Improved collaborative filtering algorithm using intuitionistic fuzzy C -means clustering and incorporating time factors

Abstract

Collaborative filtering (CF) has been widely used to help users discover items of interest. Most of the existing CF methods collect preference information from other users to predict user’s ratings. However, the user’s interest will change over time, and the traditional collaborative filtering recommendation does not take the change of user’s interest into consideration. In addition, there are also problems of information overload and data sparseness in applications. In order to solve the problem of data sparseness and take the changes of user interests into account , we propose an improved collaborative filtering algorithm using intuitionistic fuzzy C-means clustering and incorporating time factors. Firstly, we introduce intuitionistic fuzzy C-means clustering to complete the classification of items; secondly, the above classification results are combined with the user-item score matrix to obtain the user-type preference matrix; thirdly, the joint similarities are used to obtain the nearest neighbors of the target user; finally, the changes of user interest over time are taken into account when predicting the target user’s rating on unrated items, top-k items are selected as recommendation results. To verify the effectiveness of the algorithm, experiments are conducted on Movie Lens and Booking-Crossing datasets. The Improved collaborative filtering algorithm using intuitionistic fuzzy C-means clustering results show that compared with the traditional recommendation algorithms , the algorithm proposed in this paper can effectively solve data sparseness and improve the quality and accuracy of recommendation.

GIF: A General Graph Unlearning Strategy via Influence Function

“…At present, the research on graph unlearning is still at an early stage and there are several open research questions. In future work, we would like to explore the potential of graph influence function in other applications, such as recommender systems [37,42], influential subgraph discovery, and certified data removal. Going beyond explicit unlearning requests, we will focus on the auto-repair ability of the deployed model -that is, to discover the adversarial attacks to the model and revoke their negative impact.…”

Robust Collaborative Filtering to Popularity Distribution Shift

“…Generalized methods [12,21,54,56,59,66,68,69] aim to learn invariant representations against popularity bias and achieve stability and generalization ability. s-DRO [59] adopts Distributionally Robust Optimization (DRO) framework, CD 2 AN [12] disentangles item property representations from popularity under co-training networks.…”

“…COR [54] formulates feature shift as an intervention and performs counterfactual inference to mitigate the effect of out-of-date interactions. BC Loss [66] incorporates popularity bias-aware margins to achieve better generalization ability. InvCF [68] discovers disentangled representations that faithfully reveal the latent preference and popularity semantics.…”

Top-Personalized-K Recommendation

“…It is noted that when 𝑌 𝑢,𝑖 = 0, it indicates either that the item is irrelevant to the user or that it can be a hidden-relevant item of the user [53] [25], neural networks [13,35], and graph neural networks [12,64]. To train recommender models, point-wise loss (e.g., binary cross-entropy, mean squared error), pair-wise loss (e.g., BPR [50], Margin Ranking loss [61]), and listwise loss (e.g., InfoNCE [44], Sampled Softmax [67]) have been adopted [21]. After the recommender model is trained, we produce a ranking list 𝜋 with unobserved items in I − 𝑢 by sorting the ranking scores 𝑓 𝜃 (𝑢, 𝑖) in descending order.…”