M.A.BDEIRA//
SCRL 000%
SCENE 01 / 09
// SYSTEM READY
MODE DÉMO INTERACTIF // MASTER THESIS UCBET 2024 ENIC-NARIC LEVEL 7
LINKLEARN AI (LINKEDLEARN) // INTELLIGENT COMPUTER SYSTEMS // DEFENDED JUNE 23, 2024

DEEP LEARNING COURSE RECOMMENDER & CLAUDE-3-OPUS E-LEARNING ENGINE

Interactive research & full-stack demonstration of “Towards a Deep Learning-Based E-learning Platform” by Mostepha Amine Bdeira. Combines a CodeIgniter 4 LMS & Python Flask inference microservice with Keras DLRM (50-dim Latent Embeddings + Dense-100 ReLU) and Scikit-Surprise SVD trained on 3,522 Coursera courses & MovieLens (610×9,724), plus automated video quiz synthesis via YouTubeTranscriptApi + Anthropic Claude-3-Opus.

CodeIgniter 4 (PHP MVC) Python Flask (/recommend) Keras / TensorFlow DLRM Scikit-Surprise SVD (svd_model.pkl) NLTK Stopword Pipeline Claude-3-Opus API YouTubeTranscriptApi Video.js + Linked Data MySQL + Tailwind CSS
// THESIS CREDENTIALS 2023 / 2024
UNIVERSITY: Chadli Bendjedid (UCBET)
SPECIALTY: Intelligent Computer Systems
AUTHOR: Bdeira Mostepha Amine
SUPERVISOR: Mr. CHEMAM C.
JURY: Mr. BOUTOUIL A. · Mr. BENTRAD S.
01 // TRAINING CORPUS

3,522 COURSES

Coursera Dataset (137.7+ KB, NLTK filtered) × MovieLens 610 users / 9,724 items

02 // DLRM MODEL 2 ERROR

RMSE 0.8718

-20.8% lower RMSE vs Model 1 (1.1004) · MAE: 0.6668 (526,901 params)

03 // TOP-K RANKING ACCURACY

71.46% RECALL@K

Precision@K: 70.53% · F1@K: 70.48% · 50-dim Latent Factors

04 // GENERATIVE AI ASSESSMENT

CLAUDE-3-OPUS

YouTubeTranscriptApi captions → 10 auto-graded quiz questions + Video.js RDF notes

MODULE 01 // LIVE COURSE RECOMMENDATION ENGINE (FLASK + KERAS DLRM + SURPRISE SVD)

PERSONALIZED COURSERA RECOMMENDER SIMULATOR

Select a learner vector (`user_id`), switch between Model 2 (Concatenate + Dense-100 ReLU), Model 1 (Dot-Product), or Flask Surprise SVD (`svd_model.pkl`), and inspect the predicted top-N courses & JSON microservice response.

02 // NLTK SKILL TOKEN BOOST ALL SKILLS

Courses in `Coursera.csv` (`3,522` rows) are preprocessed with Python `NLTK` (`stopwords.words('english')`) to tokenize `title`, `description`, and `skills`. Click a token to boost matching courses:

GET http://127.0.0.1:5000/recommend?user_id=104&model=dlrm_v2
LATENCY: 14.2 ms RMSE: 0.8718
RANK #01 · 98.4% MATCH PRED: 4.94 / 5.0 ★

Introduction to Python Programming

University of Pennsylvania (Coursera) — Core Python syntax, data structures, functions, and algorithmic problem-solving.

Beginner python data-structures
RANK #02 · 96.9% MATCH PRED: 4.89 / 5.0 ★

The Top 5 Machine Learning Libraries in Python

Coursera Guided Project — Hands-on model training and evaluation with Scikit-Learn, TensorFlow, Keras, PyTorch, and XGBoost.

Intermediate machine-learning keras
RANK #03 · 95.2% MATCH PRED: 4.85 / 5.0 ★

100 Days of Code: The Complete Python Pro Bootcamp

Full-stack Python automation, Flask microservices, SQL integration, Pandas data analysis, and REST API engineering.

Beginner → Pro flask python
RANK #04 · 93.7% MATCH PRED: 4.78 / 5.0 ★

First Steps in Linear Algebra for Machine Learning

National Research University Higher School of Economics — Vector spaces, matrix decomposition (SVD), eigenvalues, and latent embeddings.

Intermediate linear-algebra svd
FLASK -> CODEIGNITER 4 PAYLOAD (`app/Controllers/CourseController.php`) HTTP 200 OK · application/json
{
  "user_id": 104,
  "engine": "Keras DLRM Model 2 (Concatenate + Dense-100 ReLU)",
  "embedding_dim": 50,
  "metrics": { "rmse": 0.8718, "mae": 0.6668, "precision_at_k": 0.7053, "recall_at_k": 0.7146 },
  "recommendations": [
    { "course_id": 412, "title": "Introduction to Python Programming", "predicted_rating": 4.94, "match": "98.4%" },
    { "course_id": 1089, "title": "The Top 5 Machine Learning Libraries in Python", "predicted_rating": 4.89, "match": "96.9%" },
    { "course_id": 2104, "title": "100 Days of Code: The Complete Python Pro Bootcamp", "predicted_rating": 4.85, "match": "95.2%" },
    { "course_id": 874, "title": "First Steps in Linear Algebra for Machine Learning", "predicted_rating": 4.78, "match": "93.7%" }
  ]
}
MODULE 02 // THESIS CHAPTER 3 & 4 — DEEP LEARNING RECOMMENDATION MODELS (KERAS / TENSORFLOW)

MODEL 1 (DOT-PRODUCT) VS. MODEL 2 (CONCATENATE + DENSE RELU)

WINNER: MODEL 2 (-20.8% RMSE REDUCTION)
MODEL 1 // DOT-PRODUCT EMBEDDING 516,700 TRAINABLE PARAMS

LINEAR INNER-PRODUCT MATRIX FACTORIZATION IN KERAS

Projects both User_Input (None, 1) and Course_Input (None, 1) into 50-dimensional dense latent vectors, flattens them, and computes their scalar Dot_Product (None, 1). Captures linear user-item affinity but suffers from validation loss drift after Epoch 6.

User_Embedding (Embedding 610×50)30,500 params
Course_Embedding (Embedding 9724×50)486,200 params
User_Vector & Course_Vector (Flatten)(None, 50)
Dot_Product (axes=1) → Output(None, 1)
RMSE 1.1004
MAE 0.7791
PREC@K 70.14%
RECALL@K 71.09%
F1@K 69.82%
MODEL 2 // CONCATENATE + DENSE-100 RELU ★ 526,901 TRAINABLE PARAMS

NON-LINEAR NEURAL COLLABORATIVE FILTERING (DLRM)

Concatenates the flattened 50-dim User and Course embeddings into a (None, 100) joint feature vector and passes it through a Dense(100, activation='relu') hidden layer (10,100 params) before the Dense(1) regression head (Adam lr=0.001, MSE loss).

User_Embedding (610×50) + Course_Embedding (9724×50)516,700 params
Concatenate ([User_Vector, Course_Vector])(None, 100)
dense_2 (Dense 100 units, ReLU activation)10,100 params
dense_3 (Dense 1 unit, Linear output)101 params
RMSE (-20.8%) 0.8718
MAE (-14.4%) 0.6668
PREC@K 70.53%
RECALL@K 71.46%
F1@K 70.48%
// 20-EPOCH MSE LOSS TRAJECTORY (THESIS FIG. 38 & FIG. 42)

Model 2 (Concatenate + Dense-100 ReLU) stabilizes test MSE at ~0.76 (RMSE 0.8718), whereas Model 1’s inner product overfits after Epoch 6.

Model 2 Test Loss (0.76 MSE) Model 1 Test Loss (1.21 MSE) Training Loss
12.0 MSE 2.00 MSE 0.80 MSE 0.15 MSE Model 2: RMSE 0.8718 ★ Epoch 1 Epoch 5 Epoch 10 Epoch 15 Epoch 20
MODULE 03 // AUTOMATED ASSIGNMENT GENERATION (YOUTUBETRANSCRIPTAPI + CLAUDE-3-OPUS + VIDEO.JS)

LECTURE TRANSCRIPT → AI QUIZ SYNTHESIS & VIDEO ANNOTATIONS

In Thesis Section 5.6 (Assignment Generation Page, Fig. 62–71), instructors select lecture captions extracted via YouTubeTranscriptApi.get_transcript(video_id) and invoke claude-3-opus-20240229 (max_tokens=4000, temp=0.7) to generate interactive assessments.

EXTRACTED TRANSCRIPT (`YouTubeTranscriptApi`) MySQL: Captions Fetched

[00:14.2] "In Python, variables are created the moment you first assign a value to them using the = operator."

[01:08.5] "Unlike C or Java, Python is dynamically typed: a variable can hold an integer (x = 4) and later be reassigned to a string (x = 'LinkLearn')."

[02:19.0] "Variable names are case-sensitive (age, Age, and AGE are three distinct identifiers) and must start with a letter or underscore."

02 // VIDEO.JS + LINKED DATA ANNOTATIONS store_annotation(video_id, ts, content)
@ 00:42.0s Linked Data URI: dbpedia.org/resource/Dynamic_typing
RDF URI
@ 02:00.5s Instructor Note: Remember that `type(x)` returns the runtime class of any variable.
NOTE
// CLAUDE-3-OPUS GENERATED ASSESSMENT (`claude-3-opus-20240229`)

Topic: Introduction to Python — Variables & Dynamic Typing (3 Auto-Graded Questions)

SCORE: 0 / 3
API Config: model="claude-3-opus-20240229", max_tokens=4000, temperature=0.7
// MASTER THESIS REPORT EXHIBITS (UCBET — JUNE 2024)

END-TO-END SYSTEM PIPELINE & DATABASE SCHEMA

VIEW FULL CASE STUDY & SCENE EXHIBITS
01 // CODEIGNITER 4 ↔ FLASK SVD BRIDGE

The PHP CourseController::recommend($userId) issues an HTTP GET request to the Python Flask microservice (http://localhost:5000/recommend?user_id=...), which unpickles svd_model.pkl / evaluates the Keras DLRM embeddings and returns top-N ranked course IDs.

02 // MYSQL RELATIONAL SCHEMA

Structured across users, courses (3,522 Coursera records), ratings (MovieLens user-item matrix), videos_db (YouTube IDs + fetched transcripts), annotations, and assignments.

03 // SEMANTIC WEB & LINKED DATA

Grounded in Semantic Web standards (RDF triples, OWL ontologies, SPARQL endpoints, and JSON-LD educational metadata) to connect video timestamps and course skills with external knowledge graphs like DBpedia.

LinkLearn AI Recommended Courses Dashboard
Fig. 55 — Personalized Course Recommendations Top-N Coursera recommendations served via Flask SVD & Keras DLRM.
Select Videos for Claude-3-Opus Assignment Generation
Fig. 62 — AI Assignment Generation (Claude-3-Opus) Selecting lecture transcripts to synthesize 10-question quizzes.
VideoDB YouTubeTranscriptApi Captions Fetched
Fig. 65 — VideoDB & YouTubeTranscriptApi Automated extraction of timestamped YouTube captions into MySQL.
Interactive Video Lecture Player and 23 Python Lessons
Fig. 69 — Interactive Video Lecture Player 23-lesson Python curriculum with completion tracking & annotations.
Video.js Timestamped Annotation Modal
Fig. 71 — Timestamped Video Annotation Modal Learners & instructors attach notes and Linked Data URIs to video frames.
Keras DLRM Model 2 Performance Metrics
Fig. 43 — DLRM Model 2 Evaluation Metrics RMSE: 0.8718 · MAE: 0.6668 · Precision@K: 70.53% · Recall@K: 71.46%.