Knowledge Graphs: Opportunities and Challenges

In plain words

A knowledge graph stores facts as a network: the dots are things (a person, a company, a movie) and the arrows between them say how they relate ("Bill Gates founded Microsoft"). Each fact is a small triple of head, relation and tail. Because the arrows have a direction and a meaning, a computer can follow them to answer questions, suggest items or find documents.

This paper is a survey, not a new method. It walks through where knowledge graphs already help: recommending movies by following links from what you watched to directors and genres, answering questions by hopping from one entity to the next, and making search engines understand what a query is about. It also lists uses in education, research, social media and medicine.

The second half is about what is still hard. Turning facts into vectors loses information such as entity types and long chains of relations. Building graphs from raw text is error-prone, especially across languages and across text and images. Graphs are always incomplete, and most completion methods can only link entities that are already in the graph. Merging graphs needs entities matched and disambiguated, and multi-step reasoning gets expensive on large graphs.

For an engineer, the card is a map of the field and a short reference for a few classic scoring functions, not a recipe to implement.

TL;DR

A 2023 survey organised as opportunities (KG-enhanced AI systems and application fields) and challenges (limitations of KG embeddings, knowledge acquisition, KG completion, knowledge fusion and knowledge reasoning) §1 §2.2. It gives the triple definition, the three families of triple-based embedding methods (tensor factorization, translation, neural network) with scoring functions for SME, ConvKB and an R-GCN layer §5.1 eq. 1–4, and a table of literature Hits@10 / accuracy numbers for 22 embedding models tab. 3. There are no experiments and no code; all numbers are copied from the cited works. Useful as an orientation map, not as an implementation source.

Problem

AI systems need structured world knowledge, and knowledge graphs have become a common way to provide it §1. The survey asks two questions: where KGs bring the largest benefit (AI systems in §3, application fields in §4), and which technical limits block their further development (§5) §1. It positions itself against other surveys by framing the field as opportunities vs challenges §2.2.

Method

The paper is a literature survey; its "method" is a taxonomy of seven research lines §2.2 fig. 2: five technologies (embedding, acquisition, completion, fusion, reasoning) discussed as challenges, and two areas (AI systems, application fields) discussed as opportunities.

Definitions §2.1

• A knowledge base stores facts as triples; drawn as a graph with entities as nodes and relations as edges it is a knowledge graph. The paper treats the two terms as interchangeable, and calls the schema an ontology §2.1.
• The basic unit is the triple (subject, predicate, object) or (head, relation, tail), e.g. (Bill Gates, founderOf, Microsoft). Relations need not be symmetric, so the graph is directed §2.1 fig. 1.
• RDF and labeled property graphs (LPGs) are named as the two usual ways to represent and manage KGs §2.1.
• Example KGs listed: Google Knowledge Graph (2012), DBpedia, Freebase, Facebook's entity graph, Wikidata, YAGO, WordNet §2.1.

Knowledge graph embeddings §5.1

General recipe: embed entities and relations in a dense low-dimensional space, define a scoring function for the plausibility of each triple, and maximise the plausibility of observed facts to learn the embeddings §5.1. Three families:

1. Tensor factorization §5.1.1 fig. 5. Build a binary 3-D tensor \mathcal{X} \in \mathbb{R}^{m \times m \times n} for m entities and n relations, one slice per relation; \mathcal{X}_{ijk}=1 iff the triple (e_i, r_k, e_j) is in the graph, else 0. The tensor is then represented by embedding matrices of entity and relation vectors (the factorization itself is not specified).
2. Translation-based §5.1.2. TransE embeds head, relation and tail as \mathbf{h}, \mathbf{r}, \mathbf{t} and scores a triple so that \mathbf{h} + \mathbf{r} \approx \mathbf{t}. TransH and TransR are named as extensions. The exact distance and loss are not given.
3. Neural network-based §5.1.3: SME, ConvKB and R-GCN, with the equations below.

SME, linear form eq. 1:

f_r(h,t)=(\mathbf{W}_{h1}\mathbf{h}+\mathbf{W}_{h2}\mathbf{r}+\mathbf{b}_h)^\top(\mathbf{W}_{t1}\mathbf{t}+\mathbf{W}_{t2}\mathbf{r}+\mathbf{b}_t)

SME, bilinear form eq. 2:

f_r(h,t)=((\mathbf{W}_{h1}\mathbf{h})\circ (\mathbf{W}_{h2}\mathbf{r})+\mathbf{b}_h)^\top((\mathbf{W}_{t1}\mathbf{t})\circ (\mathbf{W}_{t2}\mathbf{r})+\mathbf{b}_t)

with \mathbf{W} \in \mathbb{R}^{d \times d} weight matrices, \mathbf{b} bias vectors and \circ the element-wise product §5.1.3.

ConvKB eq. 3: the triple is stacked into a three-row matrix \mathbf{A}, convolved with a filter set \Omega, passed through ReLU g, the feature maps concatenated (O) and projected by \mathbf{w} \in \mathbb{R}^{3d}:

f_r(h,t)=O(g(\mathbf{A}*\Omega))\,\mathbf{w}

R-GCN layer eq. 4, for entity k at layer l, with N_k^r the neighbours of k under relation r and n_{k,r} a normalisation constant:

h_k^{(l+1)}=\sigma\Big(\sum_{r\in R}\sum_{i\in N_k^r}\frac{1}{n_{k,r}}W_i^{(l)}h_i^{(l)} +W_k^{(l)}h_k^{(l)} \Big)

# Build the binary KG tensor used by factorization methods  [§5.1.1]
def build_kg_tensor(triples, num_entities, num_relations):
    # one m x m slice per relation, zeros mean "triple absent"  [§5.1.1]
    tensor = zeros(num_entities, num_entities, num_relations)
    for head, relation, tail in triples:
        tensor[head, tail, relation] = 1
    return tensor

# SME linear score of one triple  [eq. 1]
def sme_linear_score(h, r, t, W_h1, W_h2, b_h, W_t1, W_t2, b_t):
    # project head with relation, and tail with relation  [eq. 1]
    left = W_h1 @ h + W_h2 @ r + b_h      # (d,)
    right = W_t1 @ t + W_t2 @ r + b_t     # (d,)
    # inner product of the two projections  [eq. 1]
    return dot(left, right)

# One R-GCN layer as written in the survey  [eq. 4]
def rgcn_layer(hidden, neighbours, norm, weights, self_weight, sigma):
    new_hidden = {}
    for k in entities:
        # self-connection term  [eq. 4]
        total = self_weight[k] @ hidden[k]
        # normalised sum over neighbours, grouped by relation  [eq. 4]
        for r in relations:
            for i in neighbours[k][r]:
                total += (weights[i] @ hidden[i]) / norm[k][r]
        new_hidden[k] = sigma(total)
    return new_hidden

Knowledge acquisition, completion, fusion, reasoning §5.2–§5.5

• Acquisition §5.2: relation, entity and attribute extraction (attribute extraction treated as a special case of entity extraction); structured sources are mapped in with languages such as R2RML §2.2.
• Completion §5.3: train a model on the graph to score candidate triples (link prediction, entity prediction) and add the high-scoring ones, e.g. ranking tails for (Tom, friendOf, ?). Human supervision is described as the gold-standard evaluation §5.3.
• Fusion §5.4: mainly entity or ontology alignment, matching the same entity across graphs; entity disambiguation must precede alignment §5.4.
• Reasoning §5.5: infer a relation r between unconnected h, t \in G. Families: logic rule-based, distributed representation-based, neural network-based. Tasks: single-hop prediction (predict one element of a triple from the other two) and multi-hop reasoning over logical queries §5.5.

Opportunities §3 §4

Summarised in tab. 1 and tab. 2:

Source: tab. 1, condensed.

Claimed advantages of KG-based systems over traditional ones: for recommenders, less data sparsity, cold-start mitigation through relations (a new film linked by (Tenet, has genre of, Sci-Fi)) and explainable recommendation paths §3.1.2; for QA, a smaller search space and multi-hop answers §3.2; for IR, semantic item representation, a reduced search space and relation-based query-document matching §3.3. These are argued qualitatively, not measured.

Hyperparameters

Not applicable: the survey trains no models and reports no hyperparameters §5.1.

Results

The survey runs no experiments. tab. 3 collects results reported by the original papers; per its footnote, all link prediction numbers are filtered. Note that datasets and metrics differ between rows.

Source: tab. 3. The authors single out QuatE (90%), RMNN (89.9%) and KBGAN (89.2%) as the only strong results, and attribute weak results to methods that ignore extra information such as entity types and relation paths §5.1.4; that comparison mixes datasets and metrics (see Pitfalls).

Other quantitative statements: in Freebase, over half of person entities lack birthplace and parents §5.3; most multi-hop reasoning work uses graphs with only 63K entities and 592K relations §5.5; MAG covers more than 248M publications, AMiner Graph more than 200M, AceKG 3 billion triples, AI-KG 800K entities from 330K articles, AIDA KG 21M publications and 8M patents across 66 industrial sectors §4.2.

Paper vs code

No official code.

Pitfalls

• Not a method source. The survey does not give losses, negative sampling, normalisation or training details for any model. Implement TransE, SME, ConvKB or R-GCN from their original papers, not from this card.
• eq. 1 and eq. 2 typeset ⊤ as an infix operator between two vectors; this card reads it as transpose followed by an inner product, giving a scalar score. The survey does not say which \mathbf{W} matrices are shared or learned per relation.
• eq. 4 indexes the neighbour weight by neighbour (W_i^{(l)}), while the text calls R-GCN a "relation-specific transformation" §5.1.3. The equation as printed has no relation-dependent weight; check the original R-GCN paper for the intended indexing before implementing. n_{k,r} is only described as "the normalization process"; its value is not stated.
• ConvKB eq. 3: filter count, filter shape and the layout of \mathbf{A} (rows vs columns) are not stated; \mathbf{w} \in \mathbb{R}^{3d} only matches a specific filter configuration that the survey does not give.
• TransE: only \mathbf{h}+\mathbf{r}\approx\mathbf{t} is given; the norm (L1 or L2), margin and loss are not stated §5.1.2.
• tab. 3 is not a fair leaderboard. It mixes FB15K, WN18, WN18RR and WN11, and Hits@10 with classification accuracy, and the numbers come from the original papers under their own protocols. Ranking QuatE, RMNN and KBGAN together as the best three §5.1.4 compares across datasets and metrics.
• Tensor orientation: §5.1.1 sets \mathcal{X}_{ijk}=1 for triple (e_i, r_k, e_j), so the first two axes are head and tail and the third is the relation. Dense storage needs m^2 n cells, which is impractical for large graphs; the survey does not discuss sparse storage.
• Naming inconsistencies in the source: the health misinformation model is "DETERRENT" in tab. 2 but "DETERREN" in §4.4; COVID-KG is cited to different references in tab. 2 and §4.4; the KEQA example writes "Leonardo" in the question and "Leonard" in the triple §3.2.
• Advantages are asserted, not measured. The benefits claimed for KG-based recommenders, QA and IR §3 carry no quantitative comparison in the survey.

Lineage

• Builds on: earlier KG surveys cited in the paper (e.g. Hogan et al., "Knowledge Graphs"; Ji et al., KG survey; Dai et al., KG embedding survey); none have cards in this wiki.
• Compared against: not applicable (survey).
• Superseded by: not known from content in this wiki.