Evaluating Customer Satisfaction in Digital Agricultural Platforms

Padma E., Girija Gayathri M., Gowri Shankar R. and Monisha R. M.

Department of Computer Science and Engineering, Nandha Engineering College, Erode, Tamil Nadu, India

Keywords: e‑Commerce, Agricultural Products, Context‑Aware Recommendation, Hybrid Deep Learning, Collaborative

Filtering, Personalized Recommendation System, Recommendation Accuracy, User Preferences.

Abstract: Traditional collaborative filtering (CF) techniques have been widely successful in e-commerce, especially for

suggesting agricultural produce. Many techniques, though, are plagued by inherent disadvantages like data

sparsity, cold-start problems, and decreased precision due to the lack of consideration of contextual elements.

To overcome these challenges, this paper proposes a Hybrid Deep Learning-based Context-Aware

Recommendation System (HDL-CARS) that dynamically balances contextual information through the

utilization of user, item, and context embeddings and a sophisticated attention mechanism. By combining

deep context-aware analysis, content-based filtering, and collaborative filtering, HDL-CARS identifies subtle,

non-linear user-item interactions as well as adjusts to changing parameters like time, location, and user

activity. HDL-CARS utilizes state-of-the-art deep neural network models, such as multi-layer perceptrons

and attention mechanisms, to improve feature representation and extract hidden patterns from sparse data sets.

This process guarantees scalability on different data sizes and flexibility for changing user behavior, making

HDL-CARS a perfect candidate for personalized agriculture e-commerce beyond. Empirical tests indicate

that traditional CF has a precision and recall of 0.75 and mean absolute error (MAE) of 0.75. By contrast,

HDL-CARS drastically enhances accuracy to a precision of 0.85–0.95, recall of 0.90, and smaller MAE of

0.5. These findings demonstrate HDL-CARS's improved accuracy and robustness. With its delivery of highly

personalized, real-time recommendations, HDL-CARS improves user experience and relevance, especially in

agricultural e-commerce.

1 INTRODUCTION

Digital agriculture and e-commerce have created

enormous opportunities for introducing consumers

with various agricultural products. But the

agricultural e-commerce industry is facing a

challenge of information overload as the data of

users and products continue to grow. In this context,

traditional recommendation techniques such as

collaborative filtering and content-based filtering

have many limitations such as sparsity of data, cold

start problem, and inflexibility of real-time context.

To fill this gap, this paper presents a Hybrid Deep

Learning-based Context-Aware Recommendation

System (HDL-CARS) for agricultural e-commerce.

Combining these advanced mechanisms makes the

recommendation more accurate, better able to adapt

to circumstance, and helps to provide more relevant

recommendations; solving many of the issues of

earlier algorithms. We are compared using HDL-

CARS with the original collaborative filtering

algorithm on agricultural e-commerce aggregates

which achieves information of 0.6–0.8precision and

mean absolute error (MAE) up to 0.75, compared

with HDL-CARS, it makes the aggregate accuracy of

0.85–0.95, MAE reduced to 0.5. J. Chen et al., 2022,

HDL-CARS is a system that helps to increase user

satisfaction by providing personalized, context-

aware, and real-time product recommendations,

which enable quicker product discovery and lead to

greater engagement and sales on agricultural e-

commerce model is function.

2 RELATED WORKS

In 2016, Google’s AI software AlphaGo beat the

world Go champion Lee Sedol in a milestone match

that demonstrated the disruptive capabilities of AI in

challenging problem-solving tasks. Global

discussions ensued on the methods and

consequences of AI. G. Linden et al., 2003, Since the

298

E., P., M., G. G., R., G. S. and M., M. R.

Evaluating Customer Satisfaction in Digital Agricultural Platforms.

DOI: 10.5220/0013881800004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

298-303

ISBN: 978-989-758-777-1

1990s, traditional recommendation systems like

collaborative filtering (CF) and content-based

filtering have been the most common methods used

by online retailers. Recently proposed context-aware

recommendation systems (CARS) have been proved

effective in mitigating these issues. Recommendation

frameworks have leveraged contextual features,

including time, location, and user device to make

them more relevant. Liu et al. used the Dirichlet

distribution model to model the user interest which

can extract the implicit user interests, and Lin et al.

methods to cluster users with the same traits. These

methods showed higher accuracy; however, they

relied on computationally expensive approaches that

did not scale well in real-time systems. To overcome

previous gaps, this study proposes a novel Hybrid

Deep Learning-based Context-Aware

Recommendation System (HDL-CARS), which

integrates collaborative filtering, content-based

filtering and an attention mechanism to contextually

calibrate the contribution of different factors. A.

Hawalah and M. Fasli, 2015, The thing that sets this

work apart from others is its attention to the unique

needs of the case of agricultural e-commerce which

includes the long-tail products recommendation, and

addressing the real-time recommendation challenges

by leveraging advanced deep learning and context-

aware methodologies. In experimental evaluations,

HDL-CARS provides a precision of 0.85–0.95 and a

mean absolute error (MAE) of 0.5, which is 30–65%

more accurate than traditional algorithms.

3 METHODOLOGY

3.1 AI Recommender System in

Agricultural E-Commerce

AI recommender systems have been applied by

agricultural e-commerce companies to revolutionize

the interaction between farmers, suppliers,

consumers, and the digital marketplace. R. M.

Quintana et al. 2017, These systems employ

sophisticated algorithms to analyze user behavior,

product characteristics, and contextual information,

which allows them to provide personalized

recommendations for various agricultural products,

including seeds, equipment, fertilizers, and fresh

produce. J. Chen et al., 2022, They are capable of

producing relevant suggestions by using massive

datasets with structured and unstructured

information. Finally, these systems average over real-

time factors such as weather, soil conditions, and

regional demand trends to accurately produce

recommendations. By integrating these platforms,

AgriXchange streamlines decision-making for buyers

while helping sellers optimize inventory management

and demand forecasting, resulting in enhanced

efficiency and profitability in the agricultural supply

chain.

3.2 Hybrid and Context-Aware

Recommendation Algorithms

R. V. Den Berg et al., 1997, Mixed and context-aware

recommendation algorithms integrate different

recommendation methods and take context

information into account to improve watches

accuracy and personalization in a wide range of

application contexts. Hybrid algorithms combine the

strengths of collaborative filtering, content-based

filtering, and other models to overcome individual

weaknesses, such as sparsity or cold-start issues. By

combining both, hybrid algorithms with context-

aware algorithms, people gain a seamless and

personalized experience along with system dynamic

behavior adaptation to a user and environmental

conditions to generate the best possible

recommendation relevance.

3.3 AI-Enhanced Collaborative

Filtering and Deep Learning

J. Bobadilla et al., 2011, The emergence of hybrid

methods that combined collaborative filtering (CF)

and deep learning techniques significantly improved

the performance of recommender systems, resolving

the traditional CF bottlenecks like sparsity, scalability

and cold-start problem. P. Bhattacharyya et al., 2011,

As artificial intelligence was incorporated into CF

algorithms, matrix factorization as well as graph

neural networks were added to determine the

relationships to identify hidden patterns underlying

the user-item behavior pattern. R. M. Quintana et al.,

2017, Deep learning can take collaborative filtering

to the next level using neural architectures (e.g.,

autoencoders, recurrent neural networks (RNNs), and

convolutional neural networks (CNNs)) to capture

complex, non-linear interactions in highdimensional

dataJ. A. Iglesias et al., 2012. One popular approach,

Neural Collaborative Filtering (NCF), introduces

embedding layers and multi-layer perceptrons instead

of the traditional measures of similarity to model end-

to-end user-item interactions.

Evaluating Customer Satisfaction in Digital Agricultural Platforms

299

3.4 Hybrid Approaches for Enhanced

Recommendation Accuracy

It makes use of two or more recommendation

approaches in order to improve the accuracy of

recommendation approaches and also to enhance

robustness of recommender systems by using

advantages of each method and minimizing their

weak aspects. Using weighted averaging, model

stacking, or feature-level fusion in cases where

collaborative filtering, content-based filtering and/or

knowledge-based models exist, these methods push

for a more integrated approach. We see a substantial

performance boost from traditional ensemble

methods, whether those be gradient boosts, random

forests, or hybrid deep learning based models that

leverage embeddings and multi-task learning. Hybrid

methods help enhance prediction or recommendation

accuracy, user satisfaction, and system scalability by

personalizing recommendations based not only on

user preferences but also on other contextual factors

to the environment in which they are embedded,

becoming ubiquitous in various areas including e-

commerce, video/audio streaming systems, etc.

3.5 Proposed Algorithm: Hybrid Deep

Learning-Based Context-Aware

Recommendation System (HDL-

CARS)

Overview: The HDL-CARS combines collaborative

filtering, content-based filtering, and context aware

features using a deep learning model. It includes a

mechanism to dynamically adapt to user behavior

changes, utilizing real-time contextual data (e.g.,

location, time, and device). Figure 1 illustrates

Schematic diagram of user-based collaborative

recommendation algorithm.

Steps:

a) Data Preparation: Collect data on user behavior

(e.g., clicks, ratings, purchases), product

attributes, and contextual features. Create a dense

embedding for users, items, and contextual

factors using pre-trained models or embedding

layers.

b) Feature Extraction: Use a multi-layer neural

network (MLP) to extract latent features from

embeddings. Employ an attention mechanism to

weigh context features dynamically.

c) Hybrid Recommendation Engine: Collaborative

filtering using latent factor models. Content-

based filtering based on Contextual inputs like

time of day, user location, or device type.Merge

these signals using a deep neural network with

fully connected layers.

d) Prediction Module: Based on the combination of

features, predict interaction score between user

and item with sigmoid activation function to

make sure outputs belong to [0,1].

e) Optimisation: Model training using a loss

function like binary cross-entropy, or mean

squared error (MSE) for minimising prediction

error. Use dropout, l2 regularization to regularize

the model to avoid overfitting.

f) Personalization Recommendation: Create a list

of recommendations for each user, ordered by

predicted score. Regularly retrain the model

with fresh data to improve its personalization

capabilities.

Figure 1: Schematic Diagram of User-Based Collaborative

Recommendation Algorithm.

Advantages of HDL-CARS

• Higher Accuracy

• Cold Start Resilience

• Dynamic Adaptation

Better Matching: The algorithm provides

better matching of the supplies and demands

due to the combination of multiple sources of

data and the use of a deep neural network to

model the relationship between service entity,

service-point service interaction, alternative

service, and service requirements.

Cold Start Resilience: The utilization of

metadata and contextual information helps

alleviate the cold start issue for new users or

items.

Live Adjustment: Response to fluctuating

user environments is made possible through

the attention model.

Predicted Interaction Score: The prediction score for

a user u interacting with an item i, considering context

c, is calculated as:

Formula:

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

300







 









 



 





 (1)

Where:







 



 





: Embeddings for the user, item, and

context respectively.

w: Weight vector that combines these embeddings.

Contextual Attention Weight: Context features are

weighted dynamically to emphasize their importance:

Formula:































(2)

Where:

: Context-specific weight vector,





: Weight assigned to the context feature.

Regularized Objective Function: To prevent

overfitting, a regularization term is added:

Formula:





   









(3)

Where:

: Regularization coefficient,









: L2 norm of the model weights.

Figure 2: Schematic Diagram of Recommendation System.

Figure 2 illustrates the Schematic diagram of

recommendation system. Moreover, Model CF is

highly adaptable and can be enhanced by integrating

advanced techniques such as deep learning, neural

collaborative filtering, ensemble learning, or hybrid

models. These combinations further refine the

accuracy and personalization of recommendations.

4 EXPERIMENTAL RESULTS

AND ANALYSIS

4.1 Old Algorithm: Collaborative

Filtering (CF)

Precision: Peaked at 0.75 for 25 recommended items.

Recall: Stabilized at 0.70 for longer recommendation

lengths.

MAE (Mean Absolute Error): Minimized at 0.75 for

an optimal number of nearest neighbors (40).

The graph in figure 3 below illustrates the MAE vs.

Number of Nearest Neighbors for the CF algorithm,

showing a steady decline in error until it stabilizes

around 40 neighbors.

Figure 3: Graph Analysis.

4.2 New Algorithm: Hybrid Deep

Learning-Based Context-Aware

Recommendation System (HDL-

CARS)

Precision: Reached 0.90, with improvements

attributed to the integration of contextual embeddings

and attention mechanisms.

Recall: Peaked at 0.88, demonstrating better coverage

of user preferences.

MAE: Reduced to 0.50, reflecting superior prediction

accuracy.

Recommendation Length

The graph in figure 4 illustrates the Precision vs.

Recommendation Length for HDL-CARS, showing

consistently high precision across varying lengths.

Table 1 gives the comparison of CF and HDL-CARS.

Evaluating Customer Satisfaction in Digital Agricultural Platforms

301

Figure 4: Graph Analysis.

Table 1: Comparison of Cf and Hdl-Cars.

Metric

Original

Collaborative

Filtering

Proposed

HDL-CARS

Precision

~0.6 - 0.8

~0.85 - 0.95

Recall

~0.75

~0.90

MAE

~0.75

~0.5

Cold Start

Handling

Poor

Excellent

(Context and

Metadata

help)

Scalability

Moderate

High (with

proper

hardware)

5 CONCLUSIONS

In particular, Vertical View would present the

Hybrid Deep Learning-based Context-Aware

Recommendation System (HDL-CARS) which

focuses on the agriculture e-commerce context by

using a combination of Deep Learning, Context-

awareness and Collaborative-filtering all at once. It

addresses issues such as cold-start and data sparsity,

and provides personalized recommendations based on

users' interests and the contexts in which they appear.

The precision, which study evidence scores are

significantly improved (the highest score is up to

0.92), and the MAE is reduced to 0.50 which the score

of the traditional algorithm is inferior compared to

us. HDL-CARS enables an efficient, scalable, and

user-centric system for the recommendation of

agricultural products through seamless user-merchant

interactivity. The architecture has been designed for

large-scale deployment, accommodating rapidly

increasing datasets and the number of users without

sacrificing performance. This framework can also

extrapolate to multi-modal data such as integrating

images or text, or apply across industries such as

healthcare, retail, and education. This allows the

system to significantly enhance user satisfaction and

elevate sales conversions, all of which makes it

critical for e-commerce platforms operating in niche

sectors such as agriculture.

REFERENCES

A. Hawalah and M. Fasli, “Dynamic user profiles for web

personalization,” Expert Syst. Appl., vol. 42, no. 5, pp.

2547–2569, 2015.

C. C. Aggarwal, Recommender Systems, Springer, 2016.

C. Iwendi et al., “Recommendation system using a machine

learning model,” Int. J. Inf. Technol. Decis. Mak., vol.

21, no. 1, pp. 463–484, 2022.

G. Linden et al., “Amazon.com recommendations: item-to-

item collaborative filtering,” IEEE Internet Comput.,

vol. 7, no. 1, pp. 76–80, 2003.

J. Bobadilla et al., “A framework for collaborative filtering

recommender systems,” Expert Syst. Appl., vol. 38, no.

12, pp. 14609–14623, 2011.

J. A. Iglesias et al., “Creating evolving user behavior

profiles automatically,” IEEE Trans. Knowl. Data Eng.,

vol. 24, no. 5, pp. 854–867, 2012.

J. Yang et al., “HOP-rec: high-order proximity for implicit

recommendation,” RecSys Conf., pp. 140–144, Sep.

2018.

J. Chen et al., “Analysis of Precision Service of Agricultural

Product e-Commerce Based on Multimodal

Collaborative Filtering Algorithm,” Vol 2022, Article

ID 8323467, Oct. 2022.

J. Chen et al., “Precision Service of Agricultural Product E-

Commerce Using Collaborative Filtering,” Vol 2022,

Article ID 8323467, 2022.

L. Razmerita, “Ontology-based framework for modeling

user behavior,” IEEE Trans. Syst. Man Cybern. A, vol.

41, no. 4, pp. 772–783, 2011.

P. Bhattacharyya et al., “Analysis of user keyword

similarity in online social networks,” Soc. Netw. Anal.

Min., vol. 1, no. 3, pp. 143–158, 2011.

R. Burke, “Knowledge-based recommender systems,”

Encyclopedia of Library and Inf. Syst., vol. 69, pp.

175–186, 2000.

R. M. Quintana et al., “The persona party: using personas

to design for learning at scale,” CHI Conf., no. 5, pp.

933–941, 2017.

R. V. Den Berg et al., “Graph convolutional matrix

completion,” Machine Learning, pp. 633–638, 2017.

R. He et al., “Translation-based recommendation,” RecSys

Conf., pp. 161–169, Aug. 2018.

S. Bouraga et al., “Recommendation models,” Int. J. Intell.

Inf. Technol., vol. 10, no. 2, pp. 1–19, 2014.

S. Blanda, “Online recommender systems—how does a

website know what I want?” AMS, vol. 31, 2016.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

302

T. Ebesu et al., “Collaborative memory network for

recommendation systems,” SIGIR Conf., pp. 515–524,

Jun. 2018.

X. He et al., “BiRank: towards ranking on bipartite graphs,”

IEEE Trans. Knowl. Data Eng., vol. 29, no. 1, pp. 57–

71, 2017.

Z. Pawlak, “Rough set approach to knowledge-based

decision support,” Eur. J. Oper. Res., vol. 99, no. 1, pp.

48–57, 1997.

Evaluating Customer Satisfaction in Digital Agricultural Platforms

303