At Travelix, our recommendation system relied on three primary datasets: Business, Users, and Reviews. Due to the large size of these datasets, we employed a preprocessing step to divide the Business data by state. We identified the top ten states based on the highest number of businesses and used them for generating recommendations. Further categorization of the Business data was carried out to focus on the three main categories of interest: Hotels & Travel, Restaurants, and Nightlife. This enabled us to create more focused and relevant recommendations for our users based on their interests and preferences. Some of the subcategories are as follows

Matrix Factorization is a technique used in machine learning and recommendation systems to decompose a matrix into two lower-rank matrices. In the context of recommendation systems, the matrix represents the user-item interactions, where the rows represent the users, the columns represent the items, and the values represent the ratings or preferences of the users for the items. By factorizing the matrix into lower-rank matrices, Matrix Factorization can identify latent factors that explain the observed user-item interactions, such as user preferences or item characteristics. Matrix Factorization can be used to make personalized recommendations to users based on their preferences and past behavior. The benefits of Matrix Factorization in recommendation systems include improved accuracy, scalability, and interpretability of the recommendations. Additionally, Matrix Factorization can handle sparse data, which is common in recommendation systems where users have only rated a small fraction of the available items.

We implement the Matrix Factorization algorithm for implicit feedback data using the MF_implicit class that takes as input the training rating matrix train_mat, the number of latent factors latent, the learning rate lr, and the regularization weight reg. The class contains the methods negative_sampling() to sample negative user-item pairs to train the model, train() to train the model for a given number of epochs, and predict() to generate the recommendation list for each user. In the train() method, the model is trained using stochastic gradient descent by minimizing the difference between the predicted and actual ratings of the user-item pairs. The predict() method generates the top 50 ranked list of recommendations for each user based on the learned latent factors.

Autoencoders and collaborative filtering are both widely used techniques in the field of machine learning and for designing Recommendation engines. Autoencoders are a type of neural network that can learn to compress and reconstruct high-dimensional data, such as images, while minimizing the reconstruction error. Collaborative filtering, on the other hand, is a technique used for recommendation systems, where the system predicts user preferences based on past behavior and similarity to other users. By analyzing user behavior and preferences, collaborative filtering can suggest products, services, or content that users are likely to enjoy.

We have implemented an autoencoder-based collaborative filtering algorithm for personalized recommendation in a travel recommendation system. The autoencoder model is trained on a sparse matrix of user ratings, with the goal of learning embeddings that capture user preferences. The learned embeddings are then used to compute similarity scores between users and items, which are used to predict the ratings of the items for each user. The model is evaluated using the root mean squared error (RMSE) metric on a held-out test set, and the top recommended items for a given user are returned based on the predicted ratings.

To implement the Autoencoder, we used 3 hidden dense layers with 64, 32 and 16 neurons during encoding and 2 hidden layers with 32 and 64 neurons during decoding. We are using the ReLU activation function in the hidden layers. We are also sandwiching the dropout layers between all the hidden layers and the last hidden layer and output layer. In the output layer we are using the “sigmoid” as the activation function. To train the Autoencoder as defined above, we are using the “Adam” optimizer and “Binary Cross-Entropy” loss function. After this we train the model taking a batch size of 32 and training the model for 200 epochs. We also implemented a learning rate scheduler to find the optimal hyperparameter during training. We shuffle the data before each epoch and use a validation set as well.

Autoencoders can help solve problems in recommendation systems by learning a low-dimensional representation of the data that captures the underlying structure and patterns in the user-item interactions. This can be useful in handling the sparsity and high dimensionality of the data, as well as improving the accuracy and relevance of the recommendations. Autoencoders can also be used in conjunction with other techniques such as collaborative filtering, content-based filtering, and hybrid methods to further improve the performance and robustness of the recommendation system.