SOAR: Social and Opinion-Aware Recommendations

Giannis Christoforidis; Apostolos N. Papadopoulos

cs.LGOct 21, 2026

SOAR: Social and Opinion-Aware Recommendations

Authors: Giannis Christoforidis, Apostolos N. Papadopoulos

Organizations: School of Informatics, Aristotle University of Thessaloniki, AUTH Campus, Thessaloniki, 54124, Greece.

Abstract

One of the enduring challenges in online platforms is recommendation. Both communities and advertisers face difficulties in analyzing complex user behavior, identifying habits and evolving preferences, and providing timely recommendations that are relevant to user interests. In particular, point of interest recommendation poses the additional challenge of identifying the factors that influence user behavior, capturing user tendencies, and generating accurate, personalized recommendations for each individual. In this work, we propose SOAR, a Social and Opinion Aware Recommendation model that incorporates both the user's friendship network and textual reviews to assess whether a location is suitable for a specific user, with a particular emphasis on the sentiment of shared reviews. Experimental results in real-world datasets demonstrate that incorporating reviews into the recommendation process improves the precision of the recommendation.

Figures & tables

Symbol	Description
$\mathcal{U},\mathcal{L},\mathcal{R},\mathcal{C}$	Set of Users, $\mathcal{U}=\{u_{1},...,u_{n}\}$ ,
	set of locations $\mathcal{L}=\{l_{1},...,l_{m}\}$ ,
	set of Reviews $\mathcal{R}=\{r_{1},...,r_{k}\}$ ,
	set of Check-ins $\mathcal{C}=\{c_{1},...,c_{o}\}$
$\mathcal{F}_{u_{i}}$	set of Users’ Friends $\mathcal{F}=\{f{u_{1}},...,f{u_{n}}\}$ ,
$c_{u_{i}}$	user’s check-in

Table 1: Frequently Used Symbols

Figure 2: Bert example.

Figure 3: Similarity Score Computation.

Figure 4: Example of Review Connection.

Figure 5: Example of Review Connection in Vector Space.

Figure 6: Example of Friend’s influence.

Statistic	Value
Total Number of Reviews	6,990,280
Total Number of Words	732,415,837
Total Number of Locations	150,346
Average Number of Words	104.78
Median Review Length (words)	75
Std Deviation of Review Length (words)	97.92

Table 2: Yelp Dataset statistical data.

Radius	Avg	Max	>200
0.500	1.83	153	0
0.525	2.94	330	2
0.550	5.58	630	5
0.575	11.30	1166	12
0.600	23.91	2248	26
0.625	51.03	4336	49

Table 3: Review statistics for different radius values

Top-K	Recall (%)			F1-score (%)			Accuracy (%)
Top-K	Main	2-hop	1km	Main	2-hop	1km	Main	2-hop	1km
Top-1	16.37	16.25	7.98	27.17	27.00	14.52	34.22	34.13	29.58
Top-5	42.22	41.68	23.56	54.72	54.34	36.32	47.66	47.53	38.11
Top-10	56.59	55.62	34.25	65.08	64.61	47.68	54.54	54.39	43.72
Top-25	73.66	72.45	48.67	74.53	74.18	59.72	62.29	62.21	50.81
Top-50	82.76	81.83	56.77	79.05	78.78	65.37	66.89	66.73	54.63

Table 4: Comparison of Recall, F1-score, and Accuracy among versions of our model 2-hop, and 1km across different top-K categories.

Model	Speed (predictions/s)	Total Execution Time
SOAR 2 HOP	7.32	36 hours
SOAR	55.56	4.6 hours

Table 5: Performance comparison between friends and friends of friends

Top-K	Recall (%)		F1 (%)		Top-K Hits (%)
Top-K	SOAR	baseline	SOAR	baseline	SOAR	baseline
Top-1	16.37	13.37	27.17	22.85	15.38	12.77
Top-5	42.22	34.44	54.72	47.56	40.68	33.57
Top-10	56.59	47.01	65.08	58.06	55.33	46.37
Top-25	73.66	63.60	74.53	68.84	73.16	63.51
Top-50	82.76	72.79	79.05	74.01	82.54	72.96

Table 6: Comparison of SOAR and baseline on Recall, F1, and top-K Hits

Figure 12: Accuracy comparison between models

Figure 13: Recall comparison between models

Figure 14: F1-score comparison between models

Figure 15: Precision comparison between models

Figure 16: Top-K Hits comparison between models

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 516 papers

ReFORM: Review-aggregated Profile Generation via LLM with Multi-Factor Attention for Restaurant Recommendation
Mar 17, 2026Moonsoo Park, Seulbeen Je, Donghyeon ParkRecommender Systems Large Language Model-Based Recommendation
University of Southern California, USA · Sejong University, South Korea
SPADE: Escaping the Popularity-Similarity Frontier to Measure Serendipitous Recommendations
Sep 25, 2026Tobias Vente, Maarten Peirsman, Noah Daniëls +2Recommender Systems
ADREM Data Lab, University of Antwerp, Middelheimlaan 1, 2020 Antwerpen, Belgium · Department of Computer Science, PO Box 68, FI-00014 University of Helsinki, Finland · Froomle, Post X, Borsbeeksebrug 34, 2600 Antwerpen, Belgium
Scaling Articulated Rationales for MLLM-based Recommendation
Sep 15, 2026Haoke Xiao, Yueyang Liu, Yuhui Zhang +17Learning to Rank Recommender Systems
Kuaishou Technology

Mar 17, 2026J/K move · Enter open · S save

ReFORM: Review-aggregated Profile Generation via LLM with Multi-Factor Attention for Restaurant Recommendation

Moonsoo Park, Seulbeen Je, Donghyeon Park

University of Southern California, USA · Sejong University, South Korea

Recommender Systems Large Language Model-Based Recommendation User Preference Modeling Attention Mechanisms

Open paper AskPDF arXiv

In recommender systems, large language models (LLMs) have gained popularity for generating descriptive summarization to improve recommendation robustness, along with Graph Convolution Networks. However, existing LLM-enhanced recommendation studies mainly rely on the internal knowledge of LLMs about item titles while neglecting the importance of various factors influencing users' decisions. Although information reflecting various decision factors of each user is abundant in reviews, few studies have actively exploited such insights for recommendation. To address these limitations, we propose a ReFORM: Review-aggregated Profile Generation via LLM with Multi-FactOr Attentive RecoMmendation framework. Specifically, we first generate factor-specific user and item profiles from reviews using LLM to capture a user's preference by items and an item's evaluation by users. Then, we propose a Multi-Factor Attention to highlight the most influential factors in each user's decision-making process. In this paper, we conduct experiments on two restaurant datasets of varying scales, demonstrating its robustness and superior performance over state-of-the-art baselines. Furthermore, in-depth analyses validate the effectiveness of the proposed modules and provide insights into the sources of personalization. Our source code and datasets are available at https://github.com/m0onsoo/ReFORM.

Abstract

Recommender Systems

Explore similar work

ReFORM: Review-aggregated Profile Generation via LLM with Multi-Factor Attention for Restaurant Recommendation

SPADE: Escaping the Popularity-Similarity Frontier to Measure Serendipitous Recommendations

Scaling Articulated Rationales for MLLM-based Recommendation