Enhancing Top-N Recommendation Using Stacked Autoencoder in Context-Aware Recommender System

Abstract

Context-aware recommender systems (CARS) are a vital module of many corporate, especially within the online commerce domain, where consumers are provided with recommendations about products potentially relevant for them. A traditional CARS, which utilizes deep learning models considers that user’s preferences can be predicted by ratings, reviews, demographics, etc. However, the feedback given by the users is often conflicting when comparing the rating score and the sentiment behind the reviews. Therefore, a model that utilizes either ratings or reviews for predicting items for top-N recommendation may generate unsatisfactory recommendations in many cases. In order to address this problem, this paper proposes an effective context-specific sentiment based stacked autoencoder (CSSAE) to learn the concrete preference of the user by merging the rating and reviews for a context-specific item into a stacked autoencoder. Hence, the user's preferences are consistently predicted to enhance the Top-N recommendation quality, by adapting the recommended list to the exact context where an active user is operating. Experiments on four Amazon (5-core) datasets demonstrated that the proposed CSSAE model persistently outperforms state-of-the-art top-N recommendation methods on various effectiveness metrics.

Appendices

Appendix

In this section, the basic architectural components of the auto-encoder have been described and the changes introduced to implement the proposed Context-specific Sentiment based Stacked AutoEncoder (CSSAE) method to fit with the recommendation problem is described in detail with an example with a table provided with the parameters of each layer and the layer input shapes.

1.1 An autoencoder consists of three components

• Encoder: An encoder is a feedforward, fully connected neural network that a compress i.e. encodes the input into a latent space representation.

• Code: This part of the network contains the reduced representation of the input that is fed into the decoder.

• Decoder: Decoder is also a feedforward network like the encoder and has a similar structure to the encoder. This network is responsible for reconstructing the input back to the original dimensions from the code.

1.2 The following are the changes made to the CSSAE model to fit with the recommendation problem

However, to consider the multiple feedback scenario, each context-based item has multiple feedback that is required as input to the network. Therefore, as shown in Fig. 2 the traditional stacked autoencoder was expanded by integrating an additional layer which holds two input layer for both the rating and sentiment data. Thus, the additional layer acts as the input layer in the context-specific sentiment based stacked autoencoder, and Table

Table 8 The following table provides the parameters of each layer and the layer input shapes

8 provides the parameters of each layer and the layer input shapes.

Note: The shape of the input layer is the total number of context-based items in the training set is 19366.

The first layer acts as a dual input layer i.e. two input layers (R_Score[0][0] and S_Score[0][0]) which are linked to their respective custom layer where elementwise_weights are trained for each node. Hence the particular node in the custom layer elementwise_weights_14[0][0] which is connected to R_Score has the value $r_{ij}^{{}} \times w_{j}^{1}$ and elementwise_weights_15[0][0] which is connected to S_Score has the value $s_{ij}^{{}} \times w_{j}^{2}$ and both which in turn is fed into the lambda layer for merging two inputs which hold $r_{ij}^{{}} \times w_{j}^{1} + s_{ij}^{{}} \times w_{j}^{2}$.

Hence corresponding item nodes in the following intermediate layer (lambda_7) has the value as mentioned in Eq. (3)

$$ r_{ij}^{I} = r_{ij}^{{}} \times w_{j}^{1} + s_{ij}^{{}} \times w_{j}^{2} $$

where $r_{ij}^{I}$ be the combined rating value of an item j rated by user i which is obtained from both $r_{ij}^{{}}$, the rating and $s_{ij}^{{}}$ the sentiment score obtained from the review specified by user i to item j.

This intermediate layer (lambda_7) which in turn is linked to consecutive M encoding layers that are being concatenated with User_sideInfm that are employed to find out the hidden representation of the items. The final encoding layer (LatentSpace) is linked to M successive decoding layers that are used to decode the hidden factors acquired from corresponding encoders. The Final decoding layer (UserScorePred) acts as an output where the actual concrete ratings of the items are predicted.

We have selected the intermediate layer (lambda_7) and its successive set of layers as an autoencoder because the nodes in the intermediate layer (lambda_7) has the merger of values of both the rating and sentiment scores as mentioned in Eq. (3)

$$ r_{ij}^{I} = r_{ij}^{{}} \times w_{j}^{1} + s_{ij}^{{}} \times w_{j}^{2} $$

Hence this intermediate layer acts as the input to the autoencoder, which encodes the input to find out the hidden representation of the items. Therefore, the latent feature vector (code) in the stacked autoencoder fits the ratings and reviews simultaneously. The final decoding layer (UserScorePred) acts as an output where the actual concrete ratings of the context-based items are predicted.

Note

In the proposed methodology, Sentiment Analysis is performed using Textblob on Users’ review on context-based items in-order to predict the Numerical sentiment score. This numerical sentiment score is given as the input into stacked autoencoder, a fully-connected neural network that was expanded by integrating an additional layer that holds two input layers for both the rating and sentiment data, where the training takes place end-to-end.