma, Sara and Jay

Using the above data, I will be explaining below how aggregation techniques actually work —

Majority Based Strategies

1. Approval Voting : Only the ratings of an item above the approval threshold are counted. Assuming threshold = 5. So the rating for A will be 1 since only one rating is above the threshold level, whereas C’s rating will be 3.
2. Borda Count : Points are counted from the ratings given to an item in an individuals’ preference list, with the bottom item getting 0 points, the next one up getting one point, etc. Group rating for item F will be 17; 0 (last for Sara) + 9 (first for Emma) + 8 (shared top 3 for Jay)
3. Copeland Rule : The number of times the item beats other items (using majority voting) minus the number of times it loses. Group rating for I is 5; as it beats 7 items (A,B,D,E,G,H,J) and looses from 2 (C,F)
4. Plurality Voting : “First Past The Post” is used; the items with high votes are chosen every time. F will be chosen first as it exhibits the highest rating, followed by C (which has the highest rating for the majority when excluding F)

Consensus Based Strategies

1. Additive Utilitarian : Sums up the item rankings following the item with the highest summation is chosen.
   Item A →3+8+2 = 13
   Item C →10+7+9 = 26
2. Average : Takes the average of individual ratings. B will have a group rating of 7.67 → (8+9+6)/3
3. Average without misery : Takes an average of individual ratings after excluding the ratings below some assumed threshold level. Assuming threshold = 5. E will have a group rating of 7.3 (average of 6,9,7), while F is excluded because Sara hates it.
4. Fairness : Rankings will be given to an item as if users are choosing them one after the other. Item F may be chosen first (highest for Jay), followed by C (highest for Emma) and then B (highest for Sara).
5. Multiplicative : Multiplies all the users’ ratings. A will have a group rating of 48 (3*8*2).

Borderline strategies

1. Least Misery : Considers the minimum rating of an item. Group rating for A will be 2 since it’s the smallest amongst(3,8,2).
2. Most Pleasure : Considers the maximum rating of an item. Group rating for A will be 8 since it’s the largest amongst(3,8,2)
3. Most Respected Person (Dictatorship) : Considers ratings given by the most respected person in the group. If Jay is the most respected person, then A’s rating will be 2. Whereas if Sara is the most respected person, then the rating will be 8

Photo by Priscilla Du Preez on Unsplash

The satisfaction of all the users in the group is another important factor to be considered. If the recommender system constantly provides items that are disliked by a user, this may lead to user dissatisfaction and may cause disruption. A satisfaction function can be defined to measure the satisfaction of a user with a new item ‘i’ after having seen a sequence of items:

where the impact on the satisfaction of new item ‘i’, given existing satisfaction ‘s’ is defined as:

Parameter δ represents satisfaction decaying over time (with δ = 0, past items have no influence; with δ = 1, there is no decay). Parameter Ԑ represents the influence of the user’s satisfaction after experiencing previous items on the impact of a new item.

After computation of users’ satisfaction, it is now time to apply this model to the aggregation strategies so that the aggregation strategy could please the least satisfied user. There are several techniques for this, which include — Strongly Support Grumpiest strategy, Weakly Support Grumpiest strategy, and weighted strategy.

Knowing about the different personalities in a group helps in providing more accurate recommendations. It will be nice to include attributes such as demographic information, users’ role in the group, their personalities, relationship type, and strengths, etc.

Finally, I would conclude that most people socialize frequently rather than hanging out alone. There might be few introverts in the group who may not come up if they are dissatisfied with the decision. Therefore knowing the personalities of each member in the group and drawing recommendations which will be applauded by everyone could be a highly accepted approach for the recommender system.