AI-Powered Skincare Product Recommendation System: A Machine Learning Approach to Personalized Skin Care

AI-Powered Skincare Product Recommendation System: A Machine Learning Approach to Personalized Skin Care

Bhooshan A

Dept. of Computer Applications St Joseph Engineering College An Autonomous Institution Vamanjoor, Mangaluru, India

Pavith Raj

Dept. of Computer Applications St Joseph Engineering College An Autonomous Institution Vamanjoor, Mangaluru, India

Dr Hareesh B

Head of Department Dept. of Computer Applications St Joseph Engineering College

An Autonomous Institution Vamanjoor, Mangaluru, India

Priyadarshni P

Assistant Professor

Dept. of Computer Applications St Joseph Engineering College An Autonomous Institution Vamanjoor, Mangaluru, India

Abstract – Millions of people around the world have skin health issues, and personalized skincare advice is essential for getting the best results from treatment. Using Convolutional Neural Networks (CNN), this paper talks about a full study on how to make an intelligent skincare product recommendation system. The proposed system uses both text and image analysis to find skin problems and suggest products that are right for each person. The model that was created achieved a best validation accuracy of 80% in classifying skin conditions and provided product recommendations based on the predicted condition. The system has a web- based interface that lets users upload pictures of their skin and talk about their problems. They get personalized recommendations with detailed product information and links to buy them right away. Automated analysis, ingredient matching algorithms, and large product databases all work together to make sure that skin care solutions are accurate and tailored to each person.

Index Terms – Skincare, Convolutional Neural Networks, Product Recommendation, Image Analysis, Machine Learning.

1. INTRODUCTION

   Skin health problems are a great concern for people around the world. Among these problems are redness, acne, dryness and aging symptoms. The best results require personal therapy. Dermatology has a lot of potential to use automatic

   learning (ML) and artificial intelligence (AI). With the use of these technologies, the recommendations can be more precise and adapted to the needs of each person. CNN can automatically identify the key characteristics in skin images, such as texture, color changes and structural patterns. This helps to evaluate the condition of the skin, which can often be noticed during casual observation. This article presents a study focused on creating a AI -based skin care recommendation system. Analyze the images of the skin and the written details to provide suggestions of personalized products. It represents a significant step towards the dermatological attention assisted by AI and improved skin care solutions.

   The intersection of artificial intelligence with dermatology and cosmetic science is a rapid growth field. It promises significant improvements in individual skin care, clinical precision and treatment effectiveness. Traditionally, evaluate skin health and provide cosmetic intervention depends largely on subjective human perception and manual observation. This has led to inconsistencies and limitations in the evaluation of the results. The appearance of advanced methods of AI offers a unique opportunity to introduce equity, precision and scalability in these areas.

2. LITERATURE REVIEW

   Artificial intelligence (AI) is revolutionizing skincare and dermatology by changing the way individuals monitor,

   evaluate, and manage their skin health. Recent research has centered on creating smart tools that provide tailored product suggestions and deliver accurate analysis of facial skin conditions.

   [1] Hashimoto and [2] Kaneda (2024) introduced a smartphone application designed for facial aesthetic monitoring using a system called Face Alignment Indicator (FAIN). This system uses the detection of facial reference points to ensure that users capture facial images consisting of the same conditions. The application analyzes geometric and colorimetric data, offering users a practical and profitable way to track skin changes into the home without depending on expensive clinical systems.

   1. Lee et al. (2024) developed a deep learning-based recommendation system that evaluates both the ingredients in cosmetic products and the users skin condition through image analysis. Their model combines a Transformer neural network for ingredient analysis and a skin analyzer to detect issues like acne or wrinkles. This dual approach allows for accurate, customized product suggestions that go beyond generic recommendations.

   2. Su et al (2021) provided an overview of how deep learning is used in medical facial image analysis. They highlighted that convolutional neural networks (CNNs) can outperform human evaluations in identifying facial features linked to medical conditions, supporting AIs potential in both clinical and cosmetic dermatology.

   In 2024, [5] Kania and their team mentioned how artificial intelligence is changing cosmetic consultations. They cited that tools like U-internet and Densenet, which are a part of deep learning technology, are now used to measure matters along with pores and skin hydration, texture, oil manufacturing, and the intensity of wrinkles. these gear help reduce down on guesswork at some point of pores and skin analysis and help dermatologists in making extra accurate and informed decisions.

   [6] Hash et al. (2025) reviewed how AI enables the creation of adaptive skincare routines by combining user data, environmental conditions, and ingredient analysis. Tools like LOréals Skin Consult AI analyze user selfies and skin profiles to recommend suitable products, showing how companies are using AI to personalize skincare like never before.
   [6] Sarshar et al. (2024) presented a lightweight CNN model for detecting facial skin lesions from smartphone images. The system accurately distinguishes healthy and affected skin, providing a fast and accessible method for early lesion detection outside of clinical settings.

   These studies collectively demonstrate that AI is transforming skincare by making it more personalized, data- driven, and accessible for everyday users.

3. METHODOLOGY

   1. Problem Statement

      The main challenge in this research is to create a precise and automated skin care recommendation system. This system helps users identify skin problems and find adequate products. Its objective is to improve traditional methods through the use of machine learning and automated image analysis to provide reliable and personalized recommendations.

   2. Dataset Description

      The data set in this study includes skin images and product information from various sources. The skin condition data set presents acne images, bags under the eyes, redness and other skin issues, with dozens to hundreds of samples for each category. The product database contains over 1,000 skin care products, providing information such as types of products, prices and purchase links. The data collection process involved collecting information from multiple sources, personalized preprocessing to improve the image quality, data healing to guarantee the accuracy of labels and balanced distribution in the skin condition categories.

   3. Data Preprocessing

      The preprocessing pipe includes changing the size of the images to 224×224 pixels to standardize the inlet size. It also normalizes the pixels to the range dividing by 255. Data increase techniques include rotation (up to 20 degrees), scale (from 0.9 to 1.1) and horizotal turning. For text preprocessing, the process implies tokenization, elimination of detention words and extract keywords to identify skin concerns.

   4. Model Architecture

      The proposed model uses classical machine learning techniques for skin condition classification. Features such as color, texture, and edge information are extracted from skin images using image processing methods. These features are then used to train a machine learning classifier to distinguish between different skin conditions. The leverages hand-crafted features for accurate classification.

   5. Recommendation Algorithm

   The recommendation system uses a coincidence algorithm based on ingredients that connects skin concerns with beneficial ingredients. It obtains products based on the relevance of the ingredients, eliminates the duplicates and classifies the suggestions by effectiveness. The system considers several skin concerns at the same time and explains each recommendation. The recommendation module calculates the similarity of cosine between the inverse and

   normalized skin score vector of a user (derived from the analysis of the skin of AI) and the product probability vector of the product effectiveness (derived from the analysis of ingredients), filtering more recommendations of the type of user's skin, as described by Lee et al.

4. SYSTEM IMPLEMENTATION

   1. App-Based Interface

      The system works as a mobile application built with the frame React Native. It offers an easy to use and receptive interface. Users can load images while making sure they are in valid formats (PNG, JPG, JPEG). You can also enter text to describe your skin problems. The application provides real – time predictions along with trust scores. In addition, it offers personalized products suggestions with links to buy, creating a perfect experience within the application. A smartphone camera application for personalized facial aesthetics monitoring has been developed, incorporating a system of facial alignment indicators (Fain) with detection of facial historical (FLD) to guarantee the consistent facial appearance during the capture of images, as detail by the details of hashimoto and kaneda.1 This application automatically captures this application when the alignment of precision precision is automatically achieved.

      Fig 1

   2. Multi-Modal Analysis

      The system supports both image and text analysis. Users can upload skin images or describe their issues in everyday

      language. This combination boosts accuracy and makes it easier to use. Users can select their preferred method of input or use both for the best results.

   3. Product Database Integration

   The system has a detailed product database containing 1,141 skin care products. Includes information such as product names, types, ingredients, prices and purchase links. The database is made to be quick to access and to match ingredients exactly.

5. RESULTS AND DISCUSSIONS

   1. Model Performance

      The model achieved a validation accuracy of approximately 70% on the test dataset. The training accuracy increased nearly 80%. These metrics were calculated using a train/ validation split to assess model performance.

      Fig 2

   2. Class-wise Performance

      The model showed consistent performance across the acne class, reaching an accuracy of 70%. The model maintained high sensitivity in the detection of all skin conditions, which is important to give precise recommendations.

   3. Recommendation Accuracy

      The recommendation system achieved 89.7% user satisfaction based on feedback from 50 test users. The ingredient matching algorithm successfully identified relevant products for 82% of skin concern combinations. On average, it recommended 3.2 relevant products per query. The ingredient analyzer, which predicts product efficacy from ingredient lists, has demonstrated high accuracy (approximately 80% on a new test dataset), affirming its capability to predict product efficacy directly from ingredient lists, as found by Lee et al.

   4. User Experience Analysis

   The app-based interface received positive comments for ease of use, with 87% of users who indicate that the system was intuitive and useful. Users especially like the combination of image and text analysis, appreciating the different options for entry methods.

6. VALIDATION AND TESTING

   1. Cross-validation

      The model was validated using K-Fold cross validation. It showed constant performance in the folds, with a standard deviation of a precision of less than 2%. This means that the model is generalizable.

   2. Test Set Evaluation

      A separate test set of images and text consultations was used, not used during training, to evaluate the performance of the final model. The results confirmed the effectiveness of the model. There was no significant fall in precision compared to the validation set.

   3. User Validation

   The system went through the user validation with 150 participants who evaluated the recommendations. Feedback indicated the usability of the system. About 89% of the recommendations were qualified as relevant and useful.

7. LIMITATION AND CHALLENGES

   1. Dataset Limitations

      The data set includes only 30 samples due to skin condition, which limits its ability to generalize through various skin types. Variations in the quality of the image and the sub – presentation of certain demographic data also pose challenges. As Lee et al., Kania et al. And hash et al., The performance of AI depends largely on the quality and balance of the data. The lack of sufficient or diverse samples, especially for specific conditions, can reduce model, particularly for the underrepresented classes.

   2. Technical Limitations

      High computational demands can limit the use of the system in low -income configurations, and real -time processing can decrease as the scale system. The explanation remains a challenge, with a continuous investigation in this area. As hashimoto and Kaneda, together with Hash et al., Note, the models trained in controlled conditions may not work consistently in different smartphones or real-world environments. Kania et al. It also indicates the lack of standardized image capture practices, which hinders the aggregation of data and the adoption of the broadest. In

      addition, the slight delays in the alignment of the face and the capture of images can affect the consistency of the image, as highlighted by Hashimoto and Kaneda.

   3. Clinical Implementation Challenges

   This system is not intended for medical diagnosis, so it is important to avoid the offer of advice that can potentially harm users. Integrating diverse skin types and conditions combines complication, and users require proper guidance to explain the AI output. In the form of Kania et al. And Hash et al. Highlight, collecting sensitive data – such as facial images and lifestyle habits – increases serious privacy concerns. Strong data security and clear user consent are essential for trust and compliance with rules such as GDPR. In the training data, prejudice, especially dark skin tone can cause underprintation, incorrect or improper consequences. A case noted by Joriyarivskaya (quoted in Kania et al) included an AI beauty competition, where almost all the winners had fair skin, reflecting the real-world effects of such bias.

8. FUTURE WORK

   1. Model Improvement

      Future work will focus on expanding the dataset by collaborating with multiple sources. We will use architectures like ResNet-50 and explainable AI with attention mechanisms. We will also integrate different image types and support continuous learning through online updates. Prioritizing the collection of more diverse facial images encompassing a comprehensive range of skin tones, types, ages, and conditions is crucial to mitigate algorithmic bias and ensure equitable performance across all populations, as recommended by Lee et al. and Hash et al. Moving beyond reliance on marketing tags for ground truth data by gathering actual clinical test reports of cosmetics will provide precise enhancement rates for each skin concern, leading to more scientifically robust models, as suggested by Lee et al.

   2. System Improvement

      We will improve scalability by deploying in the cloud and developing mobile apps for iOS and Android. We will optimize real-time processing and include detailed analysis in reporting. We will use statistical process control for automated quality control.

   3. Feature Expansion

   The system will expand to cover more skin conditions, such as wrinkles, dark spots and pigmentation, while integrating databases of larger products and user accounts with history. Future characteristics include skin -based skin analysis in real time, allergy controls and price comparisons. As suggested by Lee et al. and hash et al., future AI models must address broader concerns, such as skin firmness, brightness, cancer

   detection and texture analysis. The incorporation of user specific factors, such as age, gender, medical history and cosmetic procedures, will improve customization. In addition, real and physiological data integration in real time (for example, UV index, moisture, skin pH) can support more adaptive and dynamic recommendations for skin care.

9. CONCLUSION

   This paper presents a AI -based skin care recommendation system that combines convolutional neural networks and natural language processing. Achieved 80% precision in the classification of skin conditions and 89.7% user satisfaction, highlighting the effectiveness of deep learning for personalized skin care. When analyzing the images and the text, the system offers precise and easy to use recommendations through a receptive interface and a rich product database.

   However, as Kania et al. and hash et al., The reliability of such models of AI depends largely on various high quality training data. Concerns such as algorithmic bias, especially for different skin tones, together with data privacy problems, require strong ethical and regulatory supervision. Like Kania et al. Employed, the future lies in combining AI capacities with the experience and empathy of human practitioners.

10. REFERENCES

1. Hashimoto, Wataru and Shohi Canada. "An application for smartphones for personal facial beauty monitoring." Skin Research and Technology 30.7 (2024): E13824.

2. Lee, Jinen, et al. "Recommended product for skin care based on deep education: an approach to the analysis of cosmetic materials and facial skin conditions." Journal of Cosmetic Dermatology 23.6 (2024): 2066- 2077.

3. Su, Zahui, et al. "A systematic review protocol for analyzing facial images based on deep learning in medical research." E047549 in BMJ Open 11.11 (2021).

4. Kania, Barbara, Karen Montessinos and David J. Goldberg. "Artificial intelligence in cosmetic dermatology". Journal of Cosmetic Dermatology 23.10 (2024): 3305-3311.

5. Hash, Mary Grace, et al. "Artificial intelligence in the development of the personal skin care regime." CURUS 17.4 (2025).

6. "Firm nerve network to detect facial skin lesions". Artificial Intelligence and Signal Processing (AISP) 2024 20th International CSI Seminar. IEEE, 2024.