Situation-dependent combination of long-term and session-based Preferences in Group recommendations:an experimental Analy SIS----Group recommendations based on long-term and conversational preferences for contextual dependency combinations

First, Summary:

background: One of the main challenges of the conversation group referral system is how to make appropriate use of interaction between group members to generate user preferences, which may deviate from users ' long-term preferences. The relative importance of long-term preferences and Group-induced preferences should vary according to the specific group settings.

This article: through the experiment, the conclusion: When group discussion has no influence on group members ' preferences, long-term preference occupies greater weight. Group-induced preference has a greater weight when the group context encourages members to have more or less similar preferences.

Second, Introduction:

background: traditional recommender systems focus on personalized recommendations, but there are many scenarios that need to meet a set of user needs. For example, a group of friends or a family needs to find a restaurant, which leads to the development of a group referral system.

problem: The group recommendation system has made some progress in improving the quality of recommendation, but the Dynamics of group decision-making process has not been fully explored. In fact, most studies focus on the method of merging the static preferences of group members, ignoring the behavior of users in a particular group context, and ignoring changes in user preferences, "These changes often occur in the group decision-making process"

This paper proposes a session group re-correction model which takes into account the individual's long-term preference "scored by Project Score", and also takes into account the user's direct feedback to the project during group discussions "which reflects the user's current needs". The model is implemented in a group recommendation system that provides a chat environment, which integrates various decision support and re-notification functions.

There may be different social manifestations of group members facing group situations in the system. For example, group members can express their views according to their own ideas, or they can change their views to accept the influence of others, or they take action contrary to the team's advice.

The purpose of this article is to study how to properly combine long-term preferences and session-specific preferences in these scenarios.

Three social Impacts: (a) Independence-groups have no effect on user preferences. (b) Conversion-group-driven members have more similar preferences. (c) Anti-consistency-groups give members more different preferences.

Three variants of the Preference Portfolio strategy: (i) the importance of long-term and session-based preference is equal (ii) long-term preference is more important (iii) Conversational preference is more important.

Third, group recommendation

In previous models, the group referral system gained long-term preference through project scoring, but in group discussions, group members might deviate from the preferences they previously observed. This may be due to the impact of other group members and group decision Dynamics.

Therefore, the user's preference model needs to be generated and continuously updated using two preferences.

• The Member preference model is represented by a function,
• The group preference model is established by aggregating the utility functions of the group members.
• Sort the group referral items according to the set preference.

the preference model for each user is represented by the utility function :

Here x (i) = (X1 (i), ..., xn (i)) is an n-dimensional bool eigenvector, which represents item I. XJ (i) = 1 (XJ (i) = 0) indicates that the item has a (none) J feature. For example:

X (5) = (1,0,1,0) means that item 5 contains the 1th and 3rd features, with no 2nd and 4th features.

W represents user preferences, weights. WJ (U) indicates the importance of user u to the characteristics of the J item. All weights are added equal to 1, and greater than 0, the larger the more attention.

1, the content-based method to generate the user's utility vectors for long-term preference：

is a collection of items scored by user U, and K is the normalization factor. For example:

2. Conversational Preferences

When group decisions are made, it is assumed that the user presents a project for the group discussion and evaluates the items presented by other group members.

In the panel discussion, all the projects were divided into four groups: BS (U) (best project), LS (U) (like project), NS (U) (Neutral project), DS (U) (not like project).

Assuming that the user prefers items with greater utility, the following constraints are met:

: The project utility set of user U in Group G.

Like what:

3. Utility functions for users and groups:

User U's group session preferences:

W (g) Indicates the average session preference of all members of the group. The WG (U) needs to satisfy the constraints, while the cosine of W (g) is the most similar.

Linear combination of the original user's long-term preferences and conversational preferences:

The following indicates that the user's long-term preference is influenced by group-induced preference, resulting in a real utility vector. Gamma is the parameter that controls the three variants of the preference combination.

Overall algorithm:

Situation-dependent combination of long-term and session-based Preferences in Group recommendations:an experimental Analy SIS----Group recommendations based on long-term and conversational preferences for contextual dependency combinations