Knowledge graphs have been conceived to collect heterogeneous data and knowledge about large domains, e.g. medical or engineering domains, and to allow versatile access to such collections by means of querying and logical reasoning. A surge of methods has responded to additional requirements in recent years. (i) Knowledge graph embeddings use similarity and analogy of structures to speculatively add to the collected data and knowledge. (ii) Queries with shapes and schema information can be typed to provide certainty about results. We survey both developments and find that the development of techniques happens in disjoint communities that mostly do not understand each other, thus limiting the proper and most versatile use of knowledge graphs.
Portal Kombat : extension du réseau de propagande russe
Knowledge graphs for knowing more and knowing for sure
1. KI – Institute for Artificial Intelligence
Knowledge graphs for
knowing more and knowing for sure
Steffen Staab
@ststaab
https://www.ki.uni-stuttgart.de
https://semanux.com
https://southampton.ac.uk/research/institutes-centres/web-internet-science
2. provides an international forum for presentation and
discussion of research on information and knowledge
management, as well as recent advances on data and
knowledge bases.
2
Conference on Information and Knowledge Management
Rather: Conference on Large Language Models?
Let’s explore the role of knowledge bases/graphs!
3. 1. What is a Knowledge Graph?
2. Some Applications of Knowledge Graphs
3. Knowledge Graphs for Knowing for Sure
4. Knowledge Graphs for Knowing More
5. Large Language Models as Knowledge Bases
6. Large Language Models as AI Assistants
3
Plan for my talk
5. What is a Knowledge Graph?
A model for knowledge structures with
5
C22.0
Patient2342
treatedBy
„liver tumor“ / „PhValue 7.5“
Concepts
Entities
Relations
Labels / Values
6. Queries
• Scalability to
billions of facts
• Answering with
• facts
• predictions
• recommendations
6
What does a knowledge graph do for us?
What are the difficulties?
Example from medical project:
• Foundational Model of Anatomy:
75.000 concepts, 120.000 labels,
> 2 Mio facts
• Not even patient data yet!
7. Queries
Ontologies & Facts
• How to develop and integrate
ontologies?
• How to provide facts?
• Reasoning?
• Learning?
• Guarantees?
7
What does a knowledge graph do for us?
What are the difficulties?
Example from medical project:
• Foundational Model of Anatomy
• RadLex
• ICD-10
8. Queries
Ontologies & Facts
What can be represented?
• Provenance
• Uncertainty
• Time
• …
8
What does a knowledge graph do for us?
What are the difficulties?
Example from medical project:
• Patient history
• patient measurements
11. 02.11.2020
Steffen Staab, Universität Stuttgart, @ststaab, https://www.ipvs.uni-stuttgart.de/departments/ac/ 11
Wonderful ressource
– but not representative
13. Application 2: KG for Circular Factory
Product
Production
Co-
Design
Knowledge Graph contains knowledge about design, production and product
including plans, sensor measurements and intra-logistics
17. • Updates and deletions with dependencies [EKAW18],
also at the ontological level [KR2020]
• Federation [WWW08]
• Lacking views with deletions and updates
• Transaction locking [ESWC2013]
• Lacking recent standards (SHACL) and optimistic schemes
• Uncertainties
• Managing identities
(„does re-designed column preserve its identity?“)
• ...
17
Applications Imply Wealth of Requirements
rudimentary
available
research
18. Encyclopedic KGs
• Facts are reported often
• Who is Douglas Adams?
• What is the capital of France?
• Head of distribution of world
knowledge on the Web
• Answers with high precision
retrieval desired
Engineering KGs
• Point facts exist once
• w3476 instOf AngleGrinder
• faceGear4223
maxDeviation 0.3mm
• Processes are important
• Answers must be correct
18
Sliding scale of knowledge graph requirements
Currently fashionable
research
“we build a system”
under-researched
19. A lot of research in Knowledge Graphs builds on the
assumption that we want to query encyclopedias
but we have many other requirements in industry.
19
Observation 1
21. KG 1 KG 2 KG 3
App A App B App C
Scenario in Architecture, Engineering, Construction
(AEC)
22. 22
SOLID Project
https://solidproject.org/
• people store their data
securely in decentralized
data stores - Pods
• people control access to
the data in their Pod
• standard, open, and
interoperable data
formats and protocols
Focus:
authentication &
authorization
23. KG 1 KG 2 KG 3
App A App B App C
Can my app B work on my KG2?
24. Example: How old are the students?
Query for all students, access age
Query fails during evaluation
let students = query { SELECT ?x WHERE {?x a Student. } }
for student in students do
printfn „%A“ (student.age)
bob
alice 𝑏1
Student
University
subClass
type
studiesAt
type
211... "Bob"
matrNr name
25 "Alice"
age name
Person
[ESOP17,ISWC19]
25. Example: How old are the students?
let students = query { SELECT ?x WHERE {?x a Student. } }
for student in students do
printfn „%A“ (student.age)
Should we use this relation on this signifier?
Depends on:
1. Conceptualization of data source
2. Querying of data source
3. Software code
bob
alice 𝑏1
Student
University
subClass
type
studiesAt
type
211... "Bob"
matrNr name
25 "Alice"
age name
Person
26. Closed-world conceptualization of classes and relations
SHACL – SHApes Constraint Language
• SHACL shapes are integrity constraints
• Namespaces omitted for brevity
:StudentShape a :NodeShape;
:targetClass :Student;
:class :Person;
:property [
:path :studiesAt;
:minCount 1;
:class :University;
].
:PersonShape a :NodeShape;
:targetClass :Person;
:property [
:path :name;
:minCount 1;
:datatype xsd:string;
].
27. Closed-world conceptualization of code (1)
Type checking discovers (potential) run-time errors
let students = query { SELECT ?x WHERE {?x a Student. } }
for student in students do
printfn „%A“ (student.age)
Set of all students (StudentShape)
One value of
StudentShape
set
Not allowed since
StudentShape ⊈ ≥𝟏age.⊤
when considering
all conceptually possible RDF graphs
28. Closed-world conceptualization of code (2)
• Access: matrNr
• No error during evaluation
• Unsafe: Rejected by type checking,
conceptualization not guaranteed
let students = query { SELECT ?x WHERE {?x a Student. } }
for student in students do
printfn „%A“ (student.matrNr)
bob
alice 𝑏1
Student
University
subClass
type
studiesAt
type
211... "Bob"
matrNr name
25 "Alice"
age name
Person
29. Closed-world conceptualization of code (3)
• Query for: matrNr
• Type safe access:
matrNr inferred to be given for all values of student
let students = query { SELECT ?x WHERE {?x matrNr ?y. } }
for student in students do
printfn „%A“ (student.matrNr)
bob
alice 𝑏1
Student
University
subClass
type
studiesAt
type
211... "Bob"
matrNr name
25 "Alice"
age name
Person
[ESOP17,ISWC19]
30. 1. Use available SHACL constraints
2. Infer additional SHACL constraints from queries
3. Type check using inference
Determine type safety
let students = query { SELECT ?x WHERE {?x a Student. } }
for student in students do
printfn „%A“ (student.name)
Query shape(2) including StudentShape (1)
One value of
StudentShape
set
StudentShape ⊆ PersonShape and
PersonShape ⊆ ≥1name. ⊤ in all possible graphs
Inference (3)
[ESOP17,ISWC19]
31. KG 1 KG 2 KG 3
App A App B App C
Scenario: Can my app B work on my view of KG1?
32. 32
Shapes to Shapes
KG 1 KG 2
App B
Input Shape
Sin = { :Person ⊑ :Agent
}
Input query defining view
q = CONSTRUCT {
?x a :Person .
?y a :Agent
} WHERE {
?x a :Person .
?y a :Agent
“Every Person
is an Agent” Output Shapes
“Which data can App B expect?”
s2s(Sin, q) → Sout
view
[Seifer2023]
33. Tracing Query Concepts (and Relations)
Sin = { :Person1 ⊑
:Agent }
q = CONSTRUCT {
?x a :Person3 .
?y a :Agent
} WHERE {
?x a :Person2 .
?y a :Agent
}
Sout = { :Person3 ⊑
:Agent }
Are concepts
:Person1
:Person2
:Person3
the same?
33
Yes!
[Seifer2023]
34. Tracing Query Concepts (and Relations)
Sin = { :Person1 ⊑ :Agent
}
q = CONSTRUCT {
?x a :Person3 .
?y a :Agent
} WHERE {
?x a :Person2 .
?x a :Teacher .
?y a :Agent
}
Sout = { :Person3 ⊑ :Agent
}
Are concepts
:Person1
:Person2
:Person3
still the same?
34
NO!
[Seifer2023]
Hard problem even for restricted
query and constraint languages
35. KG problems occur at ontological and at fact level.
Knowledge Graph technologies lack crucial capabilities
for guaranteeing results.
35
Observation 2
39. 39
Finding and Exploiting Patterns of Similarity & Analogy
Stuttgart
Area
worksFor
locatedIn
Koblenz
Area
Wolv.
Area
Steffen
Frank
Ingo
birthdate
livesIn
prediction impossible prediction possible
40. Correct [2013]:
“TransE significantly outperforms state-of-the-art methods in link
prediction on two knowledge bases.”
Misleading:
“Our work focuses on modeling multi-relational data from KBs
(Wordnet [9] and Freebase [1] in this paper), with the goal of
providing an efficient tool to complete them by automatically
adding new facts, without requiring extra knowledge.”
A. Bordes et al. [TransE 2013]
Knowing More than What is Stated in a Knowledge Graph
41. Geometric Reasoning with EL Ontology A-Box
Concept assertion 𝐶(𝑎)
𝑎
𝐶
Geometric membership
[ISWC2022]
42. 42
Geometric Reasoning with EL Ontology T-Box
Box affine
transformation
Box entailment Box intersection
Box disjointedness
[ISWC2022]
43. Geometric Reasoning with EL Ontology A-Box
4
3
Concept assertion 𝐶(𝑎)
𝑎
𝐶
r(𝑎, 𝑏)
𝑇𝑟
𝑏
𝑎
Role assertion
Geometric membership
Affine transformation
between two points
[ISWC2022]
44. 44
Geometric Reasoning with Fact Attributions in ShrinkE
• Modeling primal triple as a spatial spanning (from a point to a box)
• Modeling qualifiers as a spatial (monotonically) shrinking of the box
• Qualifier implication and exclusion are geometrically modeled as
box containment and disjointedness
[ACL2023]
Check out https://kg-beyond-triple.github.io/
45. 45
Geometric Reasoning with Fact Attributions in ShrinkE
[ACL2023]
• Box embedding
• Box shrinking is a box-to-box transform that monotonically shrinks the size
46. • WD50k: excerpt from Wikidata
• JF17K: excerpt from Freebase
• WikiPeople: excerpt from Wikidata
• FB15k-237: excerpt from Freebase
• …
Datasets for evaluating knowledge graph embeddings
Many datasets, but all biased in the same direction
47. 02.11.2020
PhD thesis in preparation by Fabian Sasse, KIT 47
Selecting manufacturing measurement technology
in immature production processes
49. Knowledge Graph embedding techniques
do not complete knowledge graphs,
they perform similarity and analogical reasoning.
Evaluations of Knowledge Graph embedding methods
remain biased towards encyclopedic knowledge.
49
Observation 3
54. Statistically frequent knowledge
• Commonsense knowledge:
• “cows eat grass”
• “apples fall towards earth if
unsupported”
• Commonsense expert
knowledge
• “halting problem is undecidable”
• “3SAT is NP-complete”
“Point knowledge”
• Steffen Staab is a professor
at University of Stuttgart
54
Knowledge in text
55. • Smoothing a data set: create an approximating function that
preserves patterns in the data, while leaving out noise or fine-
scale structures. [Shortened from Wikipedia]
• Laplacian smoothing for Naïve Bayes:
argmax𝑐 𝑃 𝑐 𝑥1, … , 𝑥𝑛 ≈ argmax𝑐𝑃 𝑐 𝑃 𝑥1 𝑐 ⋯ 𝑃 𝑥𝑛 𝑐
• Smoothing for language models [ACL14]
𝑃 𝑤𝑛 𝑤𝑛−𝑘 ⋯ 𝑤𝑛−1
must not be 0 for unobserved 𝑤𝑛−𝑘 ⋯ 𝑤𝑛−1 𝑤𝑛
55
Language models smoothen probability distributions
must not be 0
56. • What other terms could appear in a masked position?
• High “temperature” → diversity of answers
• Varying answers for “Write a poem about <your name>”
56
Smoothing is the core task of Large Language Models
63. 63
Few shot in context learning on KB question answering
Tianle Li, Xueguang Ma, Alex Zhuang, Yu Gu, Yu Su,
and Wenhu Chen. 2023. Few-shot In-context Learning
on Knowledge Base Question Answering. In ACL-2023
67. Knowing for Sure
• Research required for
dealing with federated,
overlapping KGs with
multiple authorities
Knowing More
• Know what you get and
evaluate not only with
encyclopedic KGs
LLMs as knowledge bases
• Commonsense knowledge
• Frequently observed
knowledge
LLMs as AI assistants
• entering and retrieving
“point knowledge”
Do not (always) go with the flow
68. Thank you!
E-Mail
www.
Universität Stuttgart
KI – Institute for Artificial Intelligence
Universitätsstraße 32, 70569 Stuttgart
Steffen Staab
ki.uni-stuttgart.de
Analytic Computing, KI
Steffen.staab@ki.uni-stuttgart.de
Many thanks go to my
PhD students, PostDocs and
collaborators who made the work
possible portrayed in this talk
check out references!
I hire
PostDoc & PhD student
for circular factory project!
69. 1. [Potyka23] Nico Potyka, Yuqicheng Zhu, Evgeny Kharlamov and Steffen Staab.
Uncertainty-aware Knowledge Extraction from Large Language Models using
Social Choice Theory. TechReport.
2. [ISWC2022] B. Xiong, N. Potyka, T.-K. Tran, M. Nayyeri, S. Staab. For “Faithful
Embeddings for EL++ Knowledge Bases”. In: 21st International Semantic Web
Conference (ISWC2022)
3. [SIGIR23] J. Lu, J. Shen, B. Xiong, W. Ma, S. Staab, C. Yang. HiPrompt: Few-
Shot Biomedical Knowledge Fusion via Hierarchy-Oriented Prompting. In:
Proceedings of ACM SIGIR-2023, Taipei, Taiwan, July 23-27, 2023.
4. [ISWC2023] M. Nayyeri, Z. Wang, M. M. Akter, M. Mohtashim, Md R. Al Hasan
Rony, J. Lehmann, S. Staab. Integrating Knowledge Graph Embeddings and Pre-
trained Language Models in Hypercomplex Spaces. In: 22nd Int. Semantic Web
Conference (ISWC2023), Athens, GR, November 6-10, 2023.
5. [TransE 2013] Bordes, Antoine, et al. "Translating embeddings for modeling
multi-relational data." Advances in neural information processing systems 26
(2013).
References related to Knowing More
70. 1. [ISWC19] M. Leinberger, P. Seifer, C. Schon, R. Lämmel, S. Staab. Type Checking Program Code using SHACL. In: Proc.
of Int. Semantic Web Conference (ISWC-2019). Auckland, New Zealand, October 2019.
2. [Seifer2023] Philipp Seifer, Daniel Hernández, Ralf Lämmel, Steffen Staab. From Shapes to Shapes: Inferring SHACL
Shapes for Results of SPARQL CONSTRUCT Queries. TechReport.
3. [ESOP17] M. Leinberger, R. Lämmel, S. Staab. The essence of functional programming on semantic data. In 26th
European Symposium on Programming (ESOP 2017), Uppsala, SE, 22 - 29 Apr 2017, pp. 750-776.
4. [CAAD Futures 2023] D. Elshani, D. Hernandez, A. Lombardi, L. Siriwardena, T. Schwinn, A. Fisher, S. Staab, A. Menges,
T. Wortmann. Building Information Validation and Reasoning Using Semantic Web Technologies. In: Computer-Aided
Architectural Design. CAAD Futures 2023. Springer, Cham, 2023.
5. [KR2020] T. Rienstra, C. Schon, S. Staab. Concept Contraction in the Description Logic EL. In: Principles of Knowledge
Representation and Reasoning: Proceedings of the Seventeenth International Conference, KR 2020, pp. 723-732.
6. [EKAW18] C. Schon, S. Staab, P. Kügler, P. Kestel, B. Schleich, S. Wartzack. Metaproperty-guided Deletion from the
Instance-Level of a Knowledge Base. In: Proc. of EKAW 2018, 21st International Conference on Knowledge Engineering
and Knowledge Management, November 12-16, 2018, Nancy, France, Springer 2018.
7. [ESWC2013] S. Scheglmann, S. Staab, M. Thimm, G. Gröner. Locking for Concurrent Transactions on Ontologies. In: 10th
Extended Semantic Web Conference (ESWC2013), Montpellier, France, May 26-30, 2013.
8. [WWW08] S. Schenk, S. Staab. Networked Graphs: A Declarative Mechanism for SPARQL Rules, SPARQL Views and RDF
Data Integration on the Web. In: Proc. of WWW-2008, 17th Int. World Wide Web Conference, Bejing, China, April 21-25,
2008, pp. 585-594.
References related to Knowing for Sure
71. [ACL14] R. Pickhardt, T. Gottron, M. Körner, P. G. Wagner, T. Speicher, S. Staab. A Generalized Language Model as the
Combination of Skipped n-grams and Modified Kneser Ney Smoothing. In: Proc. of ACL-2014 - The 52nd Annual Meeting of the
Association for Computational Linguistics. Baltimore, June 22-27, 2014.
02.11.2020
Steffen Staab, Universität Stuttgart, @ststaab, https://www.ipvs.uni-stuttgart.de/departments/ac/ 71
Others
Editor's Notes
If it looks like a duck, walks like a duck and quacks like a duck, then it just may be a duck.
Huey, Dewey, and Louie live in Duckburg
If it looks like a duck, walks like a duck and quacks like a duck, then it just may be a duck.
Huey, Dewey, and Louie live in Duckburg
If it looks like a duck, walks like a duck and quacks like a duck, then it just may be a duck.
Huey, Dewey, and Louie live in Duckburg
750 million triples, fast growing, not easy to manage
status: proposal for funding by 20 PIs, mostly engineering, mostly from KIT
7 year excellence cluster at Uni Stuttgart
medical knowledge graphs and applications may be found on either side
Now I am gonna to present those geometric interpretations and the corresponding loss term for each axiom.
In Abox, we have two types of axioms: Concept assertion and Role assertion r(a, b).
For concept assertion, the geometric interpretation is that the point of instance a should be inside the box of the class C. That means that our loss should enforce every dimension of the point a to be between the low-left corner of box C and upper-right corner of box C.
We also have role assertion r(a, b) saying that a has a relation r with b, the geometric interpretation is that the point a, after a affine transformation of r, should be near the point of b.
The corresponding loss term can be defined by minimizing the L2 distance between the transformed point of a and the point b.
We proved that our terms satisfy the soundness guarantees that means our loss terms are zero if and only if the corresponding geometric interpretations are satisfied.
Now I am gonna to present those geometric interpretations and the corresponding loss term for each axiom.
In Abox, we have two types of axioms: Concept assertion and Role assertion r(a, b).
For concept assertion, the geometric interpretation is that the point of instance a should be inside the box of the class C. That means that our loss should enforce every dimension of the point a to be between the low-left corner of box C and upper-right corner of box C.
We also have role assertion r(a, b) saying that a has a relation r with b, the geometric interpretation is that the point a, after a affine transformation of r, should be near the point of b.
The corresponding loss term can be defined by minimizing the L2 distance between the transformed point of a and the point b.
We proved that our terms satisfy the soundness guarantees that means our loss terms are zero if and only if the corresponding geometric interpretations are satisfied.
[Hypertext2008]
Yulan talked in her keynote about voting in order to improve confidence – though I also have observed non-i.i.d. behaviour and then voting may be bad
usefulness may be an issue
In the following I will give you webpage content about a soccer club. Represent the facts that you find in this text in RDF turtle notation. Effizienter VfB siegt bei Union Berlin Die Siegesserie des VfB geht weiter. Beim 1. FC Union Berlin setzt sich die Mannschaft mit dem Brustring mit 3:0 durch. Es ist der sechste Erfolg in Serie und der erste gegen Union in der Bundesliga. Der Spielverlauf: Der VfB ging mit einer auf zwei Positionen veränderten Startformation in das Duell beim 1. FC Union Berlin. Maxi Mittelstädt und Dan-Axel Zagadou begannen für Pascal Stenzel sowie Hiroki Ito (beide Bank). Die Mannschaft mit dem an diesem Tag schwarzen Brustring startete selbstbewusst in die Partie und hatte in der Anfangsviertelstunde deutlich mehr Ballbesitz. Die höheren Spielanteile münzte der VfB schnell in die verdiente Führung um. Wer sonst als Serhou Guirassy hätte der Torschütze zum 1:0 sein können (siehe „Die Tore“). Die Jungs aus Cannstatt kontrollierten die Partie auch in der Folge, musste nach knapp einer halben Stunde aber schon wechseln. Serhou Guirassy verließ den Platz angeschlagen mit muskulären Problemen im hinteren linken Oberschenkel, Deniz Undav kam für ihn in die Partie. Der VfB war dennoch bis zum Pausenpfiff das tonangebende Team. Silas und Deniz Undav sorgen für die Entscheidung Nach dem Wiederanpfiff entwickelte sich eine umkämpfte Partie mit vielen Situationen zwischen den Strafräumen. Klare Torchancen konnte sich zunächst keines der Teams erspielen. In der 60. Minute hatte Jamie Leweling jedoch die große Chance, auf 2:0 zu erhöhen. Der 22-Jährige scheiterte in aussichtsreicher Position frei vor dem Tor an Unions Torhüter Frederik Rönnow. In der 77. Minute war Alexander Nübel auf der Gegenseite hellwach und klärte die Situation gegen den heranstürmenden Kevin Behrens. Kurz darauf sorgte der VfB mit einem Konter für das beruhigende 2:0. Der eingewechselte Silas war mit seinem dritten Saisontor erfolgreich. Den Endstand zum 3:0 stellte Deniz Undav mit einem Kopfball her. Der VfB siegte am Ende verdient, weil er seine Chancen konsequent nutzte und über die gesamte Spielzeit hinweg kaum Chancen des Gegners zuließ. Den gesamten Spielverlauf im VfB-Liveticker nachlesen. Die Tore: 16. Minute: Serhou Guirassy köpft nach einer Flanke von Anthony Rouault zum 1:0 ein. Es ist das 14. Saisontor des VfB-Stürmers. 81. Minute: Silas kommt über Karazor und Millot an den Ball, setzt sich gegen die aufgerückten Union-Verteidiger durch und schließt überlegt zum 2:0 ab. 88. Minute: Der VfB erobert in Höhe des gegnerischen Strafraums den Ball, Wooyeong Jeong flankt von rechts auf Deniz Undav, der zentral zum 3:0 einköpft. Die Stimmen: VfB-Cheftrainer Sebastian Hoeneß: „Es war eine reife Leistung von uns. Wir haben sehr erwachsen gespielt. Die Druckphasen des Gegners waren nie so sehr ausgeprägt. Dass wir das Spiel am Ende so klar auf unsere Seite ziehen, macht mich stolz. Wir haben aktuell einen Lauf und den wollen wir so lange wie möglich mitnehmen.“ Chris Führich: „Es war eine Riesenteamleistung von uns. Wir wussten, wie schwierig es ist, hier zu gewinnen. Wir haben von der ersten bis zur letzten Minute unsere Taktik durchgezogen. Es ist auch sehr wichtig gegen die lange Bälle und wuchtigen Spieler von Union gut zu stehen. Das ist uns gut gelungen und wir haben letztlich auch verdient gewonnen.“ Maxi Mittelstädt: „Es ist ein schönes Gefühl, dass wir gewonnen haben. Wir haben eine reife Leistung gezeigt und wenig anbrennen lassen. Natürlich gab es auch Phasen, die wir überstehen mussten. Wir haben einen breiten und starken Kader. Auch heute war wichtig, welche Impulse die Einwechselspieler in die Partie gebracht haben. Das Kollektiv macht uns aktuell stark. Manchmal muss man sich angesichts der Siegesserie kneifen, aber wir haben uns das auch über die vergangenen Monate und Wochen erarbeitet. Ich freue mich auf die kommenden Herausforderungen.“ Die Besonderheiten: Der gebürtige Berliner Maximilian Mittelstädt gab sein Startelfdebüt für den VfB. Neu-Nationalspieler Chris Führich machte an diesem Samstag sein 75. Pflichtspiel im Trikot mit dem roten Brustring. Der ehemalige VfBler Rani Khedira gab auf Seiten der Berliner gegen den Club aus Cannstatt sein Comeback nach einer längeren Wadenverletzung. Der Schiedsrichter der Partie Bastian Dankert und seine Assistenz René Rohde leiteten ihr jeweils 150. Bundesligaspiel. 21 Punkte nach acht Spieltagen hat der VfB in seiner Vereinshistorie noch nie auf dem Konto gehabt. Die nächsten Spiele: Am kommenden Samstag empfängt der VfB die TSG Hoffenheim in der MHPArena. Dieses Spiel ist bereits ausverkauft, ebenso wie die Heimpartie gegen Borussia Dortmund am 11. November. Der Mitgliedervorverkauf für das Pokal-Heimspiel gegen den 1. FC Union Berlin am Dienstagabend, 31. Oktober, 18 Uhr läuft, genauso wie für die Heimbegegnung gegen den SV Werder Bremen am Samstag, 2. Dezember, 18:30 Uhr. Zum VfB-Onlineshop.