Restricted Boltzmann Machines for Recommender Systems with Implicit Feedback

Explicit and implicit feedback exhibits different characteristics of users' preferences with both pros and cons. However, a combination of these two types of feedback provides another paradigm for recommender systems (RS). Their combination in a user preference model presents a number of challenges but can also overcome the problems associated with each other. In order to build an effective RS on combination of both types of feedback, we need to have comparative data allowing an understanding of the computation of user preferences. In this paper, we provide an overview of the differentiating characteristics of explicit and implicit feedback using datasets mined from Last.fm, an online music station and recommender service. The datasets consisted of explicit positive feedback (by loving tracks) and implicit feedback which is inherently positive (the number of times a track is played). Rather than relying on just one type of feedback, we present techniques for extracting user preferences from both. In order to compare and contrast the performances of these techniques, we carried out experiments using the Taste recommender system engine and the Last.fm datasets. Our results show that implicit and explicit positive feedback complements each other, with similar performances despite their different characteristics.

In this paper, we present our study and characterisation of explicit and implicit feedback on Last.fm, an online music station and recommender service. The dataset consisted of explicit positive feedback (through loved tracks) and implicit positive feedback (the number of times a track is played). As one would expect, our analysis shows that explicit feedback is very scarce. However, we also found that the rate at which a user provides explicit feedback decreases with time, and that overall leaving explicit feedback has a negative effect on the user's behaviour.

Restricted Boltzmann Machines (RBMs) have been developed for a lot of applications in the past few years, and many of its variants have also appeared. In this paper, RBM model and its learning algorithm with contrastive divergence algorithm will be introduced firstly. Then three important variants of RBM are presented in details, which are sparse RBM, discriminative RBM, and the Deep Boltzmann Machines (DBM). All the variants including original RBM are tested on MNIST handwriting digit dataset for classification task. Our empirical results demonstrate the advantage of RBM models and show that compared with other variants, the DBM is the best one in terms of the classification accuracy.

The two main tasks addressed by collaborative filtering approaches are rating prediction and ranking. Rating prediction models leverage explicit feedback (e.g. ratings), and aim to estimate the rating a user would assign to an unseen item. In contrast, ranking models leverage implicit feedback (e.g. clicks) in order to provide the user with a personalized ranked list of recommended items. Several previous approaches have been proposed that learn from both explicit and implicit feedback to optimize the task of ranking or rating prediction at the level of recommendation algorithm. Yet we argue that these two tasks are not completely separate, but are part of a unified process: a user first interacts with a set of items and then might decide to provide explicit feedback on a subset of items. We propose to bridge the gap between the tasks of rating prediction and ranking through the use of a novel weak supervision approach that unifies both explicit and implicit feedback datasets. The key aspects of the proposed model is that (1) it is applied at the level of data pre-processing and (2) it increases the representation of less popular items in recommendations while maintaining reasonable recommendation performance. Our experimental results - on six datasets covering different types of heterogeneous user's interactions and using a wide range of evaluation metrics - show that, our proposed approach can effectively combine explicit and implicit feedback and improve the effectiveness of the baseline explicit model on the ranking task by covering a broader range of long-tail items.

Contrastive Divergence has become a common way to train Restricted Boltzmann Machines; however, its convergence has not been made clear yet. This paper studies the convergence of Contrastive Divergence algorithm. We relate Contrastive Divergence algorithm to gradient method with errors and derive convergence conditions of Contrastive Divergence algorithm using the convergence theorem of gradient method with errors. We give specific convergence conditions of Contrastive Divergence learning algorithm for Restricted Boltzmann Machines in which both visible units and hidden units can only take a finite number of values. Two new convergence conditions are obtained by specifying the learning rate. Finally, we give specific conditions that the step number of Gibbs sampling must be satisfied in order to guarantee the Contrastive Divergence algorithm convergence.

Recommender systems are increasingly being used in university or online education. However, recommender systems still have not found major usage in K12 education. This may be because of the unique challenges that recommender systems face when used by a young and diverse population. However, recommender systems for K12 education could provide many benefits for students and teachers, such as the simplification of personalized learning. Some of the issues with K12 educational recommender systems may be solved by the use of deep learning and implicit feedback. As such, we investigated the use of deep recommendation algorithms and implicit feedback for K12 educational recommender systems. To do this, we compared metrics for highly cited traditional and deep recommendation algorithms trained on explicit and implicit data. We found that recommendation algorithms using deep learning as a group do not differ in performance compared to traditional recommendation algorithms. We also found that the use of implicit feedback led to higher performance than using explicit feedback. The best performing algorithm used both deep learning and implicit feedback. We conclude that deep learning can be a benefit for K12 recommender systems, particularly when the ordered sequence of items is carefully accounted for. We also conclude that researchers and developers must carefully consider which feedback contributes the most information for learning.

In the recommendation system, compared with the explicit feedback, implicit feedback has the characteristics of large quantity, easier tracking and easier access. Even in many practical application scenarios of the recommendation system, there is only implicit feedback without explicit feedback Based on implicit feedback to recommend there has been a problem, that is usually implicit feedback provides only reflect the noise signal of user preferences, because although the observed interactions at least reflect the user’s interest in the item, but not observed the interaction may be missing data, and is not necessarily negative samples, so it’s natural lack of negative feedback In this work, this paper improves the neural colla-borative filtering algorithm (NCF) by extracting the popularity characteristics as the basis for extracting negative samples from implicit feedback data. To a certain extent, it solves the problem that negative sampling introduces a lot of noise in previous work. At the same time, we found that the gradient vanishing in the training process when using the traditional loss function Binary Cross Entropy (BCE). In order to solve this problem, a new loss function, BCE-Max, was proposed in this paper. Based on the above improvements, we proposed a neural collaborative filtering recommendation algorithm (PNCF) based on popularity characteristics. By comparing the performance of PNCF, NCF, traditional collaborative filtering and some matrix factorization recommendation algorithms in the evaluation indexes such as HR@10, NDCG@10 and avg_popularity@10, we find that PNCF has certain advantages over these recommendation algorithms. Meanwhile, We found that compared with NCF, the experimental results of PNCF improved by about 1.2% and 2% onHR@10 and NDCG@10 respectively, and even improved by about 2% on HR@1 and NDCG@I.

Recommendation systems have been the backbone of the digital age, helping users choose relevant information from the pool of data. However, modern recommendation engines have an inherent flaw in the way they take user feedback. The implicit feedback, like history, can be misleading and might not provide the true sentiment of the user, and explicit feedback is incoherent. This paper aims to tackle this flaw by integrating user sentiment into recommender systems using ECG signals from a wearable device. Physiological signals like ECG have been known to contain the emotional information of the user. We propose a combination of classification and recommendation, wherein the AdaBoost classifier uses ECG signals from a wearable device for emotion classification into positive or negative feedback depending on the valence-arousal values, with an accuracy of 0.98. This feedback is then fed to the recommender engine as implicit feedback of the user giving precision and recall of 0.72 and 0.68, respectively. Therefore, the proposed approach allows for accurate and seamless implicit feedback that incorporates user emotions in recommendation systems.

E-commerce platforms usually train their recommender system models to achieve personalized recommendations based on user behavior data. User behavior can be categorized into implicit and explicit feedback. Explicit feedback data have been well studied. However, the implicit feedback data still have many issues, such as the multiple types of behavior data, lack of negative feedback, and lack of the ability to express the real user preference. Targeting these problems of implicit feedback, we propose a TDF-WNSP-WLFM (time decay factor-weight of negative sample possibility-weighted latent factor model) based on the latent factor model for product recommendation. Our method mainly focuses on reconstructing the implicit rating matrix to enable the algorithm to perform better. The TDF-WNSP-WLFM algorithm is tested on two public user behavior datasets from Taobao and REES46, two big e-commerce platforms. Our algorithm compares favorably with other known collaborative filtering methods.

The current study investigates the effects of more explicit versus more implicit corrective feedback on beginner Mandarin learners’ perception and production of Mandarin tones. The effects of oral corrective feedback have been extensively investigated for various domains of second language acquisition (see Mackey & Goo, 2007). For phonological errors, implicit feedback has shown to be particularly salient to learners (Mackey, Gass, & McDonough, 2000) and therefore potentially more effective for tone learning. However, for lower proficiency beginners, explicit corrective feedback has been shown to be more effective (Li, 2009, 2014). Using a mixed methods design, the current study investigated the acquisition of Mandarin tones in beginner, adult learners (n = 41) enrolled in a 14-week, one-on-one, synchronous computer-mediated communication course. Learners were divided into two experimental groups: a more implicit feedback group and a more explicit feedback group. Pretests and posttests assessed changes in tone perception and production. Upon completion of the course, learners in the more implicit feedback group had greater improvement in tone production compared to the more explicit feedback group (d = .75). Both learners and the instructor indicated a preference toward recasts for tone feedback. No statistically significant differences were found for tone perception.

Transfer on Reinforcement Learning (RL) is a promising method on learning new skills and adapting new situations for humanoid robots as tasks or environments change. Within the normal process of Transfer Learning in RL, the inter-task mapping is manually defined, which lacks generalization ability. Therefore, how to automatically learning intertask relations becomes a hot topic. Considering the limited computational resource of a physical humanoid robot, high learning efficiency regarding to both fast speed algorithm and low sample complexity should be emphasized in skills transfer. According to this view, in this research, the inter-task relations are modelled using a three-way Restricted Boltzmann Machine (RBM), which is turned out to be a powerful model in capturing the similarity between samples from source task and target task. Since standard Contrastive Divergence (CD) algorithm commonly used for RBM learning suffers from the inputindependent problem and may lead the learning process timeconsuming or inapplicable, a Cyclic Contrastive Divergence (CCD) learning algorithm is employed. In order to evaluate the performance, experiment that transfer the skill of walking on flat surface to the skill walking on slope surface is conducted on our physical robot platform, PKU-HR5.1, and the result indicates that the method is feasible and efficient.

In this study it was attempted to investigate whether different CF strategies, including implicit, explicit, and emergent CF, can differently affect the accuracy and fluency of oral production among Iranian EFL learners. In addition, it explored the learners' attitudes towards how they felt about the CF types applied in the classroom. For these purposes, the researchers selected 54 homogeneous pre-intermediate learners on the basis of the PET results and randomly assigned them into three experimental groups: the Implicit group, the Explicit group, and the Emergent group. While the Implicit and Explicit groups received just implicit feedback and explicit correction for their erroneous oral production, respectively, the Emergent group took CF from implicit to explicit. Oral narrative and picture description tasks and semi-structured interviews were used to collect the quantitative and qualitative data for the study. The results of inferential statistics indicated significant differences among the feedback types in both past-tense and future-tense accuracy. Moreover, the Emergent group had a better performance compared with the other groups with respect to accuracy. However, no significant difference was revealed among the feedback types with respect to fluency. The results of the content analysis also indicated that the learners mostly preferred to receive emergent feedback and participate in the process of error correction. The findings of this study can raise researchers', teachers', and teacher trainers' awareness of the function of various CF types.

Recommender systems are becoming a widespread technology used to promote cross-selling. Collaborative filtering is one of the main paradigms employed to offer recommendations to users. However, while most collaborative filtering methods require explicit user feedback, such as ratings, it is a well-established fact that users rate only a small portion of all available products. Subsequently, the rating system often acquires insufficient explicit feedback, thus leading to unsatisfactory recommendations. We propose a novel approach in the implicit feedback recommender system domain that combines clustering and matrix factorization to yield good results while using only implicit feedback on users purchase history and without requiring any parameter. We use a high-dimensional, parameter-free, divisive hierarchical clustering technique and, based on the clustering results, create personalized recommendations of high interest for each user. This easy to implement and very effective technique can be applied to any data sets where we can identify users with a purchase history.

Recommendation with implicit feedback has been extensively studied in recent years. It is more difficult to provide users with stable and accurate recommendation compared to recommendation with explicit feedback, due to the reason that interactions from implicit feedback datasets do not clearly indicate the level of user preference. Most existing methods dealing with implicit feedback have achieved excellent performance by focusing on other aspects rather than directly inferring user preference. In this paper, we offer accurate recommendation to users by addressing the problem of directly inferring user preference from implicit feedback with such less information and huge uncertainty. We propose a novel mixture-preference model (MPBMF), which introduces a set of pseudo-preference values to surmise the true user preference. More specifically our proposed model can be described as a Gaussian mixture model in which each single model is trained with pseudo-preferences which show the user's different views for items. Then the predicted user preference is estimated by the models under different pseudo-preferences with different contributions. We conduct extensive experiments on three real-world datasets, and the superior performance demonstrates the effectiveness of our model.

A gas path fault diagnosis scheme for turborfan engines based on deep belief network (DBN) is presented. The scheme is constructed according to the diagnosis principles of gas path faults and is composed of a turbofan engine reference model and a DBN diagnosis model. The DBN diagnosis model is a stacked network of several restricted Boltzmann machines (RBM) and was trained with the contrastive divergence algorithm and the back propagation algorithm. To optimize the DBN performance, the orthogonal tests L25 (5 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7</sup> ) were adopted to determine the hyper-parameters, such as learning rate, hidden layer number, hidden layer neuron number, etc. The proposed DBN-based scheme was applied to diagnose the gas path faults of a turbofan engine model and compared with BP-based and SVM-based schemes. The results show that the fault diagnosis accuracy of the DBN-based scheme is as high as 96.59%, and the DBN-based scheme has dramatic performance advantages over the other two schemes.