Green knowledge management aims to integrate green or environmental aspects into all dimensions of knowledge management (KM). The need for this has increased greatly because of growing global environmental challenges. As such, green knowledge management can potentially help KM better support the UN Sustainable Development Goals (SDGs). But despite its potential, the research literature addressing green knowledge management is currently very limited, and not well developed.

A recent paper¹ attempts to add some rigor to explorations of green knowledge management by developing and validating a proposed scale for its measurement in organizations. The scale development steps followed established guidelines². The first step involved a literature review and interviews with managers. The collected information was then used to draft a scale which was proofread and refined by experts from industry and academia. After pilot testing, the scale was finalized, a comprehensive survey was initiated, and the collected data were subjected to validation through different statistical tests.

The finalized scale lists a range of factors against five dimensions, as shown below. The paper authors advise that organizations can use it as a checklist to ensure nothing is overlooked when creating their green measurement models.

Shortcomings of the scale

I recommend that the scale is used only to initiate the exploration of green knowledge management in organizations because some shortcomings mean it should not be used as-is.

One of these shortcomings relates to a bias towards explicit knowledge systems, as opposed to tacit knowledge processes. For example, the ‘knowledge storage’ dimension of the scale would be best changed to ‘knowledge retention’ and expanded to include tacit knowledge loss prevention measures such as mentoring.

Another shortcoming relates to the apparent lack of specialist environmental input into the scale. Although the paper extensively references environmental management and sustainability literature, all of the authors come from university departments related to business, economics, and organizational management, rather than environmental science and management departments. Sadly, despite advocating for the effective engagement of the best available knowledge in organizational decision-making, the KM field can itself be knowledge-siloed. As I’ve previously discussed in RealKM Magazine, there’s much that the KM field could learn from the knowledge of environmental scientists and managers. For example, specialist environmental knowledge in regard to the measurement of outcomes could assist the further development of the ‘knowledge application’ dimension of the scale, and in regard to stakeholder knowledge engagement could assist the further development of the ‘knowledge acquisition’ dimension of the scale.

Header image source: Created by Bruce Boyes with Perchance AI Photo Generator.

References:

Mariano’s five clusters of knowledge loss and five related interventions are summarized in the following sections and illustrated in Figure 1 below.

Five clusters of knowledge loss

1. Hanging – When organizational members become repositories of key organizational knowledge, the likelihood of disruptions to this knowledge, or ‘hanging’, may increase if proper retention mechanisms are not in place. If passage of time, infrequent use, or low perceived value influence the extent to which knowledge is used, hanging may occur. Hanging may be particularly detrimental when multiple individuals contribute to organizational processes and, therefore, the loss of one contribution may have a critical impact on overall performance.

2. Fading – ‘Fading’ relates to knowledge retained in storage facilities such as records, archives, collective electronic infrastructure, and databases. The disbandment of storage facilities such as when they break-up or cease to function; or their deterioration―a symptom of reduced quality, strength, passage of time or fall in disuse―are likely to determine the loss of crucial organizational knowledge. Since organizational knowledge may dissipate in the long term―in terms of content as well as the rationale behind it―deliberate and properly planned maintenance strategies are crucial.

3. Disengaging – When organizational members move to other roles, departments, subsidiaries, or geographical locations, this ‘disengaging’ may have potential detrimental consequences to organizational knowledge. Such detrimental consequences may include poor data handover; reduced knowledge accessibility and coordination; disruptions to knowledge flows; disappearances of important contacts; and misplaced, lost nodes, or broken links in networks of relationships.

4. Dissolving – In ‘dissolving’, knowledge is permanently lost. It may be the case of departing organizational members, including internal replacement, quitting, or retirement of employees in an aging workforce, or replaced management teams. Knowledge loss due to turnover includes the parting of subject matter expertise and governance knowledge; loss of knowledge about business relationships and social networks; loss of knowledge of business systems, processes or value chains; and loss of institutional memory.

5. Vanishing – ‘Vanishing’ relates to the complex combinations of collective and physical spaces where organizational activities take place. It may be the case when mergers, acquisitions, of restructuring impose radical changes that have a disruptive influence on the amount and quality of knowledge possessed at the organizational level, creating knowledge asymmetries, and loss of know-how.

Five potential interventions to help mitigate knowledge loss

1. Reminding – ‘Reminding’ addresses with the first cluster of knowledge loss i.e., hanging, It relates to individual knowledge. Reminding is proposed to be performed before (ex-ante) the accidental loss of knowledge may create disruptions at a more collective level. Strategies include prompt mechanisms to avoid knowledge loss due to passage of time or infrequent use of knowledge such as the use of how-to lists to accomplish a routine task; self-training mechanisms to re-acquire knowledge loss due to infrequent use; or the introduction of automated or less automated reminders to be sent periodically to provide the necessary information to accomplish certain tasks.

2. Refreshing – ‘Refreshing’ addresses the second cluster of knowledge loss i.e., fading. It relates to organizational knowledge embedded in storage facilities. Refreshing is proposed to be performed concurrently to keep the storage facilities fully functional. Strategies include the facilitation of both access and maintenance of the storage facilities.

3. Re-acquiring – ‘Re-acquiring’ addresses the third cluster of knowledge loss i.e., disengaging. It relates to knowledge found in the individuals’ networks of relationships. Re-acquiring is proposed to be performed concurrently to keep or strengthen the networks of relationships to find valuable knowledge that otherwise would be lost. Thus, the major aim of re-acquiring would be to facilitate the access to key nodes and links in available networks of relationships, including gatekeepers, brokers, central nodes and, at times, peripheral nodes. Strategies include the facilitation of virtual or less virtual networking sites; datasets of expertise or demographic inventories; or the development of relationships that may help increase knowledge exchange, especially at an intergenerational level.

4. Re-building – ‘Re-building’ addresses the fourth cluster of knowledge loss i.e., dissolving. Re-building is proposed to be performed ex-post, through reflection attempts and performance of actions to restore the lost knowledge, with the overall aim to strengthen the capacity to re-build the knowledge that had dissipated. When the organization realizes that knowledge may have been dissolved, actions to rebuild this knowledge, and to prevent similar disruptions need to be taken. These actions include the re-building of tacit, explicit, or relational knowledge, depending on the knowledge type that has been dissolved, including the overlapping of knowledge to have it always available at multiple locations, for instance implementing a master pool strategy to ensure that multiple individuals possess the needed knowledge.

5. Re-inventing – Finally, ‘re-inventing’ addresses the last cluster of knowledge loss i.e., vanishing. Re-inventing is proposed to be performed when knowledge has been permanently lost, often unknowingly. The major aim is to facilitate the recreation or optimization of knowledge and related processes to favor knowledge reinstatement.

Header image source: Gerd Altmann on Pixabay.

Reference:

The findings of reviews like these are very important for KM practice. For example, some of the eight emerging innovative concepts in KM identified in Schenk’s review are receiving much less attention from KM practitioners than others, which is a concern, and the significant implications flowing from Serenko’s review include the need to consider that KM may exist as a cluster of divergent schools of thought rather than a single discipline.

Now, coinciding with its 20^th anniversary, the journal Knowledge Management Research & Practice (KMRP) has published a new narrative review³ titled “The future of knowledge management: an agenda for research and practice.” KMRP is focused on the interplay between research and practice, so the findings of the review are of particular relevance to evidence-based practice in KM. The first issue of KMRP was published in May 2003, and the twentieth volume at the end of 2022. The review is authored by John Edwards and Antti Lönnqvist, who are both members of the KMRP Editorial Board.

In this RealKM Magazine article, I summarise Edwards and Lönnqvist’s review and raise further important perspectives in additional commentary, including alerting to significant issues that I consider are missing from the review. The structure of sections 1 to 5 below is as follows:

• Sections 1 and 2: Edwards and Lönnqvist first look at selected past KM activity, including descriptions, predictions, initiatives and other research agendas.
• Section 3: Edwards and Lönnqvist state that their review looks backwards only in order to look forward, so they then merge their retrospective into a consideration of the current states of KM literature and KM practice, raising some issues that they consider need to be taken into account for the future.
• Section 4: Finally, Edwards and Lönnqvist synthesize an agenda that they contend is focused on research on practice, not just research and practice.
• Section 5: I alert to significant issues that I consider are missing from Edwards and Lönnqvist’s review.

1. What is KM?

2. Looking backwards to look forward

2.1. Past descriptions of KM

2.2. “Pre-KMRP” predictions about the future of KM

2.3. Past research agendas

2.4. Past initiatives

3. Aspects and issues

3.1. Is KM dying?

3.2. Repeating the same mistakes

3.3. KM and other new(ish) technologies

3.4. Problem-driven research

3.5. Scope

Edwards and Lönnqvist’s review:

• KM can operate on several levels, including at least those shown in Figure 1.
• The placing of the levels is indicative, and is not clear-cut, and linkages have been omitted deliberately to avoid blinkered thinking.
• It could be argued that the focus of KM research on the organization has been taken too far, with most published KM research having this focus.
• Within organizations, trying to do KM at the group level can be more of a hindrance than a help – the classic criticism relating to organizational silos.
• There are wider societal implications of what an organization does.
• All KM needs to be grounded in personal KM, as all aspects of KM, including organizational KM, can be seen as stemming from the personal relevance of KM.
• The gap between academics and practitioners appears to have widened since 2013, with Serenko’s landmark 2021 review²¹ identifying that there is a need for knowledge brokers that may deliver the KM academic body of knowledge to practitioners.

3.6. Fake knowledge

4. Agenda

4.1. For those doing research

4.2. For journal editors and reviewers

4.3. For those in and around KM

5. What’s missing?

Header image source: Javier Allegue Barros on Unsplash.

References and notes:

The valuable role of arts and culture in knowledge management (KM) is increasingly being recognized. For example,

• Dr Arthur Shelley (a RealKM Director) has developed The Organizational Zoo as a creative metaphor¹ for human behaviors in organizations.
• Stephanie Barnes highlights the benefits of creativity and arts-based interventions (ABIs) in her radical KM approach².
• Goal 14 of the Agenda Knowledge for Development³ is “The arts and culture are central to knowledge societies. Literature, the performing arts and the visual arts are key elements of a knowledge society, as are religion and spirituality.”
• Laura Rademaker, Joakim Goldhahn, Mr Gabriel Maralngurra, Mr Kenneth Mangiru, Paul S.C.Taçon, and Sally K. May describe how galleries of First Nations rock art are vast repositories of knowledge⁴ from which we can all learn.
• Dr Ann Russell uses metaphor to explore the the hidden curriculum in art education, where knowledge, skills, and values are implicitly imparted to students⁵.

In response, RealKM Magazine has commenced a new series of articles on arts and culture in KM. Articles in the series will initially be published weekly, and contributions are invited – see below.

The series will explore relationships and interactions between all aspects of KM and arts and societal culture in their broadest sense, with “arts” and “culture” defined⁶ as follows:

Contributions are invited

Submissions to the RealKM Magazine arts and culture in knowledge management series are warmly invited. Potential contributions can include:

1. Direct contributions of arts or cultural materials that work with knowledge in some way

For example, any of the following:

• poetry or story texts
• video or audio recordings of dance, music, theatre performances, or intangible cultural heritage⁷
• images of paintings, sculptures, architecture, or cultural heritage
• short or long films

that play a role in knowledge creation, storage, use, or sharing, together with an accompanying text explanation of how this happens or is achieved.

If you’re interested in potentially contributing in this way, please feel free to make contact to discuss your ideas.

2. Informative or educational articles that explore aspects of arts and culture in KM

This type of article submission should meet the requirements of the RealKM Magazine editorial guidelines. In particular, these articles should be backed by sound research on accepted or emerging KM practice, or describe a real and specific case scenario. All research references must also be linked open access articles or publications.

Some exceptions to the requirement for articles to be backed by sound research can however be made, for example to deliver epistemic justice⁸ in regard to indigenous or other forms of knowledge that are currently underrepresented in the research literature. Other articles in this series and their references can also be potentially be used as references for your article.

Please note that the requirement for articles to be backed by sound research is not to say that professional expertise has no value in comparison to research literature. Professional expertise certainly has value, and indeed, is also one of the four sources of evidence in evidence-based KM. Rather, this requirement is to emphasize that RealKM Magazine‘s purpose in the wider KM landscape is research communication, and there are numerous other forums for expressions of professional expertise.

If you’re interested in potentially contributing in this way, please feel free to make contact to discuss your ideas.

Header image: Colorful pathways of the white matter by Francois Rheault, Sherbrooke University, in The Neuro Bureau Brain Art Competition 2018, CC BY-SA 4.0.

References:

Recent research¹ has identified the development of knowledge graphs as an important aspect of artificial intelligence (AI) in knowledge management (KM). The same research also recommended the training of knowledge scientists who can build knowledge graphs that represent background knowledge and that complement training data.

To assist in advancing AI in KM, this 6-part series of articles has provided an introduction to knowledge graphs.

Part 1 defines knowledge graphs and summarizes their applications. Two definitions are put forward – one general, and the other technical. The “knowledge” in the term “knowledge graphs” refers to what Nonaka and Takeuchi call “explicit knowledge,” that is, something that is known and can be written down. The recent applications of knowledge graphs include organizing knowledge over the internet, data integration in enterprises, and artificial intelligence.

Part 2 charts the history of knowledge graphs, an awareness of which is considered very important.

Parts 3, 4, and 5 then draw on Hogan and colleagues’ comprehensive tutorial article² and other research to provide an introduction to the technical aspects of knowledge graphs. To keep the discussion accessible, Hogan and colleagues’ present concrete examples for a hypothetical knowledge graph, which are reproduced in parts 3, 4, and 5. This hypothetical knowledge graph relates to tourism in Chile.

Part 3 outlines graph data models and the languages used to query and validate them.

Part 4 present deductive formalisms by which knowledge can be represented and entailed.

Part 5 describes inductive techniques by which additional knowledge can be extracted.

Knowledge graphs serve as a common substrate of knowledge within an organisation or community, enabling the representation, accumulation, curation, and dissemination of knowledge over time. In this role, knowledge graphs have been applied for diverse use-cases, ranging from commercial applications – involving semantic search, user recommendations, conversational agents, targeted advertising, transport automation, and so on – to open knowledge graphs made available for the public good.

General trends include the use of knowledge graphs to integrate and leverage data from diverse sources at large scale, and the combination of deductive (rules, ontologies, etc.) and inductive techniques (machine learning, analytics, etc.) to represent and accumulate knowledge.

Header image source: Crow Intelligence, CC BY-NC-SA 4.0.

References:

In this section of their comprehensive tutorial article¹, Hogan and colleagues discuss two forms of symbolic learning for knowledge graphs: rule mining for learning rules and axiom mining for learning other forms of logical axioms.

The supervised techniques discussed so far learn numerical models that are hard to interpret; for example, taking the graph of Figure 19, knowledge graph embeddings might predict the edge as being highly plausible, but the reason lies implicit in a complex matrix of learned parameters. Embeddings further suffer from the out-of-vocabulary problem, where they are often unable to provide results for inputs involving previously unseen nodes or edge-labels. An alternative is to use symbolic learning to learn hypotheses in a logical (symbolic) language that “explain” sets of positive and negative edges. Such hypotheses are interpretable; furthermore, they are quantified (e.g., “all airports are domestic or international”), partially addressing the out-of-vocabulary issue.

Rule Mining

Rule mining, in the general sense, refers to discovering meaningful patterns in the form of rules from large collections of background knowledge. In the context of knowledge graphs, we assume a set of positive and negative edges as given. The goal of rule mining is to identify new rules that entail a high ratio of positive edges from other positive edges, but entail a low ratio of negative edges from positive edges. The types of rules considered may vary from more simple cases, such as , to more complex rules, such as , indicating that airports near capitals tend to be international airports; or , indicating that flights within the same country denote domestic flights.

Per the international airport example, rules are not assumed to hold in all cases, but rather are associated with measures of how well they conform to the positive and negative edges. In more detail, we call the edges entailed by a rule and the set of positive edges (not including the entailed edge itself) the positive entailments of that rule. The number of entailments that are positive is called the support for the rule, while the ratio of a rule’s entailments that are positive is called the confidence for the rule. The goal is to find rules with both high support and confidence.

While similar tasks have been explored for relational settings with Inductive Logic Programming (ILP)², when dealing with an incomplete knowledge graph, it is not immediately clear how to define negative edges. A common heuristic is to adopt a Partial Completeness Assumption (PCA)³, which considers the set of positive edges to be those contained in the data graph.

An influential rule-mining system for graphs is AMIE⁴, which adopts the PCA measure of confidence and builds rules in a top-down fashion. For each such rule head, three types of refinements are considered, which add an edge with: (1) one existing variable and one fresh variable; (2) an existing variable and a node from the graph; (3) two existing variables. Combining refinements gives rise to an exponential search space that can be pruned. First, if a rule does not meet the support threshold, then its refinements need not be explored as refinements (1–3) reduce support. Second, only rules up to fixed size are considered. Third, refinement (3) is applied until a rule is closed, meaning that each variable appears in at least two edges of the rule (including the head); the previous rules produced by refinements (1) and (2) are not closed.

Later works have built on these techniques for mining rules from knowledge graphs. Gad-Elrab et al.⁵ propose a method to learn non-monotonic rules – rules with negated edges in the body – to capture exceptions to base rules. The RuLES system⁶ also learns non-monotonic rules and extends the confidence measure to consider the plausibility scores of knowledge graph embeddings for entailed edges not appearing in the graph. In lieu of PCA, the CARL system⁷ uses knowledge of the cardinalities of relations to find negative edges, while d’Amato et al.⁸ use ontologically entailed negative edges for measuring the confidence of rules generated by an evolutionary algorithm.

Another line of research is on differentiable rule mining, which enables end-to-end learning of rules by using matrix multiplication to encode joins in rule bodies. Along these lines, NeuralLP⁹ uses an attention mechanism to find variable-length sequences of edge labels for path-like rules. DRUM¹⁰ also learns path-like rules, where, observing that some edge labels are more/less likely to follow others, the system uses bidirectional recurrent neural networks (a technique for learning over sequential data) to learn sequences of relations for rules. These differentiable rule mining techniques are, however, currently limited to learning path-like rules.

Axiom Mining

Aside from rules, more general forms of axioms – expressed in logical languages such as Description Logics (DLs) – can be mined from a knowledge graph. We can divide these approaches into two categories: those mining specific axioms and more general axioms.

Among works mining specific types of axioms, disjointness axioms are a popular target; for example, the disjointness axiom DomesticAirport ⊓ InternationalAirport ≡ ⊥ states that the intersection of the two classes is equivalent to the empty class, i.e., no individual can be instances of both classes. Völker et al.¹¹ extract disjointness axioms based on (negative) association rule mining, which finds pairs of classes where each has many instances in the knowledge graph but there are relatively few (or no) instances of both classes. Töpper et al.¹² rather extract disjointness for pairs of classes that have a cosine similarity – computed over the nodes and edge-labels associated with a given class – below a fixed threshold. Rizzo et al.¹³ propose an approach that can further capture disjointness constraints between class descriptions (e.g., city without an airport nearby is disjoint from city that is the capital of a country) using a terminological cluster tree that first extracts class descriptions from clusters of similar nodes, and then identifies disjoint pairs of class descriptions.

Other systems propose methods to learn more general axioms. A prominent such system is DL-Learner¹⁴, which is based on algorithms for class learning (a.k.a. concept learning), whereby given a set of positive nodes and negative nodes, the goal is to find a logical class description that divides the positive and negative sets. Like AMIE, such class descriptions are discovered using a refinement operator used to move from more general classes to more specific classes (and vice versa), a confidence scoring function, and a search strategy. The system further supports learning more general axioms through a scoring function that determines what ratio of edges that would be entailed were the axiom true are indeed found in the graph; for example, to score the axiom ∃flight-.DomesticAirport ⊑ InternationalAirport over Figure 19, we can use a graph query to count how many nodes have incoming flights from a domestic airport (there are three), and how many nodes have incoming flights from a domestic airport and are international airports (there is one), where the greater the difference between both counts, the weaker the evidence for the axiom.

Next part: (part 6): Summary and conclusion.

Header image source: Crow Intelligence, CC BY-NC-SA 4.0.

References:

In this section of their comprehensive tutorial article¹, Hogan and colleagues introduce two types of graph neural network (GNN): recursive and convolutional.

Rather than compute numerical representations for graphs, an alternative is to define custom machine learning architectures for graphs. Most such architectures are based on neural networks given that a neural network is already a directed weighted graph, where nodes serve as artificial neurons, and edges serve as weighted connections (axons). However, the topology of a traditional (fully connected feed-forward) neural network is quite homogeneous, having sequential layers of fully connected nodes. Conversely, the topology of a data graph is typically more heterogeneous.

A graph neural network (GNN) is a neural network where nodes are connected to their neighbours in the data graph. Unlike embeddings, GNNs support end-to-end supervised learning for specific tasks: Given a set of labelled examples, GNNs can be used to classify elements of the graph or the graph itself. GNNs have been used to perform classification over graphs encoding compounds, objects in images, documents, and so on; as well as to predict traffic, build recommender systems, verify software, and so on. Given labelled examples, GNNs can even replace graph algorithms; for example, GNNs have been used to find central nodes in knowledge graphs in a supervised manner.

Recursive Graph Neural Networks

Recursive graph neural networks (RecGNNs) are the seminal approach to graph neural networks. The approach is conceptually similar to the abstraction illustrated in Figure 17, where messages are passed between neighbours towards recursively computing some result. However, rather than define the functions used to decide the messages to pass, we rather give labelled examples and let the framework learn the functions.

In a seminal paper², Scarselli and colleagues proposed what they generically call a graph neural network (GNN), which takes as input a directed graph where nodes and edges are associated with static feature vectors that can capture node and edge labels, weights, and so on. Each node in the graph also has a state vector, which is recursively updated based on information from the node’s neighbours – i.e., the feature and state vectors of the neighbouring nodes and edges – using a parametric transition function. A parametric output function then computes the final output for a node based on its own feature and state vector. These functions are applied recursively up to a fixpoint. Both parametric functions can be learned using neural networks given a partial set of labelled nodes in the graph. The result can thus be seen as a recursive (or even recurrent) neural network architecture. To ensure convergence up to a fixpoint, the functions must be contractors, meaning that upon each application, points in the numeric space are brought closer together.

To illustrate, assume that we wish to identify new locations needing tourist information offices. In Figure 18, we illustrate the GNN architecture proposed by Scarselli and colleagues for a sub-graph of Figure 15, where we highlight the neighbourhood of In this graph, nodes are annotated with feature vectors (n_x) and hidden states at step t (h_x^(t)), while edges are annotated with feature vectors (a_xy). Feature vectors for nodes may, for example, one-hot encode the type of node (City, Attraction, etc.), directly encode statistics such as the number of tourists visiting per year, and so on. Feature vectors for edges may, for example, one-hot encode the edge label (i.e., the type of transport), directly encode statistics such as the distance or number of tickets sold per year, and so on. Hidden states can be randomly initialised. The right-hand side of Figure 18 provides the GNN transition and output functions, where N(x) denotes the neighbouring nodes of x, f_w(·) denotes the transition function with parameters w, and g_w′(·) denotes the output function with parameters w′. An example is also provided for Punta Arenas (x = 1). These functions will be recursively applied until a fixpoint is reached. To train the network, we can label examples of places that already have tourist offices and places that do not have tourist offices. These labels may be taken from the

knowledge graph or may be added manually. The GNN can then learn parameters w and w′ that give the expected output for the labelled examples, which can subsequently applied to label other nodes.

Convolutional Graph Neural Networks

Convolutional neural networks (CNNs) have gained a lot of attention, in particular, for machine learning tasks involving images. The core idea in the image setting is to apply small kernels (a.k.a. filters) over localised regions of an image using a convolution operator to extract features from that local region. When applied to all local regions, the convolution outputs a feature map of the image. Multiple kernels are typically applied, forming multiple convolutional layers. These kernels can be learned, given sufficient labelled examples.

Both GNNs and CNNs work over local regions of the input data: GNNs operate over a node and its neighbours in the graph, while (in the case of images) CNNs operate over a pixel and its neighbours in the image. Following this intuition, a number of convolutional graph neural networks (ConvGNNs) – a.k.a. graph convolutional networks (GCNs) – have been proposed, where the transition function is implemented by means of convolutions. A benefit of CNNs is that the same kernel can be applied over all the regions of an image, but this creates a challenge for ConvGNNs, since – unlike in the case of images, where pixels have a predictable number of neighbours – the neighbourhoods of different nodes in a graph can be diverse. Approaches to address these challenges involve working with spectral or spatial representations of graphs that induce a more regular structure from the graph. An alternative is to use an attention mechanism to learn the nodes whose features are most important to the current node.

Aside from architectural considerations, there are two main differences between RecGNNs and ConvGNNs. First, RecGNNs aggregate information from neighbours recursively up to a fixpoint, whereas ConvGNNs typically apply a fixed number of convolutional layers. Second, RecGNNs typically use the same function/parameters in uniform steps, while different convolutional layers of a ConvGNN can apply different kernels/weights at each distinct step.

Next part: (section 5.4): Inductive knowledge – Symbolic learning.

Header image source: Crow Intelligence, CC BY-NC-SA 4.0.

References:

Machine learning can be used for directly refining a knowledge graph; or for downstream tasks using the knowledge graph, such as recommendation, information extraction, question answering, query relaxation, query approximation, and so on. However, machine learning techniques typically assume numeric representations (e.g., vectors), distinct from how graphs are usually expressed. In this section of their comprehensive tutorial article¹, Hogan and colleagues explore how graphs can be encoded numerically for machine learning.

A first attempt to represent a graph using vectors would be to use a one-hot encoding, generating a vector of length |L| · |V| for each node – with |V| the number of nodes in the input graph and |L| the number of edge labels – placing a one at the corresponding index to indicate the existence of the respective edge in the graph, or zero otherwise. Such a representation will, however, typically result in large and sparse vectors, which will be detrimental for most machine learning models.

The main goal of knowledge graph embedding techniques is to create a dense representation of the graph (i.e., embed the graph) in a continuous, low-dimensional vector space that can then be used for machine learning tasks. The dimensionality d of the embedding is fixed and typically low (often, e.g., 50 ≥ d ≥ 1000). Typically the graph embedding is composed of an entity embedding for each node: a vector with d dimensions that we denote by e; and a relation embedding for each edge label: (typically) a vector with O(d) dimensions that we denote by r. The overall goal of these vectors is to abstract and preserve latent structures in the graph. There are many ways in which this notion of an embedding can be instantiated. Most commonly, given an edge a specific embedding approach defines a scoring function that accepts e_s (the entity embedding of node ), r_p (the relation embedding of edge label p) and e_o (the entity embedding of node ) and computes the plausibility of the edge: how likely it is to be true. Given a data graph, the goal is then to compute the embeddings of dimension d that maximise the plausibility of positive edges (typically edges in the graph) and minimise the plausibility of negative examples (typically edges in the graph with a node or edge label changed such that they are no longer in the graph) according to the given scoring function. The resulting embeddings can then be seen as models learned through self-supervision that encode (latent) features of the graph, mapping input edges to output plausibility scores.

Embeddings can then be used for a number of low-level tasks. The plausibility scoring function can be used to assign confidence to edges (possibly extracted from an external source) or to complete edges with missing nodes/edge labels (a.k.a. link prediction). Additionally, embeddings will typically assign similar vectors to similar terms and can thus be used for similarity measures.

A wide range of knowledge graph embedding techniques have been proposed. The most prominent are:

• translational models where relations are seen as translating subject entities to object entities
• tensor decomposition models that extract latent factors approximating the graph’s structure
• neural models based on neural networks
• language models based on word embedding techniques.

Translational Models

Translational models interpret edge labels as transformations from subject nodes (a.k.a. the source or head) to object nodes (a.k.a. the target or tail); for example, in the edge the edge label bus is seen as transforming to and likewise for other bus edges. A seminal approach is TransE². Over all positive edges TransE learns vectors e_s, r_p, and e_o aiming to make e_s + r_p as close as possible to e_o. Conversely, if the edge is negative, then TransE attempts to learn a representation that keeps e_s + r_p away from e_o.

TransE can however be too simplistic, as it will try to give similar vectors to all target locations, which may not be feasible given other edges. To resolve such issues, many variants of TransE have been investigated, typically using a distinct hyperplane (e.g., TransH) or vector space (e.g., TransR, TransD) for each type of relation. Recently, RotatE³ proposes translational embeddings in complex space, which allows to capture more characteristics of relations, such as direction, symmetry, inversion, antisymmetry, and composition. Embeddings have also been proposed in non-Euclidean space; e.g., MuRP⁴ uses relation embeddings that transform entity embeddings in the hyperbolic space of the Poincaré ball mode, whose curvature provides more “space” to separate entities with respect to the dimensionality.

Tensor Decomposition Models

A second approach to derive graph embeddings is to apply methods based on tensor decomposition. A tensor is a multidimensional numeric field that generalises scalars (0-order tensors), vectors (1-order tensors), and matrices (2-order tensors) towards arbitrary dimension/order. Tensor decomposition involves decomposing a tensor into more “elemental” tensors (e.g., of lower order) from which the original tensor can be recomposed (or approximated) by a fixed sequence of basic operations. These elemental tensors can be seen as capturing latent factors in the original tensor. There are many approaches to tensor decomposition, including rank decompositions.

Leaving aside graphs, consider an (a × b)-matrix (i.e., a 2-order tensor) C, where each element (C)_ij denotes the average temperature of the ith city of Chile in the jth month of the year. Since Chile is a long, thin country – ranging from subpolar to desert climates – we may decompose C into two vectors representing latent factors – x (with a elements) giving lower values for cities with lower latitude, and y (with b elements), giving lower values for months with lower temperatures – such that computing the outer product6 of the two vectors approximates C reasonably well: x⊗y ≈ C. If there exist x and y such that x ⊗ y = C, then we call C a rank-1 matrix. Otherwise, the rank r of C is the minimum number of rank-1 matrices we need to sum to get precisely C, i.e., x₁ ⊗ y₁ + . . . x_r ⊗ y_r = C. In the temperature example, x₂ ⊗ y₂ might correspond to a correction for altitude, x₃ ⊗ y₃ for higher temperature variance further south, and so on. A (low) rank decomposition of a matrix then sets a limit d on the rank and computes the vectors (x₁, y₁, . . . , x_d, y_d) such that x₁ ⊗ y₁+· · ·+x_d ⊗ y_d gives the best d-rank approximation of C. Noting that to generate n-order tensors we need to compute the outer product of n vectors, we can generalise this idea towards low rank decomposition of tensors; this method is called canonical polyadic decomposition⁵.

DistMult⁶ is a seminal method for computing knowledge graph embeddings based on rank decompositions, but does not capture edge direction. Rather than use a vector as a relation embedding, RESCAL⁷ uses a matrix, which can capture edge direction. However, RESCAL incurs a higher cost in terms of space and time than DistMult. Recently, ComplEx and HolE both use vectors for relation and entity embeddings, but ComplEx uses complex vectors, while HolE uses a circular correlation operator (on reals) to capture edge-direction.

Neural Models

A number of approaches instead use neural networks to learn knowledge graph embeddings with non-linear scoring functions for plausibility.

An early neural model was semantic matching energy⁸, which learns parameters (a.k.a. weights: w, w′) for two functions – f_w(e_s, r_p) and g_w′(e_o, r_p) – such that the dot product of the result of both functions gives the plausibility score. Both linear and bilinear variants of f_w and g_w′ are proposed. Another early proposal was neural tensor networks⁹, which maintains a tensor W of weights and computes plausibility scores by combining the outer product e_s ⊗ W ⊗ e_o with r_p and a standard neural layer over e_s and e_o. The tensor W yields a high number of parameters, limiting scalability. Multi layer perceptron¹⁰ is a simpler model, where e_s, r_p, and e_o are concatenated and fed into a hidden layer to compute the plausibility score.

More recent models use convolutional kernels. ConvE¹¹ generates a matrix from e_s and r_p by “wrapping” each vector over several rows and concatenating both matrices, over which (2D) convolutional layers generate the embeddings. A disadvantage is that wrapping vectors imposes an arbitrary two-dimensional structure on the embeddings. HypER¹² also uses convolutions, but avoids such wrapping by applying a fully connected layer (called the “hypernetwork”) to r_p to generate relation-specific convolutional filters through which the embeddings are generated.

The presented approaches strike different balances in terms of expressivity and the number of parameters that need to be trained. While more expressive models, such as neural tensor networks, may better fit more complex plausibility functions over lower dimensional embeddings by using more hidden parameters, simpler models and convolutional networks that enable parameter sharing by applying the same (typically small) kernels over different regions of a matrix, require handling fewer parameters overall and are more scalable.

Language Models

Embedding techniques were first explored as a way to represent natural language within machine learning frameworks, with word2vec¹³ and GloVe¹⁴ being two seminal approaches. Both approaches compute embeddings for words based on large corpora of text such that words used in similar contexts (e.g., “frog,” “toad”) have similar vectors.

Approaches for language embeddings can be applied for graphs. However, while graphs consist of an unordered set of sequences of three terms (i.e., a set of edges), text in natural language consists of arbitrary-length sequences of terms (i.e., sentences of words). Along these lines, RDF2Vec¹⁵ performs biased random walks on the graph and records the paths traversed as “sentences,” which are then fed as input into the word2vec model.

RDF2Vec also proposes a second mode where sequences are generated for nodes from canonically-labelled sub-trees of which they are a root node. Conversely, KGloVe is based on the GloVe model. Much like how the original GloVe model considers words that co-occur frequently in windows of text to be more related, KGloVe uses personalised PageRank to determine the most related nodes to a given node, whose results are then fed into the GloVe model.

Entailment-aware Models

The embeddings thus far consider the data graph alone. But what if an ontology or set of rules is provided? One may first consider using constraint rules to refine the predictions made by embeddings. More recent approaches rather propose joint embeddings that consider both the data graph and rules. KALE¹⁶ computes entity and relation embeddings using a translational model (specifically TransE) that is adapted to further consider rules using t-norm fuzzy logics. But generating ground rules can be costly.

An alternative approach, adopted by FSL¹⁷, observes that in the case of a simple rule, such as the relation embedding bus should always return a lower plausibility than connects to. Thus, for all such rules, FSL proposes to train relation embeddings while avoiding violations of such inequalities. While relatively straightforward, FSL only supports simple rules, while KALE also supports more complex rules.

Next part: (section 5.3): Inductive knowledge – Graph neural networks.

Header image source: Crow Intelligence, CC BY-NC-SA 4.0.

References:

The second meeting of the KMGN Research Community earlier this year looked at knowledge hiding research. In regard to the terminology around knowledge hiding and hoarding, concerns were expressed by Stuart French and others about the emotive connotations of the term “knowledge hiding.” This is because such language suggests a deliberate intention to not share knowledge, when there could be organizational constraints that lead a person to choose to not share their knowledge. In response, I advised that the less emotive term “knowledge withholding” is also being used in a number of papers.

Related to this, a newly published paper¹ reports the findings of a systematic review that has sought to clarify the current research knowledge base around the concepts of knowledge withholding, hiding, and hoarding, and also the additional identified behaviors of knowledge-sharing hostility, knowledge contribution loafing, and knowledge disengagement. The concept of knowledge withholding is seen as an overall umbrella concept under which the other concepts sit. The study also provides recommendations to guide future research.

For their systematic review, researchers Gonçalves, Curado, and Oliveira used a protocol based on the PRISMA guidelines². The review covered articles published between 2000 and 2021. After a keyword search and the application of inclusion criteria, 84 research articles and six conceptual articles were identified for content analysis. The results show exponential growth in knowledge withholding-related research in recent years, with particular significance in the three years up to 2021.

Theoretical basis

The results of Gonçalves and colleagues study show that 41% of the reviewed articles address the social exchange theory as a theoretical background for knowledge withholding research, bridging knowledge hiding and knowledge hoarding. The social exchange theory addresses social behavior through economic principles of cost-benefits in social sharing, resulting in an analysis of risks and benefits. The theory, while more focused on the economic rather than psychological aspects of social exchange, is the most adopted perspective used in knowledge withholding research, given its discussion of emergent properties and the anticipation of benefit in the exchange. Such a perspective supports a view of knowledge as a source of power.

Similarly, 18% of the research papers use the conservation of resources theory for both knowledge hiding and knowledge hoarding. The conservation of resources theory explains the human motivation to engage in behaviors that drive the conservation or the pursuit of new resources when psychological stressors are at play. The approach of psychological ownership of knowledge was found in 11% of the papers, paving the way for research focused on defensive or territorial behaviors towards a sense of property protection.

Building on the psychosocial aspects of knowledge results also shows a focus on both self-determination theory (7%) and social learning theory (8%) driving knowledge withholding research. Given the psychological needs for competence, autonomy, and social relations as part of intrinsic motivations that drive engagement to action and imitation as a process of learning through socialization, the two theoretical perspectives drive research considering both the supervision role and psychological safety importance in knowledge hiding behavior.

Considering the social exchange theory focus on the importance of communication and mutual gains, findings are consistent with the social nature of tasks and the psychological meaning behind knowledge. Creativity and innovation appear in 15.6% of the articles, bridging theoretical rational au pair with knowledge creation as a socially driven cycle.

Knowledge withholding constructs

Considering overall knowledge withholding-related phenomena, most papers address the knowledge withholding concept of knowledge hiding. Given the conceptual evolution of such a multidimensional construct, Gonçalves and colleagues argue that the higher prevalence of research focused on this knowledge withholding-related construct happens due to its conceptual development.

Knowledge hoarding, while scarcer in the sample (7.7%), is also related to discussions of its robustness as a phenomenon. Nevertheless, the covered papers show several instances where overlapping constructs operate with linguistic differences – in particular between knowledge hiding and knowledge hoarding, and knowledge hiding and knowledge withholding. For example, one study discusses knowledge hoarding as the intentional concealment of knowledge, even upon request. Such rationale, in turn, presents a clash of overlapping definitions commonly attributed to knowledge hiding.

Other positions on different knowledge withholding-related constructs consider hostility toward sharing knowledge. Results also show no evidence of a specific agenda shaping or justifying the growth in knowledge withholding-related phenomena in recent years. Gonçalves and colleagues argue, however, that the progressive awareness of individual-level and organizational-level consequences resulting from knowledge withholding is driving the recent growth of empirical works – an argument reflected in the practical implications of most empirical works found in the sample.

Regardless of overlapping similarities, Gonçalves and colleagues results allow them to build on existing differences that support distinctive behavioral aspects among the identified knowledge withholding-related constructs. Based on the reviewed literature, they propose a differentiation between knowledge withholding-related constructs presenting four dimensions:

1. Individual motivation to protect existing knowledge.
2. Individual behavioral intention to conceal knowledge.
3. Knowledge availability.
4. Request to share knowledge with others.

Discussing their dimension proposal, evidence shows that knowledge disengagement is influenced by feelings of safety, social availability, and job engagement driven by the engagement theory. Similarly, knowledge contribution loafing is related to similar feelings of safety influenced by leadership in organizations. Thus, whereas knowledge disengagement is a consequence of lower levels of engagement, knowledge contribution loafing acts as a behavioral response motivated by knowledge protection.

Evidence also shows that the quasi-accidental nature of knowledge hoarding contrasts with the continuous withholding behavior behind knowledge hiding. Therefore, knowledge hiding portrays a concealment behavior in circumstances where knowledge is requested from others; whereas knowledge hoarding acts as a concealment behavior driven by the protection of knowledge when others do not request knowledge.

Finally, knowledge availability is related to knowledge withholding-related constructs, namely knowledge disengagement and knowledge hiding. Gonçalves and colleagues argue that perceived threats related to a diminished level of available knowledge can motivate and further exacerbate several knowledge withholding behaviors. However, in the case of knowledge disengagement, evidence suggests that individuals showing such behaviors do not share or actively withhold knowledge. The low level of engagement translates into a lack of intention to share knowledge with others, even when communication channels are available in organizations.

Article source and license: Clarifying knowledge withholding: A systematic literature review and future research agenda, CC BY 4.0.

Header image source: krakenimages on Unsplash.

References:

To help provide sound evidence-based foundations for moving forward, KM researchers are starting to explore how generative AI such as ChatGPT can assist KM. One of the first papers¹ to be published takes an initial look at how generative AI may serve as a new context for management theories and concepts, including KM theories. Paper authors Korzynski and colleagues advise that their paper is an opinion piece article and does not refer to empirical data. However, they contend that it’s conclusions can inform further empirical research studies.

Korzynski and colleagues alert that generative AI can play an important role in KM, i.e. the process of collecting, compiling, analyzing and disseminating knowledge within an organization. As a result, the new technology sheds new light on KM theories and frameworks and offers new potential for research studies. Recent studies show that AI can be used, for example, for KM activities including information retrieval, topic modeling, text mining, as well as personalization of e-learning environments, or automatic document summarization.

Given this potential, Korzynski and colleagues go on to discuss how ChatGPT and other generative AI tools can affect the most popular theories of KM.

Nonaka and Takeuchi’s theory of knowledge multiplication points out that the creation of knowledge is a dynamic process where tacit and explicit knowledge is converted as a result of the interplay of socialization, externalization, combination and internalization. Generative AI can help facilitate each of these processes. For example, in socialization, ChatGPT can make the transfer of tacit information through in-person interactions easier by becoming a platform for virtual, distributed teams, allowing their members to share and exchange knowledge and information regardless of location. Previous studies have already confirmed the positive role of AI in knowledge sharing; however, thanks to the conversational mode of ChatGPT, it may bring some new opportunities. Specifically, compared to other chatbots, ChatGPT can generate answers to open-ended questions and provide more personalized responses by adjusting to a user’s language over time. Generative AI may also affect how people assimilate and process new information.

Kolb’s learning model establishes a framework that consists of four stages: concrete experience, reflective observation, abstract conceptualization and active experimentation. Previous AI research has shown that conceptualization is one of the AI capabilities. Compared to previous AI systems, ChatGPT has access to more diverse and larger datasets and is trained on advanced deep learning techniques such as transformer models, which have significantly improved the ability of AI systems to generate coherent and contextually appropriate responses.

Korzynski and colleagues further advise that generative AI can also bring to light several aspects signaled already in the literature. Some of them include issues such as whether the sharing of best practices is still crucial to accumulating and using knowledge in an organization or whether social capital – considered as the resources contained in social relationships – is a critical factor in facilitating knowledge generation. In concluding their KM discussion, Korzynski and colleagues state that ChatGPT raises new questions and offers new perspectives in the creation of not only new frameworks and best practices, but also novel theories and models.

Article source: Generative artificial intelligence as a new context for management theories: analysis of ChatGPT, CC BY 4.0.

Header image source: Jud Mackrill on Unsplash.

Reference: