School of Computing Science

Deploying Large Language Models for Industrial Recommender Systems: From Effectiveness Improvement to Efficiency Optimization

Group: Information Retrieval (IR)
Speaker: Jianghao Lin, Shanghai Jiaotong University
Date: 09 December, 2024
Time: 15:00 - 16:00
Location: 422 SAWB - https://uofglasgow.zoom.us/j/84369571835?pwd=bWTbM95W8lWzGK6NsJZltvUNLoOJT2.1

Title: Deploying Large Language Models for Industrial Recommender Systems: From Effectiveness Improvement to Efficiency Optimization

Abstract: The advent of large language models (LLMs), particularly with the release of ChatGPT in late 2022, has trigerred significant advancements in recommender systems [1]. Early research on LLM-enhanced RecSys predominantly concentrated on improving the performance and effectiveness of recommendations, by leveraging the open-world knowledge and reasoning capabilities of LLMs. However, approaching the end of 2024, we would like to highlight that the focus of research is shifting towards a critical issue: efficiency optimization. As LLMs become increasingly integrated into real-world large-scale recommendation platforms, the need for optimizing their computational efficiency, response time, and scalability has become paramount. That is, from effectiveness improvement to efficiency optimization.
In this talk, we will first introduce two of our earlier works on the effectiveness improvement of LLM-enhanced RecSys [3][4]. Then, we will further introduce two of our new research works on the efficiency optimization for LLM-enhanced RecSys [5].

[1] How Can Recommender Systems Benefit from Large Language Models: A Survey
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2306.05817&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156426923%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=t%2FNCy%2BJpWegeo8rJ4C1I1EeGRr7hKb3b3jjNuAmCem0%3D&reserved=0

[2] Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2306.10933&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156436481%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=K8PSuf4uXdWGent2TcRiNBUDl5KF1o%2BvUa1xoM6fbwc%3D&reserved=0

[3] ReLLa: Retrieval-enhanced Large Language Models for Lifelong Sequential Behavior Comprehension in Recommendation
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2308.11131&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156445784%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=1a7%2FzsWNGZy9ycM2YqblhXwxjQhXwuTCKG5%2BOYysQmM%3D&reserved=0

[4] A Decoding Acceleration Framework for Industrial Deployable LLM-based Recommender Systems
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2408.05676&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156454662%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=%2B7Y2%2F%2FjWK0UyioHyjZJsojcK%2FonHXyNROG6ft9WnGzs%3D&reserved=0

[5] LIBER: Lifelong User Behavior Modeling Based on Large Language Models
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2411.14713&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156463235%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=7jeVbLnaUhofN1Y8HMB5BDZbT%2FI4UDV1NEyK8ms%2Febg%3D&reserved=0

Bio: Jianghao Lin is a forth-year PhD student at Shanghai Jiao Tong University supervised by Prof. Weinan Zhang. His research interests focus on large language models and their applications in real-world data mining scenarios including recommender system and information retrieval. He has published 23 papers in data-mining-related research areas, resulting in 624 citations (Google Scholar) with an H-index of 11. He works closely with various corporations and has achieved several industrial deployments for recommender systems.

[FATA Seminar] Bigraphs as a theoretical base for verifying interactive systems

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Cyril ALLIGNOL and Celia PICARD, Ecole Nationale de l'Aviation Civile
Date: 10 December, 2024
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

Abstract: Interactive systems are multiplying, including in critical aeronautical systems such as aircraft cockpits and air traffic control systems. At the same time, UIDLs, the languages that enable the development of such systems, are also on the rise. However, to date, these languages offer no guarantees, neither in terms of compilation nor on the systems developed. Hence the need for a verified programming framework for UIDLs. This is the main subject of our research.

In this talk, we will present the results we have obtained so far and our ongoing work, namely: using bigraphs to express the semantics of UIDLs; defining a formal minimal UIDL based on bigraphs; formalising the bigraph theory in the Coq proof assistant. We will also discuss our future work around the verification of interactive systems, including the verification of interactive properties.

-----------------------------------

This event is part of the FATA Weekly Seminar, which takes place every Tuesday from 3:00 - 4:00 PM in Room 422, Sir Alwyn Williams Building and on Zoom https://uofglasgow.zoom.us/j/83611964233?pwd=CgRyzxK8Z9fP2ULTb5ONWZeUYx2t2E.1

[FATA Seminar] Formally verified hardening of C programs against fault injection

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Basile PESIN, Ecole Nationale de l'Aviation Civile
Date: 10 December, 2024
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

Abstract: Fault attacks allow malicious actors to modify the behavior of a program by
physically injecting a fault in the hardware. They typically target sensitive
applications such as cryptography services, authentication or boot-loader and
firmware updater. They can be defended against by adding countermeasures, that
is control flow checks and redundancies, either in the hardware, or in the
software running on it. In particular, software countermeasures may be added
automatically during compilation.

In this talk, we will describe a formally verified implementation of this
approach in the CompCert verified compiler for the C language. We proposed a
toolkit to implement countermeasures as transformations of a middle-end
representation of CompCert, RTL.  We applied this toolkit to two existing
countermeasures that protect the control flow of the program.  We proved that
these countermeasures are correct, that is, they do not change the observable
behavior of the program during an execution without fault injection. We then
modeled the effect of a fault on the behavior of the program as an extension of
the semantic model of RTL. We used this new model to formally prove the efficacy
of the countermeasure: all attacks are caught. In addition to this formal
reasoning, we evaluated the protected program using Lazart, a tool for symbolic
fault injection.

-----------------------------------

This event is part of the FATA Weekly Seminar, which takes place every Tuesday from 3:00 - 4:00 PM in Room 422, Sir Alwyn Williams Building and on Zoom https://uofglasgow.zoom.us/j/83611964233?pwd=CgRyzxK8Z9fP2ULTb5ONWZeUYx2t2E.1

GIST Seminar - Bike to the Future: Cyclists and Automated Vehicles

Group: Human Computer Interaction (GIST)
Speaker: Ammar Jamal Said Al Taie, University of Glasgow
Date: 12 December, 2024
Time: 13:00 - 14:00
Location: Sir Alwyn Williams Building, 423 Seminar Room

Summary:

Cyclists are vulnerable road users who must share the road with motorised vehicles. They encounter vehicles across a range of traffic scenarios, including roundabouts or lane merging. These encounters result in space-sharing conflicts that require some social interaction to be resolved. For example, a driver may wave their hand to signal a cyclist to proceed, or a cyclist may extend their arm to indicate their intention to merge lanes with a driver. Autonomous vehicles will soon be on our roads, and these social interactions will diminish. This would create new ambiguities for cyclists (and other road users) in future traffic. In this talk, I will present work conducted over the past three years to understand the scenarios that trigger AV-cyclist interaction, design potential interfaces that facilitate these interactions, and evaluate these interfaces using novel study designs. The contributions are in the interface designs, and new methodologies to design and evaluate them. This is important to ensure safer, more pleasant future cycling.

Bio:

My area of research is Autonomous Vehicle-Cyclist interaction. This often involves utilising unconventional technologies, such as new displays on the car's exterior. I am a "hands-on" researcher; most of my work is conducted in real-world settings using new technologies such as eye-tracking. In my free time, I enjoy running, cycling, reading graphic novels and playing mario kart!

Deploying Large Language Models for Industrial Recommender Systems: From Effectiveness Improvement to Efficiency Optimization

Group: Information Retrieval (IR)
Speaker: Jianghao Lin, Shanghai Jiaotong University
Date: 09 December, 2024
Time: 15:00 - 16:00
Location: 422 SAWB - https://uofglasgow.zoom.us/j/84369571835?pwd=bWTbM95W8lWzGK6NsJZltvUNLoOJT2.1

Title: Deploying Large Language Models for Industrial Recommender Systems: From Effectiveness Improvement to Efficiency Optimization

Abstract: The advent of large language models (LLMs), particularly with the release of ChatGPT in late 2022, has trigerred significant advancements in recommender systems [1]. Early research on LLM-enhanced RecSys predominantly concentrated on improving the performance and effectiveness of recommendations, by leveraging the open-world knowledge and reasoning capabilities of LLMs. However, approaching the end of 2024, we would like to highlight that the focus of research is shifting towards a critical issue: efficiency optimization. As LLMs become increasingly integrated into real-world large-scale recommendation platforms, the need for optimizing their computational efficiency, response time, and scalability has become paramount. That is, from effectiveness improvement to efficiency optimization.
In this talk, we will first introduce two of our earlier works on the effectiveness improvement of LLM-enhanced RecSys [3][4]. Then, we will further introduce two of our new research works on the efficiency optimization for LLM-enhanced RecSys [5].

[1] How Can Recommender Systems Benefit from Large Language Models: A Survey
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2306.05817&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156426923%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=t%2FNCy%2BJpWegeo8rJ4C1I1EeGRr7hKb3b3jjNuAmCem0%3D&reserved=0

[2] Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2306.10933&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156436481%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=K8PSuf4uXdWGent2TcRiNBUDl5KF1o%2BvUa1xoM6fbwc%3D&reserved=0

[3] ReLLa: Retrieval-enhanced Large Language Models for Lifelong Sequential Behavior Comprehension in Recommendation
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2308.11131&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156445784%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=1a7%2FzsWNGZy9ycM2YqblhXwxjQhXwuTCKG5%2BOYysQmM%3D&reserved=0

[4] A Decoding Acceleration Framework for Industrial Deployable LLM-based Recommender Systems
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2408.05676&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156454662%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=%2B7Y2%2F%2FjWK0UyioHyjZJsojcK%2FonHXyNROG6ft9WnGzs%3D&reserved=0

[5] LIBER: Lifelong User Behavior Modeling Based on Large Language Models
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fabs%2F2411.14713&data=05%7C02%7Cz.yi.1%40research.gla.ac.uk%7C935b3a5e18114fdc0bb008dd14868218%7C6e725c29763a4f5081f22e254f0133c8%7C1%7C0%7C638689290156463235%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=7jeVbLnaUhofN1Y8HMB5BDZbT%2FI4UDV1NEyK8ms%2Febg%3D&reserved=0

Bio: Jianghao Lin is a forth-year PhD student at Shanghai Jiao Tong University supervised by Prof. Weinan Zhang. His research interests focus on large language models and their applications in real-world data mining scenarios including recommender system and information retrieval. He has published 23 papers in data-mining-related research areas, resulting in 624 citations (Google Scholar) with an H-index of 11. He works closely with various corporations and has achieved several industrial deployments for recommender systems.

[FATA Seminar] Bigraphs as a theoretical base for verifying interactive systems

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Cyril ALLIGNOL and Celia PICARD, Ecole Nationale de l'Aviation Civile
Date: 10 December, 2024
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

Abstract: Interactive systems are multiplying, including in critical aeronautical systems such as aircraft cockpits and air traffic control systems. At the same time, UIDLs, the languages that enable the development of such systems, are also on the rise. However, to date, these languages offer no guarantees, neither in terms of compilation nor on the systems developed. Hence the need for a verified programming framework for UIDLs. This is the main subject of our research.

In this talk, we will present the results we have obtained so far and our ongoing work, namely: using bigraphs to express the semantics of UIDLs; defining a formal minimal UIDL based on bigraphs; formalising the bigraph theory in the Coq proof assistant. We will also discuss our future work around the verification of interactive systems, including the verification of interactive properties.

-----------------------------------

This event is part of the FATA Weekly Seminar, which takes place every Tuesday from 3:00 - 4:00 PM in Room 422, Sir Alwyn Williams Building and on Zoom https://uofglasgow.zoom.us/j/83611964233?pwd=CgRyzxK8Z9fP2ULTb5ONWZeUYx2t2E.1

[FATA Seminar] Formally verified hardening of C programs against fault injection

Group: Formal Analysis, Theory and Algorithms (FATA)
Speaker: Basile PESIN, Ecole Nationale de l'Aviation Civile
Date: 10 December, 2024
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

Abstract: Fault attacks allow malicious actors to modify the behavior of a program by
physically injecting a fault in the hardware. They typically target sensitive
applications such as cryptography services, authentication or boot-loader and
firmware updater. They can be defended against by adding countermeasures, that
is control flow checks and redundancies, either in the hardware, or in the
software running on it. In particular, software countermeasures may be added
automatically during compilation.

In this talk, we will describe a formally verified implementation of this
approach in the CompCert verified compiler for the C language. We proposed a
toolkit to implement countermeasures as transformations of a middle-end
representation of CompCert, RTL.  We applied this toolkit to two existing
countermeasures that protect the control flow of the program.  We proved that
these countermeasures are correct, that is, they do not change the observable
behavior of the program during an execution without fault injection. We then
modeled the effect of a fault on the behavior of the program as an extension of
the semantic model of RTL. We used this new model to formally prove the efficacy
of the countermeasure: all attacks are caught. In addition to this formal
reasoning, we evaluated the protected program using Lazart, a tool for symbolic
fault injection.

-----------------------------------

This event is part of the FATA Weekly Seminar, which takes place every Tuesday from 3:00 - 4:00 PM in Room 422, Sir Alwyn Williams Building and on Zoom https://uofglasgow.zoom.us/j/83611964233?pwd=CgRyzxK8Z9fP2ULTb5ONWZeUYx2t2E.1

GIST Seminar - Bike to the Future: Cyclists and Automated Vehicles

Group: Human Computer Interaction (GIST)
Speaker: Ammar Jamal Said Al Taie, University of Glasgow
Date: 12 December, 2024
Time: 13:00 - 14:00
Location: Sir Alwyn Williams Building, 423 Seminar Room

Summary:

Cyclists are vulnerable road users who must share the road with motorised vehicles. They encounter vehicles across a range of traffic scenarios, including roundabouts or lane merging. These encounters result in space-sharing conflicts that require some social interaction to be resolved. For example, a driver may wave their hand to signal a cyclist to proceed, or a cyclist may extend their arm to indicate their intention to merge lanes with a driver. Autonomous vehicles will soon be on our roads, and these social interactions will diminish. This would create new ambiguities for cyclists (and other road users) in future traffic. In this talk, I will present work conducted over the past three years to understand the scenarios that trigger AV-cyclist interaction, design potential interfaces that facilitate these interactions, and evaluate these interfaces using novel study designs. The contributions are in the interface designs, and new methodologies to design and evaluate them. This is important to ensure safer, more pleasant future cycling.

Bio:

My area of research is Autonomous Vehicle-Cyclist interaction. This often involves utilising unconventional technologies, such as new displays on the car's exterior. I am a "hands-on" researcher; most of my work is conducted in real-world settings using new technologies such as eye-tracking. In my free time, I enjoy running, cycling, reading graphic novels and playing mario kart!

Yougang Lyu IR Seminar

Group: Information Retrieval (IR)
Speaker: Yougang Lyu, University of Amsterdam
Date: 16 December, 2024
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

TBC

Xiaoyu Zhang IR Seminar

Group: Information Retrieval (IR)
Speaker: Xiaoyu Zhang, Shandong University
Date: 06 January, 2025
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

TBC

Francesco L. De Faveri IR Seminar

Group: Information Retrieval (IR)
Speaker: Francesco L. De Faveri, University of Padua
Date: 13 January, 2025
Time: 15:00 - 16:00
Location: Sir Alwyn Williams Building, 422 Seminar Room

TBC