Conference Agenda

In Association Football, attacking players have traditionally been ranked according to the number of goals they score, and / or the number of goals they create for others. The disadvantage of this ranking system is clear: it does not account for the strengths of an opposing team. Here, we address by establishing the concept of a ‘par’ for each opponent: the average number of attacking contributions against that team. In this way, we are able to standardise measures of each player’s attacking contribution.

We use these standardised data to construct models of player influence in the teams of the 2011/12 English Premier League season. This allows us to assess not just a player’s level of contribution, but a duo’s level of contribution to the team’s attacking output. The resulting scores represent the combined strength of each attacking duo, allowing coaching staff to identify the strongest and weakest attacker combinations. Interestingly, the mean of these ‘duo scores’ appears to be a good predictor of a team’s position in the final Premier League standings. In fact, just one club recorded a top-half mean duo score, but a bottom-half final standing and we offer an explanation for this.

Viewing players as nodes in a network, with weighted, directed, edges between them illustrating influence, proves instructive for the team as a whole. For instance, algorithms grounded in Max-Plus Algebra can facilitate the identification of the strongest attacking duos, triumvirates and larger groups, all of which correspond to circuits in the network. The simplicity of this work makes it ripe for wider applicability and nuanced model development. Moreover, team structures are crucial to organisational success and, since our tools to identify key employees and groups are easily transferable to non-sporting organisations, we expect this work to be beneficial to management on a wider scale.

Collaboration is of fundamental importance to modern scientific and technological development, and expertise diversity has emerged as an important factor in predicting the success of collaboration. While expertise diversity has typically been seen as the knowledge attribute of a group (e.g., between collaborators), we provide an additional theoretical and empirical conceptualization that considers whether collaborators’ network knowledge is similar to the knowledge domain of their research output. Specifically, experience diversity or similarity can be assessed in at least two ways: between collaborators (dyadically) and between the researchers themselves and the substance of the research they engage in. We refer to these concepts as expertise diversity and divergent ideation. Expertise diversity measures the extent to which collaborators have distinct professional backgrounds, while divergent ideation represents how much the new research product diverges from what collaborators have previously worked on. While extensive research has explored the relationship between expertise diversity and innovation, few studies have examined how expertise diversity and the nature of the project (divergent ideation) jointly influence the success of first-time collaborations. This study addresses this gap by examining how these two factors—expertise diversity and divergent ideation—jointly influence both the impact of collaborative output and the likelihood of sustained collaboration.

To answer our research question we use data from 158,012 first-time partnerships recorded in the US Patent Office between the years of 1976 and 2012. We use natural language processing to estimate areas of expertise and to develop measures of expertise diversity based on inventors’ collaboration networks. Specifically, we use a Doc2Vec model on granted US utility patents to produce a 300-dimension embedding that represents the content of each patent. First, we use the embeddings of an inventor’s inventions to estimate their area of expertise (i.e. inventor’s topical focus). Topical focus is operationalized as the centroid of inventors’ prior work and is computed as the mean vector representation across all their outputs. Next, we operationalize expertise diversity as the cosine distance between a pair of inventors’ topical foci and divergent ideation as the average cosine distance between a pair of inventors’ prior expertise and the embedding location of their invention. Finally, we examine two distinct measures of partnership success—success in generating a high impact research output and success in establishing a lasting collaborative relationship.

Results show that collaborations exhibiting a high degree of expertise diversity produce more impactful products, while collaborations exhibiting a low degree of expertise diversity are more likely to collaborate again. Further, collaborations exhibiting a high degree of expertise diversity and a low-to-moderate degree of divergent ideation are most likely to create the highest impact inventions, but they are less likely to sustain their collaboration. We conclude our study by outlining the implications of our findings to the literature on diversity and technological innovation.