Group recommendations

4 minute read

Updated:

What if the user is a group?

  • We have to recommend for a group of people → one recommendation
  • Movie, tv program, restaurant, experience, house, TV, holiday to a GROUP

Main steps for group recommendations

  1. Obtain individual preferences – need a user profile/ model for each individual user
    • Content based – need to look at the parameters and put this within the user model
    • Social (collaborative) – we look for user’s interest and opinions about particular items
  2. Aggregation strategies – to aggregate and come up with one recommendation → focus of the lec
  3. Present the recommendations to the group with appropriate justification

Example

  • We have 4 users and 10 items
  • For each of the user, we have preferences for each item – 1 to 20
  • Screenshot 2020-11-23 at 12 50 50 pm
  • Q - How would you select the best item?
    • A – with the highest score amongst many users

Aggregation strategies: Average

  • Take the average for each item and take the item with the highest average
  • If we want to recommend more than one, we take it in order
  • 10 /20 is a nice compromise for the group
  • Disadvantage – the users will be disadvantaged bc strong values tend to diminish the bad choices – like 15 and 20

Aggregation strategies: Multiplicative

  • Multiple individual scores
  • Multiply all the user rating and select the highest
  • image

Aggregation strategies: Approval voting

  • Counts scores above a certain threshold, seen as approval value
  • Threshold value above > and we count how many
  • So for item 1=1, item 2 = 0, item3 = 3

Aggregation strategies: Plurality voting

  • Choose the item with highest number of votes. → per user!!
    • e.g. So u1 and u4 voted item 9 -> got 2 votes, hence chosen
  • Skewed - few users and few items. And also low scores are observed by the strong ones
  • image → item 9 !!

Aggregation strategies: Least Misery

  • Considers the minimum scores for each item, and chooses the item with the maximum (minimum-scored) value (least miserable)
  • LM looks at the lowest item, and out of those chooses the least miserable
  • User will fall into disadvantage bc its with majority
  • !image → 5 is the best choice
    • User 1 is the most unhappy, but won’t be toooo bad
  • Can be combined with other strategies e.g. average – its easy but it can disadvantage people. However, if we do average without misery, then we take everything below 4, and item 5 will have stronger confidence now

Aggregation strategies: Most pleasure

  • Considers the maximum of the scores for each item, and chooses the item with the maximum value – most pleasure
  • So item 4 (rating = 20)

Aggregation strategies: Borda

  • To ignore actual values and start looking at the ranking
  • Considers item ranking in the individual preferences, then calculates the item ranking for the group
  • image
  • We rank the item in order and give the value starting from 1-10. Highest gets 10 and lowest gets 1
  • Then we add all the items and see what the highest is

Aggregation strategies: Copeland Rule

  • Counts how often an item beats another item (majority voting) and how many time loses – takes the difference as the value
  • → We compare the item1 to every other items in both sides of the column
  • image
    • e.g. Item 2 – 5,3,6,5. And we compare this to item 1 (1,3,4,9) & item 3 (3,7,15,8)
    • 5 beats 1, 6 beats 4, 5 beats 3 -> so it in total, beat = 3
  • beats – loses = copeland. Then choose item with highest copeland value.

Aggregation strategies: Heuristics

  • So far, item 5 seems the strongest as it was recommended by most of the aggregation strategies
    • It won’t make people too miserable, and it gives everybody a reasonable choice assuming that user 3 isn’t the dominant user (authority)
  • Fairness: take the best option for everyone
  • Authority: take the best option for the most influential user

How do we know which is the best option - Evaluation

  • You recommend and see how individual members of the group feels about it.
  • Reading – when people was given the recommendation for the group and asked how satisfied they were, and what their friends thought
  • Two keys factor of people’s satisfaction with the group recommendations
    • Affective state – am I happy or grumpy – people take the recommendation based on their mood and like/ dislike
    • Relationships with other users -

Impact of relationships on accepting group recommendations

  • Several relationships that it might affect
  • You are not 100% satisfied, but okay with the rec
    • Communal sharing -Best friend – if you have to compromise for someone you like, you’d be happy
    • Authority ranking - High respect – if the recommendation was chosen by the most influential user and that person has priority, they take it even though its not their choice
  • Unhappy and more sensitive about not being taken into account
    • Market pricing - Someone you compete with – will be fighting for fairness
    • Equality matching - If you think the rec was unequal
  • → When you explain the final choice, need to bear in mind the relationship between people

Summary

  • Obtain individual preferences
  • Aggregation strategies → key !!
    • Most system use hybridisation of these strategies
    • Evaluation is important and need to tune the strategies or heuristics → will be able to explain better
    • Opinions will be influenced by affect and relationships
  • Presenting the recommendations to the group

Leave a comment