Inspiration Seeds in Adaptive Ideation

• Inspiration Seeds are adaptive strategies for ideation systems that present curated example ideas during brainstorming, distinguishing between seekers and avoiders.
• The paper proposes a simple behavioral classifier using inspiration request frequencies and random forest models to identify user types early in the session.
• Empirical findings show that on-demand inspirations boost novelty for seekers but reduce it for avoiders, highlighting the benefit of personalized deployment.

“Inspiration deployment” in large-scale ideation refers to the design of systems that show example ideas, inspirations, or prompts during brainstorming. “Seeker or Avoider? User Modeling for Inspiration Deployment in Large-Scale Ideation” argues that a central assumption in much prior work is incomplete: inspirations are not uniformly helpful. Instead, ideators can be distinguished by their orientation toward inspiration itself. The paper defines two types—Inspiration seekers and Inspiration avoiders—and shows that the availability of inspiration improves novelty for seekers while reducing novelty for avoiders. It further proposes an early behavioral classifier, based primarily on inspiration-request behavior, that can identify these types quickly enough to support adaptive inspiration deployment in online ideation systems (Mackeprang et al., 2020).

1. Conceptual framing and research problem

The paper addresses personalization of inspirations in brainstorming systems. Earlier work had already shown that inspiration effects depend on factors such as timing, semantic similarity, and attention to inspirations, but had not centered individual preference differences toward inspirations themselves. The paper’s claim is that large-scale ideation platforms often operationalize inspiration deployment as a single interface or recommendation policy—such as always showing examples, providing an “inspire me” button, or optimizing timing and similarity globally—even though users differ systematically in whether inspiration helps them (Mackeprang et al., 2020).

This distinction matters because a one-size-fits-all inspiration policy can improve performance for some users while reducing it for others. The paper therefore introduces a simple user model built around two ideator types. Inspiration seekers are ideators who actively look for inspirations during ideation and derive their strategies from them. Inspiration avoiders are ideators who feel distracted by inspirations and prefer to follow their own ideation strategies. The contribution is not merely typological. It is also operational: the paper argues that these types can be inferred early enough from behavior to support adaptive interface decisions.

A plausible implication is that inspiration deployment should be treated as a prior personalization problem rather than only a recommendation-ranking problem. The paper explicitly suggests a hierarchical strategy: first decide whether a user should receive inspiration access at all; only then optimize timing, similarity, or ranking.

2. Exploratory formation of ideator types

The ideator types were first developed in a co-located brainstorming study with 15 participants. This study consisted of three in-person “brainwriting pool” sessions with five people each. Participants brainstormed for 30 minutes on the “Smart Coating” challenge, wrote ideas on paper, placed them in a shared pool, and could pick up and read others’ ideas as inspiration. Afterward, they completed a seven-question questionnaire about their ideation experience, including when they used inspirations, how they used them, and whether they found them inspiring or distracting. Two authors also directly observed the sessions (Mackeprang et al., 2020).

Across the 15 participants, the study produced 225 ideas and 220 reads of others’ ideas. Thematic analysis of the questionnaires surfaced three themes: social pressure, inspiration integration, and attitude toward inspirations. The decisive theme was attitude toward inspirations. Some participants described others’ ideas as useful scaffolds that helped them notice missing areas, switch application domains, or think more diversely. Others described inspirations as distracting, especially when the other ideas were far outside their own thinking. One participant explicitly rejected using inspiration at all.

These observations yielded the initial seeker/avoider distinction. The distinction was qualitative, but it was behaviorally grounded. Seekers were observed as drawing on others’ ideas to reorient or extend their own thinking. Avoiders were observed as either ignoring available inspirations, using them minimally, or reporting that they were disruptive. The paper’s later quantitative analysis should therefore be read as an operationalization of a qualitatively grounded distinction rather than an arbitrary clustering exercise.

3. Behavioral user model and early classification

To translate this distinction into a model suitable for online systems, the paper analyzed logs from three existing web-based ideation datasets totaling 193 sessions; after excluding users with fewer than three idea submissions, 173 sessions remained. In those studies, participants worked on ideation challenges using a web interface with an inspiration-request button. The logs recorded session start and end, browser-tab focus, idea submissions, and inspiration requests. The key signal chosen for operationalization was simply the number of inspiration requests (Mackeprang et al., 2020).

Dataset	Sessions	Duration and challenge
DS1	89	15 minutes, Smart Coating
DS2	30	25 minutes, Smart Coating
DS3	74	20 minutes, Bionic Radar

After filtering, the distributions of inspiration requests appeared almost bimodal, with medians of 6 in DS1, 3 in DS2, and 2 in DS3. On that basis, the paper defined a rule-based model for completed sessions: participants requesting more than the median number of inspiration requests were assigned to Inspiration Seeker; participants requesting inspiration at most once were assigned to Inspiration Avoider. The avoider definition was explicitly expanded from zero requests to “at most once” because single-request sessions often contained one early click followed by abandonment, which the authors interpreted as feature testing rather than sustained inspiration use. Applying this rule to the historical data classified 80 users as seekers and 51 as avoiders; the remainder were labeled Undetermined.

The paper then examined how early classification could be made by segmenting sessions into one-minute “inspiration sequences.” The qualitative result was that seekers could often be classified faster than avoiders, and by around 10 minutes the rule-based partial classification was better than chance in all datasets. Early improvement was nonlinear, motivating a machine-learning model based on finer-grained temporal bins.

The final reported classifier was a random forest-based regressor built on 15-second request-count bins. For example, a 10-minute session yields 40 bins. The authors used decision tree regressors and random forest regressors rather than direct classifiers in order to model a continuous transition between user types and select thresholds from ROC analysis. The random forest contained 200 decision trees with no maximum depth limit, and classes were weighted by frequency because seekers heavily outnumbered avoiders in Session A of the main experiment. Evaluation used ROC curves, AUC, and thresholds chosen to maximize the ratio of true positives to false positives.

The explicit evaluation metrics were:

$$\begin{aligned} \text{Accuracy: }&\frac{TP+TN}{TP+FP+FN+TN}\ \text{Precision: }&\frac{TP}{TP+FP}\ \text{Recall: }&\frac{TP}{TP+FN} \end{aligned}$$

Here, $TP$ denotes users correctly classified as seekers, $TN$ users correctly classified as avoiders, $FP$ users predicted to be seekers who are actually avoiders, and $FN$ users predicted to be avoiders who are actually seekers. The headline result is that the random forest achieved 73% accuracy after three minutes of ideation relative to final-session seeker/avoider labels. The paper interprets this not as recovery of an absolute psychological type, but as early prediction of the user’s full-session behavioral label. Predictive power rose most strongly in the first 3–4 minutes and then plateaued.

4. Main online experiment and empirical findings

The main online experiment recruited 380 participants from Amazon Mechanical Turk and used a two-stage design. Session A served as a classification session; Session B tested the effect of inspiration availability conditional on ideator type. Session A lasted 10 minutes, used the “Bionic Radar” challenge, and gave every participant access to the inspiration button. Because the median number of inspiration requests was not known in advance during live assignment, the study used fixed thresholds: avoider = 0 or 1 inspiration requests; seeker = 4 or more inspiration requests; undetermined = 2 to 4 requests; unmotivated = 3 or fewer idea submissions (Mackeprang et al., 2020).

Session A yielded the following classifications: 201 seekers, 59 avoiders, 83 undetermined, and 37 unmotivated. Participants made 3,018 inspiration requests in total, with a median of 5. Only seekers and avoiders proceeded to Session B. Session B lasted 15 minutes, used a different challenge—“Fabric Display”—and implemented a $2\times2$ factorial design: ideator type (seeker vs. avoider) by condition (on-demand inspiration available vs. baseline with no inspiration button). Inspirations were drawn one at a time from pools of 200 prior ideas per challenge, manually rated by two authors for novelty and value and sorted by summed novelty + value, so that all participants in the on-demand condition received high-quality inspirations. Of the 134 participants who moved to Session B, 81 sessions were usable after filtering software errors and nonsense or misunderstood responses.

The paper measured fluency as the number of ideas generated and inspiration use as the number of button clicks. For Session B, each generated idea was rated by external crowdworkers on novelty and value using five-point Likert scales, with at least three ratings per idea. Ratings were normalized within worker batch. Rather than averaging all ideas, the analysis focused on each participant’s maximum novelty and maximum value—the rating of that participant’s most novel idea and most valuable idea in the session. The rationale was that brainstorming often aims to increase the probability of producing standout ideas.

The most important empirical result concerns novelty. The multiple linear regression predicting maximum novelty showed a significant interaction between ideator type and inspiration availability. The paper reports a significant interaction at $p = 0.043$. The reported coefficient table was:

Term	Estimate	$p$
Grand Mean	0.83	$< 0.001$
Seeker vs. Avoider	-0.11	$0.096$
on-demand vs. baseline	-0.02	$TP$0
Seeker vs. Avoider : on-demand vs. baseline	0.34	$TP$1

The model used 122 observations with $TP$2. The text also states a significant main effect for ideator types at $TP$3, although the table lists the Seeker vs. Avoider coefficient at $TP$4. The stable substantive conclusion is the interaction: avoiders increase their novelty ratings when moving from on-demand to baseline, while seekers increase their novelty when moving from baseline to on-demand.

By contrast, the paper found no significant main or interaction effects on fluency, and no significant main or interaction effects on maximum value. The differential effect of inspiration availability is therefore concentrated in novelty rather than idea count or value. This is a narrow but practically important result: inspiration policies should not be evaluated only by fluency, and their strongest effect may be on originality rather than usefulness.

The paper also tested cross-challenge stability. In Session B’s on-demand condition, users classified as seekers in Session A requested significantly more inspirations than avoiders did, with a linear model reporting $TP$5, $TP$6, $TP$7, and estimate $TP$8. All Session A seekers in Session B on-demand remained seekers by the rule-based classification. Avoiders were less stable: among 20 avoiders in Session B on-demand, 11 remained avoiders, 5 became undetermined, 3 became seekers, and 1 became unmotivated. This suggests that ideator type is user-linked but not perfectly fixed across challenges.

5. Design implications for large-scale ideation systems

The paper’s principal design conclusion is that large-scale ideation systems should not treat inspiration availability as universally beneficial. Because seekers benefit from on-demand inspiration while avoiders are negatively affected in novelty terms, personalization should begin with whether inspiration access should be provided at all (Mackeprang et al., 2020).

The most informative behavioral signal in the paper is inspiration-request frequency, especially in the first few minutes. Clear avoider signals include zero requests, one request very early followed by no more, and continued ideation without returning to the inspiration feature. Clear seeker signals include repeated inspiration requests and request frequency above a threshold such as 4+ in a 10-minute classification session or above the session median in a completed-session rule. The fact that the random forest reaches 73% accuracy after only three minutes means intervention is possible before most of the ideation session is over.

The practical intervention logic proposed by the paper is correspondingly simple. For likely seekers, systems should keep the inspiration button visible and accessible and may benefit from adaptive recommendation. For likely avoiders, systems should suppress, hide, fade out, or disable inspiration access rather than assuming that more examples are helpful. For undetermined users, the system should preserve a cautious default until more evidence accumulates. The paper explicitly mentions future adaptive UI interventions such as disabling or fading out the inspiration button for avoiders while preserving support for seekers.

A further implication concerns recommendation architecture. The paper’s results do not primarily concern semantic similarity or timing of delivered inspirations, but they recast those issues as secondary. The pre-study identified inspiration integration as a salient theme, with seekers using inspirations as scaffolds or guidelines to broaden the challenge. Avoiders, by contrast, reported distraction even when inspirations were high quality. Because Session B deliberately used inspirations sorted by novelty + value, the negative effect on avoiders cannot be dismissed as a low-quality recommendation problem. The implication is that even high-quality inspirations may be counterproductive for the wrong user.

6. Limitations, threats to validity, and future directions

The paper presents the seeker/avoider distinction as promising rather than definitive. First, ideator type is not fully stable across challenges. Seekers were stable in the observed data, but some avoiders shifted categories between Session A and Session B. This suggests that the model should be treated as probabilistic and context-sensitive rather than as a permanent user trait (Mackeprang et al., 2020).

Second, classification relied almost entirely on inspiration requests. Some apparent avoiders may have ignored the feature because they misunderstood it rather than because they disliked inspiration. The paper therefore recommends better tutorials and explicit comparison of behavioral classification with self-reported inspiration preference. Third, the online experiment suffered substantial attrition and filtering, and Session B had only 81 usable sessions. Fourth, the reported classifier uses only a single behavioral channel even though richer signals existed in earlier logs, such as browser-tab focus and idea-submission timing. The paper suggests that factors such as confidence, challenge comprehension, prior brainstorming experience, and motivation may improve classification and clarify exceptions. Fifth, cross-challenge generalization was weaker than within-session prediction, which limits claims of robust domain transfer.

The proposed future directions align directly with adaptive inspiration systems. The paper suggests testing ideator types across more diverse challenges, explicitly measuring confidence, challenge comprehension, and motivation, comparing behavioral labels with self-assessed preference for inspiration, moving from static conditions to real-time adaptation, and exploring sequence-sensitive models beyond random forests, including recurrent neural networks or Bayesian networks.

Taken together, the paper reframes inspirations in ideation systems as a personalization problem. Inspirations are not uniformly beneficial, and the earliest phases of interaction already contain enough behavioral information to support adaptive deployment. The resulting model is intentionally simple—frequency of inspiration requests, rule-based thresholds, and a random forest over 15-second request bins—but it is sufficient to show that “inspiration seeker” and “inspiration avoider” are operationally meaningful categories for the design of large-scale ideation platforms (Mackeprang et al., 2020).