This paper introduces an ongoing project for mining potential business opportunities from the existing business innovation cases. Crevate (a consulting company) has collected thousands of cases from news articles and columns that are in textual narratives. We aim to transform the cases into knowledge graphs that cover what kinds of ideas (e.g., untact) are applied to which industrial areas (e.g., cafe). By using link prediction methods, we will be able to evaluate the prominence of combinations (e.g., untact cafe) between ideas and domains. This study focuses on explaining the collected cases and knowledge graphs with running examples.
In the business intelligence area, data sources are frequently textual narratives. Crevate, a business consulting company, has collected thousands of business innovation cases in short textual narratives. Also, a few tools for supporting the potential idea discovery follows fixed scenarios, and its results are in narratives [Wang and Ohsawa, 2011]. Furthermore, a few studies [Segura et al., 2018; Marjanovic, 2016] attempted to employ storytelling for helping business users explore visual analytics results. However, most of the existing studies focused on discovering bursty topics in social media texts [Yan et al., 2015; Lee et al., 2018; Lee et al., 2017]. There have not been studies that consider narrative characteristics consistently from collecting existing business innovation cases, via discovering potential business opportunities, to providing the opportunities in explainable ways.
In this study, we introduce an ongoing project that aims to propose a representation model for the business innovation cases in the narrative form. Based on the representation, we focus on (i) analyzing narratives for business innovation ⇤ Corresponding Author.
Copyright © 2020 by the paper’s authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). cases, (ii) discovering the potential business opportunities, and (iii) representing the opportunities as narratives again. 2
Crevate1 has collected 3000 business innovation cases to inspire its staff and clients. The staff of Crevate have investigated the cases introduced by news articles. For example, ‘a company applied bending machines on cafes.’ They have made descriptions for each case that consist of 5 to 6 sentences by summarizing the news articles. Also, they have titled the cases (e.g., automated cafe), classified the cases according to merchandise types (e.g., physical products, services, etc.), and tagged industrial areas and keywords. This dataset includes short narratives with titles, summaries, categories, and tags. The staff have also composed a three-layered taxonomy of the innovation cases. The lowest layer includes categories, such as ‘feeling tastes of the world,’ ‘feeling local tastes,’ ‘including numerous options,’ and so on. The middle layer consists of ‘experiencing the world,’ ‘experiencing the locality,’ ‘providing options to customers,’ and so on. The highest layer contains ‘extending experiences,’ ‘maximizing options,’ and so on. Therefore, we can use the dataset for evaluating various tasks in the narrative analysis (e.g., title generation, summarization, tagging, classification, labeling, etc.). Each case has been annotated with 15 attributes, as presented in Table 1.
The cases consist of three parts: (i) what (approaches), (ii) where (domains or industrial areas), and (iii) how (specific methods). In the example of Table 1, ‘Universal Yums2’ applied the subscription and delivery services on snacks by providing diverse local tastes. As a similar example, ‘Mouth3’ applied the subscription and delivery on indie foods by providing local foods from all over America on a website.
As shown in the summary and description in Table 1, these cases are simple and formalized narratives. Also, the title, industry, and tags provide us guidelines for analyzing the narratives. With Part-Of-Speech (POS) tagging, we can find the three parts from sentences in the summary and description. Objects of sentences will be related to domains of the cases (e.g., a box of snacks). From verbs, we can find the
Year Nation Company Image Video Link Tags
The name of services/products Abstract of descriptions for services/products Content and characteristics of services/products
seven categories (e.g., products, services, etc.) Industrial areas of services/products The year of release Service regions The name of company Sample images Promotion videos Promotion websites Keywords related to the innovation cases Categories on the lowest layer in the taxonomy Categories on the middle layer Categories on the highest layer
Snack boxes with tastes of the world This service monthly delivers a box of snacks which are representative of a country.
Universal Yums sends its staff to on-spot-surveys for finding local snacks. The staff tastes and scores the local snacks to compose monthly snack boxes. Scoring criteria are distinct local characteristics, diversity of tastes, and so on.
Service
2014 United States of America Universal Yums – – https://www.universalyums.com/ Subscription; Delivery; Snacks
approaches employed in the cases (e.g., delivers). Lastly, adverbs will reflect the specific methods for applying the approaches (e.g., monthly). WordNet can also help us find the three components from the descriptions. For example, since ‘snack’ is a hyponym of ‘Food and beverage,’ we can assume that it will indicate the domain of the case. In further research, we will also consider an automated extension of the dataset by crawling new articles. 3
By extracting approaches (what), domains (where), and methods (how) from the business innovation cases, we can model relationships between the three components. As shown in the two examples, lots of cases among the dataset share the approaches and domains, even sometimes a part of the methods. If we represent relationships between components in a case, we will be able to search the cases more semantically than before. Also, representing the relationships between cases will enable us to examine correlations between the approaches, domains, and methods (e.g., synergies and trade-offs). These relationships can be modeled as a knowledge graph. Fig. 1 presents the knowledge graph-based representations for ‘Universal Yum’ and ‘Mouth.’
We define the representation model that has 6 kinds of nodes and 7 sorts of relations. The types of nodes are as follows.
• Case: This node indicates the business innovation case itself. The other types of nodes are to describe the properties of this node. • Domain: This node presents industrial areas of the case. Domains can be described in multiple layers (e.g., Snacks 2 Food & Beverage). Also, a case can have multiple domains (e.g., ‘functional foods’ and ‘cosmetics’). • Approach: We also annotate which approaches are used in the cases. As with the above, one case can apply multiple approaches (e.g., ‘delivery’ and ‘subscription’). • Method: Methods means specific strategies that are used in the cases. ‘Universal Yum’ has conducted ‘on-thespot surveys’ for finding snacks that have ‘local characters’ to satisfy ‘diversity of tastes.’ • Category: This node presents the three-layered taxonomy of the innovation cases. • Metadata: Excluding the semantic information, the cases include additional information, such as company names, release years, service regions, and so on.
The types of relations are as follows.
• ‘deals with’ connects cases and domains of the cases. • ‘applies’ links cases and their approaches. • ‘is based on’ annotates methods of the cases. • ‘is included in’ links subclasses and superclasses. • ‘is run by,’ ‘is running business in,’ and ‘is released at’ commonly indicate additional information of the cases. By representing the cases in a single knowledge graph, we can examine which approaches and domains frequently appear together, as displayed in Fig. 2. On the relations, we can assign weights based on frequency of the cases. Chen et al. [2019a] also used a network model for idea mining. However, their model only combines two concepts (e.g., ‘leaf’ and (a) Snack boxes of ‘Universal Yum’ (b) Indie food market of ‘Mouth’ ‘spoon’) to discover novelty (e.g., ‘leaf-shaped spoon’). On the other hand, we depict components of innovation cases as detail as possible. 4
Crevate has merely provided a search engine for the existing innovation cases by using the approaches and domains. This retrieval has been conducted by a three-step procedure. Users insert business domains that they want. The search engine shows a list of candidate approaches to the users. Finally, after the users choose one of the approaches, the search engine provides innovations cases that have conducted on the selected domain with the chosen approach. This service can provide only existing cases.
Nevertheless, by analyzing the knowledge graph, we can
examine the potentials of the cases, whether they have been
tried or remained as terra incognita. There have been
various studies [Rossi et al., 2020] for predicting links in
knowledge graphs based on: tensor factorization
[Trouillon et al., 2017], network embedding [Chen et al., 2019b;
Kazemi and Poole, 2018], and affinity propagation [Wang et
al., 2018]. To discover the potential opportunities, we have
to consider both affinity between nodes and structures of the
knowledge graph. Let suppose that ‘cafe’ and ‘bookstore’ are
commonly connected to ‘providing cozy spaces’ with high
affinity. If ‘cafe’ and ‘untact services’ also have high affinity,
we can expect that ‘untact services’ is applicable to
‘bookstore,’ as well. Therefore, we will apply metapath-based
network embedding methods
This paper presents backgrounds and approaches of the ongoing project. In further research, we will focus on clarifying our methodologies considering characteristics of the dataset.
This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ICT Consilience Creative program (IITP-2019-2011-1-00783) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation).