Recommendation Using Restricted Boltzmann Machine

Restricted Boltzmann Machines are stochastic neural networks which focuses on learning the underlying probability distribution. It uses a set of Visible and Hidden nodes. It is based on energy minimisation across a set of nodes. Add Stochasticity and Hidden nodes to a Hopfield Network, and you would get a Restricted Boltzmann Machine.

The Restricted part of RBMs come from the fact that the Pairwise interactions are only allowed within a set of observable and unobservable states. The visible layer or the hidden layers do not have recurrent connections within them, nor are they self connected. This property makes the algorithm able to always reach convergence, also making it possible for RBMs to be trained by efficient optimisation algorithms like contrastive divergence.

Hopfield networks memorise the data, and hence cannot be used for recommendation purposes.

Link to the collab notebook : RBM_Recommender.ipynb

Scoring System Analysis

Approach:

1. Match Interests: Study materials that align with the student’s interests will receive a higher score.
2. Match Course: If the material subject is related to the student's course, the material will also get a higher score.
3. Quiz Performance: Higher average quiz scores could indicate that the student is more capable of handling harder study materials.
4. Past Engagement: If the student has viewed certain materials, those materials will not be recommended again.

Synthetic Data Used

Student Table

student_id	name	course	year	avg_quiz_score	interests
1	John Doe	Computer Science	3	78	AI, Blockchain
2	Jane Smith	Mechanical Engineering	2	65	Environmental Science, AI
3	Sam Lee	Civil Engineering	4	85	Sustainability, Data Science
4	Alex Chen	Computer Science	1	55	AI, Fintech
5	Maria Perez	Electrical Engineering	3	70	Robotics, Blockchain
6	Tom White	Physics	2	63	Quantum Computing, Physics
7	Sophia Brown	Biology	4	88	Biotechnology, Data Science
8	Liam Green	Mathematics	1	57	Mathematics, AI
9	Olivia Black	Cybersecurity	3	74	Cybersecurity, IoT
10	Emma Gray	Mechanical Engineering	2	68	Mechanical Engineering, Robotics

Material Table

material_id	subject	difficulty_level	popularity_score	content_length
101	AI	Medium	65.3	15.2
102	Blockchain	Hard	72.5	22.6
103	Sustainability	Medium	88.7	35.1
104	Robotics	Hard	91.4	27.8
105	Environmental Science	Easy	79.2	33.4
106	Data Science	Hard	86.1	12.3
107	Cybersecurity	Medium	73.4	30.5
108	Physics	Hard	67.8	28.9
109	Mathematics	Medium	94.5	14.8
110	Chemistry	Hard	60.2	37.2
111	Astronomy	Hard	77.1	18.9
112	Quantum Computing	Medium	82.9	24.6
113	Biotechnology	Medium	68.4	10.5
114	Machine Learning	Medium	90.0	20.1
115	Renewable Energy	Hard	84.3	36.4
116	Ethics in AI	Hard	78.8	19.8
117	Web Development	Medium	92.7	12.7
118	Embedded Systems	Easy	63.5	31.5
119	IoT	Hard	70.1	25.9
120	Cloud Computing	Easy	89.6	11.4

Engagement Table

student_id	material_id	rating	viewed
1	101	4	Y
2	102	5	N
3	103	3	Y
4	104	4	Y
5	105	5	N
6	106	2	Y
7	107	3	Y
8	108	5	N
9	109	4	Y
10	110	4	Y
1	111	3	N
2	112	5	Y
3	113	2	Y
4	114	4	Y
5	115	3	N
6	116	5	Y
7	117	2	Y
8	118	4	N
9	119	4	Y
10	120	5	Y
1	101	2	Y
2	102	3	N
3	103	5	Y
4	104	4	Y
5	105	3	N
6	106	2	Y
7	107	4	Y
8	108	5	N
9	109	3	Y
10	110	2	Y

Motivation

I could have simply built a recommender system using collaborative or content filtering. But this small work is my way of paying a tribute to Geoffrey Hinton and John Hopfield.

The RBM is the perfect way to do that. The limitations of the Hopfield Networks led to the development of probabilistic models like RBMs. Also, Geoffrey Hinton has a paper where he explains how to use RBMs for building a recommendation system.

Whats Next?

I plan to write my own implementation of a RBM. So stay tuned for that!