Deep FM: Advanced Recommendation System Architecture Explained

Deep FM enhances recommendation systems by combining factorization machines with deep neural networks for capturing complex relationships.

### Understanding Deep FM: An Advanced Approach to Recommendation Systems
* 📚 Today, we will explore **deep FM**, an advanced version of factorization machines, and its application in **recommendation systems**.
* 🔍 Prior to this discussion, concepts such as **collaborative filtering**, **metric factorization**, **NCF**, and **factorization machines** were covered to provide foundational knowledge.
* 📈 Factorization machines extend linear regression by incorporating **user-item interactions** alongside user and item features, which is **crucial for effective recommendations**.
* 🤖 Deep FM operates as a neural network variant of factorization machines, enabling the **capture of non-linear relationships in data**, enhancing the **recommendation process**.
* 🔗 The architecture of deep FM is similar to NCF, where the input features are transformed into **one-hot encodings**, followed by an embedding layer that generates **fixed-size embeddings** for each feature.
* ⚙ These embeddings are then processed through two main components: the **factorization machine part** and the **deep neural network (DNN) part**, which together contribute to the **final output**.
* 📊 The outputs from both components are concatenated, and a **sigmoid activation function** is applied to produce the final recommendations, illustrating the **synergy between factorization machines and deep learning techniques**.
* 💡 Overall, deep FM integrates the strengths of **embedding layers and neural networks** to enhance the performance of recommendation systems **significantly**.

uh, so hi everyone. Today we'll be discussing what deep FM is and how they can be used for recommendation systems. So let's get started. So prior to this, I have already explained what collaborative filtering is, what metric factorization, NCF, and factorization machine are. So today we will be talking about deep FM. So before moving on to deep FM, we must know that deep FMs are considered as an advanced version of factorization machines. So this is something that I've already explained in my previous videos. So first of all, we'll try to recapsulate what factorization machines are. So factorization machines are nothing but an extension of linear regression, where apart from capturing the user-item interaction, we're also using the user and the item features as well while making the recommendation system. So deep FM can be considered as a neural network version of factorization machine. So in the case of factorization machine, we are trying to capture higher-order relationships. By using deep FM, we would be trying to capture non-linear relationships as well. So to understand

what more about factorization machine how they were what is the equation you can follow up in a previous video so we're talking about the architecture of deep FM so deep FM like saying is pretty much similar to ncf so ncf is for ncif can be taken as an extension of collaborative filtering and Matrix factorization similarly a deep FM is an extension of factorization machine so we'll start off with all the features that we get assume that we have features field one field I field J till field M now this will include the user item interaction alongside user and item features also the next step is to convert these features into one or encode so if we have categorical features in this particular data set they would be converted into space into one hot encodes if the continuous switches will leave them as well as them as it is the next layer is an embedding layer so these one word encoded features will be then fed to dense embedding layer where we will be generating embeddings for each of the one hot encoded feature along the

current switches for each feature we will be generating an embedding of other fixed size. Now, once we have generated embeddings for each of the features, we do divide this architecture into two parts: one is the factorization machine part, and the other is the DNN part. So, let's first of all start with the input and output. So, as we convert that, deep FM is an extension of a factorization machine. The input is quite similar to what goes into the factorization machine, that is, one of the encoded features. The next layer is the embedding layer, as I have already told you. Now, once the embedding layer is generated, the embeddings are then passed onto two parts: one is the factorization machine, that is, the left side of the diagram, as told here, where we are following this particular equation: y equals to W 0 plus summation w i into x i, the summation summation uh these VI dot VJ. VI represents the latent vector x i into x j. So, to understand this whole equation, you need to go to my previous video and understand what this equation is all about.

about because I have already explained a factorization machine. Here, I have given you a level key. What is y-wise label? W is bias-wise weights. x_i is the features for one art encoder feature set. nww.red within latent vectors. A DOT product is basically the meaning that we are generating now. This is the left-hand side. Moving on to the right-hand side is a DNN. So, this is nothing but very similar to the ncf multi-layer perceptron that we developed. So, what happens here is that we're feeding all these embeddings into a general neural network, which comprises multiple hidden layers, and the final layer is a sigma. The final output goes through a sigma of activation function. So, basically talking about the final output, the output from the factorization machine, that is the left part, and the DNN, that is the right part, are then combined together, concatenated, and then a sigmoid activation function is applied. Deep FM can be taken as embedding layer and sigmoid over factorization machine plus DNN. That's it. It is nothing complicated and very, very

similar to NCO, where NCF is an extension for a metric factorization.

DeepFM for recommendation systems explained

Welcome to Data Science In Your Pocket! Dive deep into advanced AI and Data Science topics not thoroughly covered elsewhere. Discover comprehensive insights and answers to complex questions, all in one place. Subscribe for in-depth tutorials and expert knowledge on cutting-edge data science and AI.

We have tutorials on Generative AI, Reinforcement Learning, NLP, Time Series, Graph Analytics and other major Data Science domains


Data Science in your pocket

Understanding Deep FM for Recommendation Systems

Deep FM is an advanced version of factorization machines, aiming to capture both higher-order and non-linear relationships in user-item interactions.

The architecture of Deep FM is similar to NCF, with a focus on extending factorization machines.

The process involves encoding features, generating embeddings, and using them to make recommendations.

Deep FM: Combining Factorization Machine and DNN

Deep FM generates fixed-size embeddings for features and divides the architecture into two parts: the factorization machine and the DNN.

The factorization machine uses a specific equation to generate outputs from the embeddings, while the DNN feeds the embeddings into a neural network with multiple hidden layers.

The final output is a combination of the factorization machine and DNN outputs, followed by a sigmoid activation function.

NCF: An Extension for Metric Factorization

NCF is an extension of the NCO framework, specifically designed for metric factorization.

It aims to enhance the performance of factorization methods by incorporating additional metrics.

This approach provides a more comprehensive solution for various applications in data analysis and machine learning.