Group recommendations are becoming prominent, since there are several applications in which users form groups for activities (e.g., tourism, restaurants, movies). SQUIRREL is a framework for sequential group recommendations, i.e., it may be used by a group in several rounds, providing a diferent recommendation in each round. The goal is that, after some rounds, all group members will be fairly satisfied with the recommendations, even if, individually, some users may not be very satisfied in every round. In this paper, we present SQUIRREL Movies, an implementation of SQUIRREL in the movies domain.
Group recommendations have become popular, mainly
due to the efortless forming of groups in social media [
A fair sequential group recommender should ofer good
results for the current round and also ensure the group
members’ satisfaction over a sequence of
recommendations, i.e., the recommender should observe that: ( ) the
recommended items should be as relevant as possible
to the group members at each round, and ( ) no group
member is dissatisfied after a sequence of rounds.
Standard group recommendation approaches fall short of
achieving both sequential group recommendation
objectives [
The SQUIRREL framework [
We apply SQUIRREL for a sequence of rounds to pro- the group the generated group recommendation list .
Let denote a set of movies, a set of users, and ⊆ a group of users. For each user ∈ , denotes an ordered list of recommended movies that have been generated by a single recommender system at a specific recommendation round for . At round , SQUIRREL chooses an appropriate aggregation function based on the current state of the group, to combine the individual members’ recommendation lists into one list . as ℛ = ( duce group recommendation lists for a group , defined 1 , 2 , ...,
). At each round, we calculate each group member’s individual satisfaction score, i.e., how relevant the items in the group list are for each member, as well as the aggregated group satisfaction score. Additionally, we calculate the disagreement between the group members, i.e., the variance between their satisfaction scores.
SQUIRREL can be defined as a Markov decision process (, A, , )
, aiming to maximize the accumulative
reward after each recommendation round , where:
• S describes the group state, expressed via the
satisfaction scores of the group members. We keep an individual
state for each group member at each round , defined
as
= (, )
, where
is the overall satisfaction
of user at round . Briefly, it describes how relevant
are the data items in the group recommendation list
compared to the best case scenario for the user , which
,
is their individual recommendations .
• A is a set of distinct actions, i.e., diferent aggregation
functions employed by the SQUIRREL model. SQUIRREL
employs 6 aggregation functions (see [
′ to transition from state to state = , = ) is the probability ′ during round , under the action . • (, ′) is the reward gained from transitioning from state to state ′. We define two reward functions. First, we examine only the overall satisfaction of the group by averaging the individual satisfaction of all group members. Second, in addition to the overall satisfaction, we also consider the disagreement between group member, defined as the diference in the satisfaction scores between the most and least satisfied group member.
The goal of the model is to find a policy (|) that takes action ∈
during state ∈ to maximize the expected discounted cumulative reward after rounds: (1) where with 0 ≤ ≤ 1 .
[()], =0 () = ∑ (, ′),
At the beginning of round , the group is given to a single-user recommender system that produces a recommendation list for each group member . These lists are then given to the SQUIRREL model, where the agent observes the state of the environment and selects an appropriate action to aggregate the lists . This results in the transitioning of the model to the next state +1 , where we update the overall satisfaction of the users and calculate the reward +1 . Finally, SQUIRREL returns to
In addition to the features described above and a Web UI, we have added some new functionalities to SQUIRREL Movies, such as recording the state of each group at each recommendation round, retrieving external movie data from MovieLens using MovieLens id, and providing userfriendly visual and textual explanations regarding the group recommendations. We cover only the latter next.
Providing explanations. SQUIRREL Movies
provides both textual and visual (through graphs)
explanations of the recommendations returned both for each
round, as well as across rounds. For example, a user can
change the reward function and get a brief, intuitive
explanation of each aggregation function. In addition to
textual explanations, the user can be presented a line
chart with all the individual group members satisfaction
not only for the current round, but also for the
previous rounds. Further visual explanations include several
bar charts with diferent group fairness measures (e.g.,
MaxMin) [
Dataset and code. We have utilized the 20M
MovieLens Dataset [
The data are available on https://github. com/mariaStratigi/SQUIRREL and the code of SQUIRREL Movies is publicly available on https://github.com/
We have presented SQUIRREL Movies, demonstrating the adaptation of the SQUIRREL sequential group recommendations using reinforcement learning in the movies domain. SQUIRREL Movies provides visual and textual explanations, justifying its recommendations, and retrieves movie data from external sources, to help users decide on (2) the movie to watch next. SQUIRREL Movies is just one of the many possible domain adaptations of SQUIRREL.
Acknowledgement. This work has received funding from the Hellenic Foundation for Research and Innovation (HFRI) and the General Secretariat for Research and Technology (GSRT), under grant agreement No 969.