LLMs have revolutionized the understanding and generation of natural language, ofering new possibilities for enhancing recommendation systems. In previous studies, LLMs exploit their global knowledge to provide zero- or few-shot recommendations. In this work, we aim to highlight the opportunities that LLMs pose to enrich the field of recommender systems combined with local knowledge. We propose to view recommender systems combined with LLMs from a broader perspective, recognizing them not merely as another method to replace existing recommendation approaches, but rather as a complementary and powerful approach to enhance and augment the overall recommendation process.
1. Introduction
like Generative Pre-trained Transformer (GPT)1, LLaMA
[
Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop Proceedings (CEUR-WS.org)
https://openai.com/research/gpt-4
ploiting the language understanding and generation
capabilities to provide explainable recommendations [
LLMs that difer from approaches that only rely on the
global knowledge of pre-trained models and directly
recommend items based on that. This is often not
applicable in a real-world scenario, where you have a fixed
set of items to choose from and want to ensure factual
recommendations. Moreover, our goal is to provide
textual metadata about a dedicated set of items, which we
call local knowledge, for our recommendations, but use
the global knowledge of LLMs to support other steps
of the pipeline of a system. To further investigate this,
we conducted a pilot study involving students of a
Masmethod can furthermore overcome problems when the
global knowledge does not hold any information about
an item in one’s own stock. Figure 1 highlights these
diferent components and emphasizes that modern
recommender systems consist of a whole system
architecture of which the recommendation engine is only one
employed exclusively by the language model integration
framework or the recommendation engine, independent
of any access to the global knowledge of the LLM; such
knowledge can be stored through diverse methods (e.g.
vector stores, text files, etc.). This local knowledge is used
as the basis to retrieve items for recommendations based
on e.g. a similarity search. LLMs in general additionally
often face the issue of hallucination [
erating summaries or topics holds benefits both in terms We anticipated a lot of potential in using LLMs in vari- of eficiency and efectivity. In a large local data base ous stages of recommender systems to support finding of news articles, processing quickly becomes expensive, suitable recommendations from local knowledge. This when assessing the similarity between a user profile’s section highlights the potentials that should be further content or a natural language user query and the fullexplored in our pilot study and reflect on challenges that text articles on every invocation. This process can be come with them. alternatively executed on a concise summary, containing significantly fewer tokens while incurring only minimal 2.1. Input Enhancement and information loss. Generating topics holds the benefit of
Restructuring providing further information that can be exploited for
When creating recommendations for a user, we can pro- recommendations that better fit the user’s needs. Topics
cess diferent types of signals as input for a system [
Challenges arise when the input structure difers from 2.3. Further Use Cases
the meta-information about target items. An example We want to mention additional use cases that tackle open
setting would be recommending job listings based on problems in the recommender systems field or can be
resumes. Job listings can follow a structure with dif- combined with traditional approaches for future
endeavferent fields that hold information about tasks etc. In ours. When using LLMs to hold conversations with users,
such cases, LLMs ofer promising potential to restructure we can use the capabilities of LLMs to extract the intent
resumes and create synthetic job listings or vice versa. of users and build profiles dynamically where the system
Additionally, LLMs can help imputing missing values generates questions in real time that aim at acquiring
spewithin structured content, such as generating a short cific missing information or infers it from a conversation
pitch or personal description from working experiences lfow. The conversational setting furthermore enables the
and interests in resumes. Experimental investigations in incorporation of real time feedback and iterative
adapthis direction can include prompt engineering to align tion and refinement of recommendations. This process
the enhancement or restructuring process across multi- involves presenting recommendations based on prompts
ple user sessions, given the non-deterministic nature of or the chat history and allowing users to provide ratings
generated responses. or rankings for the suggested items. The system
lever2.2. Content Extraction and Sample ages this feedback as additional information about user
Prompt Generation preferences, subsequently enhancing the precision of
following recommendations, resulting in more personalised
Another notable application revolves around the gener- and contextually relevant recommendations.
ation of supplementary content based on target items, A major challenge for all mentioned scenarios is that
which serves various purposes. An example is the usage one has to engineer a set of prompts or prompt templates
of LLMs to extract topics or summaries from target items. that make sure that the model executes the concerning
In the context of news, LLMs have shown to be on level task as intended and delivers responses consistently in
with human-created summaries [
New insights into query enhancement and restruc- understands its users better. Task (2) was very similar, turing were gained through four diferent projects: (1) but this time the goal was to identify reading motives. matching a set of resumes with a dataset of job listings, (2) The proposed solution was a combination of embedding creating a system, which is able to recommend companies all articles with the help of Ada8 and extracting reading based on certain preferences (i.e. to understand compet- motives from articles with the help of GPT to match them itive suppliers in the marketplace), (3) recommending with user preferences. clothes to wear on a particular occasion, and (4) building 4. Conclusion a recommender system to recommend books based on a given book. Project (1) utilized LLMs to generate struc- In conclusion, our work highlights the abundant potentured JSON Files from resumes and to complete null fields tial of leveraging local knowledge in conjunction with based on other content. They used a creative approach LLMs to advance recommender systems. We propose the to generate job listings from resumes to match them with need for a comprehensive exploration of this potential, the jobs within the job listing. In addition query exten- transcending the early trends of using LLMs to replace sion was utilized to enhance or complement the resumes by traditional recommenders or focusing on explainable recaccessing the global knowledge of the LLM. In task (2) ommendations. Recognising modern recommender systhe assumption was that a user would use a rather sim- tems as multifaceted ecosystems, where the recommenplistic prompt to search for competing companies. They dation algorithm represents only one component, we enhanced this query by prompting GPT for an enhanced propose investigating mentioned aspects such as input version of the given query. The approach was to provide enhancement and restructuring, content extraction, and GPT with the original prompt, including a keyword list sample prompt generation in future research endeavors. of the user preferences based on the behavior within the By doing so, we can benefit from the ever-evolving capasession. Subsequently, the user is provided with multi- bilities of LLMs, while also mitigating their drawbacks ple recommendations for generated search prompts from through the incorporation of trusted local knowledge. which they can choose the one that best reflects their This approach promises to uncover new advancements intent. The selected prompt is then used for the further in personalised and knowledge-aware recommendations, recommendation process. Project (3) solved the given enriching the overall user experience and amplifying the problem by providing the user with a chat interface. The benefits of using LLMs in recommender systems. user can chat with the system and submit their prefer
humanism-initiative-on-chatgpt/ 5The pilot study was designed as an optional track within the lecture to comply with the responsible research practices at TU Wien. 6https://openai.com/blog/openai-api 7https://python.langchain.com/