AI Can Enhance Creativity in Social Networks

TL;DR

This study trained a model that predicts people's ideation performances using semantic and network-structural features in an online platform and built SocialMuse, which maximizes people's predicted performances to generate peer recommendations for them.

Abstract

Can peer recommendation engines elevate people's creative performances in self-organizing social networks? Answering this question requires resolving challenges in data collection (e.g., tracing inspiration links and psycho-social attributes of nodes) and intervention design (e.g., balancing idea stimulation and redundancy in evolving information environments). We trained a model that predicts people's ideation performances using semantic and network-structural features in an online platform. Using this model, we built SocialMuse, which maximizes people's predicted performances to generate peer recommendations for them. We found treatment networks leveraging SocialMuse outperforming AI-agnostic control networks in several creativity measures. The treatment networks were more decentralized than the control, as SocialMuse increasingly emphasized network-structural features at large network sizes. This decentralization spreads people's inspiration sources, helping inspired ideas stand out better. Our study provides actionable insights into building intelligent systems for elevating creativity.

Figures (7)

• Figure 1: Data collection and experimentation framework. (A) The initial (round 1) network structure with 6 alter nodes and 18 ego nodes. Each ego is connected to two alters out of the six. Pre-recorded ideas of the same alters are shown to the egos of different study conditions (each with 18 egos). The entire arrangement is replicated across multiple trials with independent participants. (B) The protocol for each of the five rounds of idea generation. During attempt 1, an ego (blue node) generates ideas independently. During attempt 2, they view the ideas of the two followee alters (red nodes) and submit possible inspired ideas. Then, in the rewiring stage, the egos rate the ideas of all six alters and update which two alters to follow in the next round.
• Figure 2: Architecture of SocialMuse. In this stylized example, Ego $E8$ has completed the idea-generation activities for round $t-1$. In the rewiring stage of round $t-1$, we seek to recommend ego $E8$ two alters to follow for inspiration in round $t$. Seven other egos completed round $t$ before ego $E8$. (A) We first sense existing semantic and network-structural context factors. (B) Using a brute-force approach, we consider all possible alter-pairs that can be recommended to $E8$ and generate semantic and network features for each alternative. The alter-pair of $\{A1, A6\}$ is being considered in the illustration, as shown with dashed lines. (C) We use a trained model to separately predict ego $E8$'s marginal distinct idea counts in round $t$ corresponding to the features from each alter-pair alternative. (D) We recommend the highest-scoring alter-pair to the ego.
• Figure 3: Manipulating the availability of peer recommendations across conditions. In the rewiring stage of each round, the treatment egos receive AI-generated recommendations to follow two alters out of $6$ for inspiration in the following round. The control egos do not receive any recommendations. Importantly, the egos retain the agency to pay heed to the recommendation or not. In this illustration, solid lines show the two alter connections an ego chooses to make.
• Figure 4: Performance comparison between treatment and control conditions. The egos in the treatment condition significantly outperform the control egos in (A) marginal distinct idea counts per round, (B) non-redundant idea counts per round, (C) Creativity Quotient per round, and (D) Maximum SemDis scores (i.e., scores of best ideas) in a given round. Whiskers denote $95\%$ C.I.
• Figure 5: Gini Coefficient analysis. (A) Gini Coefficients are significantly lower in the treatment condition than in the control. (B) Gini Coefficients versus network size. The Gini Coefficients in the treatment condition drop significantly from the control condition at large network sizes. Whiskers and shaded regions denote $95\%$ C.I.