• No results found

Effect of Cluster Refinement

7.2 ClustMF Method

7.3.2 Effect of Cluster Refinement

In this study, we compare the effect of adding cluster refinement to the ClustMF. The results from this study are presented in Table 7.1. The column corresponding to ClustMF NCR (No Cluster Refinement) is the ClustMF method which does not refine the clusters. We can see that adding the cluster refinement based on the ranking crite- rion further helps to improve the recommendation performance in terms of the hit-rate.

7.3.3 Comparison With Other Approaches

Table 7.2 shows the overall recommendation performance of the MPCF methods for the top-N recommendation task in terms of HR and ARHR in comparison to the other

Table 7.1: Effect of Cluster Refinement (HR). Dataset ClustMF NCR ClustMF

ML100K 0.3001 0.3150

Netflix 0.1923 0.1984

Yahoo 0.0780 0.0802

state-of-the-art methods (Section 4.4). For all the results presented, the number of top-N items chosen is 10 (i.e., N = 10). The presented results in Table 7.2 show that ClustMF performs better than the rest of the methods across all the datasets. For ML100K dataset, the performance improvement is quite substantial.

7.4

Conclusion

In this thesis chapter, we proposed a new method called ClustMF which combines the benefits of neighborhood models and latent factor models for top-N recommendation task. The user preferences and item characteristics were modeled using latent vectors, similar to the standard latent factors model. To bring in the benefits of the neighborhood based models, the user and item biases were modeled at the item and user clusters level respectively. Thus, for an item to be ranked higher in the top-N list, along with the dot product of the corresponding user and latent factors giving a high score, the corresponding user and item cluster biases also play a role. That is, the item must be similar to the other items rated by the user and the user must be similar to the other users who have rated the item. A comprehensive set of experiments on multiple datasets showed that, the proposed method outperforms rest of the state-of-the-art methods for top-N recommendation task.

Table 7.2: Comparison of performance of Top-N recommendation algorithms with ClustMF Method ML100K Params HR ARHR UserKNN 50 - - - 0.2704 0.0957 ItemKNN 20 - - - 0.2778 0.1037 PureSVD 10 - - - 0.2888 0.1298 BPRMF 32 0.01 0.001 - 0.2916 0.1287 ClustMF 64 10/10 0.01/2.0 0.005 0.3150 0.1317 Method Netflix Params HR ARHR UserKNN 50 - - - 0.1360 0.0434 ItemKNN 20 - - - 0.1151 0.0328 PureSVD 50 - - - 0.1845 0.0796 BPRMF 400 0.01 0.001 - 0.1890 0.0872 ClustMF 320 100/50 0.001/2.0 0.001 0.1984 0.0896 Method Yahoo Params HR ARHR UserKNN 200 - - - 0.0627 0.0219 ItemKNN 300 - - - 0.0633 0.0225 PureSVD 100 - - - 0.0774 0.0310 BPRMF 160 0.01 0.001 - 0.0760 0.0288 ClustMF 160 100/100 1e-4/0.001 0.005 0.0802 0.0310

Columns corresponding to “params” indicate the model parameters for the corresponding method. For UserKNN and ItemKNN methods, the parameter is the number of neighbors. For PureSVD method, the parameter is the number of latent factors. For BPRMF method, the parameters are the number of latent factors used, regularization constant and the learning rate. For ClustMF method, the parameters correspond to the number of latent factors, number of user/item clusters, regulariza- tion constant and learning rate. Underlined numbers represent the best performing model measured in terms of HR for each dataset.

Chapter 8

Conclusion

8.1

Thesis Summary

Recommender Systems are prevalent and are widely used in many applications. In particular, in the recent years recommender systems have gained popularity via their usage in e-commerce applications to recommend items so as to help the users in iden- tifying the items that best fit their personal tastes. Computationally, recommender systems represent a set of methods that produce recommendations of relevant items (songs, movies, books etc.,) to the users based on the preferences captured from their consumption history. There are two main tasks associated with recommender system. The first is to predict the preference for a given user-item pair in the form of a numer- ical rating. This is known as rating prediction problem. The second is to generate a ranked list of items for users that best suits their personal preferences. This is known as top-N recommendation problem. These problems are typically solved using one of the two classes of methods. First class of methods are called content based methods, and they build models based on the intrinsic properties associated with the users and the items. The second class of methods called collaborative filtering utilizes the preference information available in the form of explicit ratings or implicit feedback. Specifically these models utilize the user/item co-rating information to learn relationships between the users and the items. At a high level, collaborative filtering based approaches can be further classified into two classes. The first class of methods known as neighborhood based methods, explicitly compute/learn the user and/or item neighborhood using the

co-rating data and then uses these neighborhoods to compute the recommendations. In the second class of methods, known as model based approaches learns an explicit model from the data and this model is then used to generate the recommendations. In this thesis, we have presented three different approaches that addresses the rating prediction and top-N recommendation tasks for recommender systems.

Factored Item Similarities Method for top-N Recommendation

The existing state-of-the-art top-N recommendation algorithms suffer from data sparsity issue associated with the user-item preference data and thus fails to capture meaningful relations between users who do not have enough co-rated items. In this thesis to ad- dress this, we presented a factored item similarity based method (FISM) for the top-N recommendation problem. FISM learns the item similarities as the product of two ma- trices, allowing it to generate high quality recommendations even on sparse datasets. The factored representation is estimated using a structural equation modeling approach, which leads to better estimators as the number of factors increases. We conducted a comprehensive set of experiments on multiple datasets at different sparsity levels and compared FISM’s performance against that of other state-of-the-art top-N recommenda- tion algorithms. The results showed that FISM outperforms the rest of the methods and the performance gaps increases as the datasets become sparser. For faster recommen- dation, we showed that sparsity can be induced in the resulting item similarity matrix with minimal reduction in the recommendation quality.

Modeling Global and Local Preferences of Users in Collaborative Fil- tering

A novel user modeling approach for collaborative filtering called MPCF that is applicable to both rating prediction and top-N recommendation tasks is presented in this thesis. MPCF models the users as a combination of global preference and local preferences components to better capture the high level and finer preferences of users corresponding to their different preferences for item features. The recommendation score is computed as a sum of the scores from the components representing global preference and local preferences. We presented two different approaches to model the different use cases

which affect the user preferences at a finer level. The first approach (MPCFi) models the strong preferences of users corresponding to a subset of item features along with a set of local preferences that correspond to the remaining set of item features on which the users are willing to make a tradeoff, whereas the second approach (MPCFs) models the users preferences on all items as part of the global preference component and each of the local preferences correspond to tradeoffs that the users are willing to take on the item features. The results showed that the proposed methods outperform rest of the state-of- the-art methods in terms of both the rating prediction and the top-N recommendation performance.

Combined Neighborhood and Latent Factor Models for Top-N Recom- mender Systems

Neighborhood methods are good in capturing the localized neighborhood of users/items, thereby providing direct context to users for the recommended items while the Latent Factor models are good in capturing the overall global relations between the users and items. Thus, there is a need to combine these two class of methods to gain both the benefits associated with them. In this thesis we have developed a new method called ClustMF, which is designed to capture the benefits of both the neighborhood based models and the latent factor models. The benefits of latent factors models are utilized by modeling the users and items similar to the standard MF based methods and benefits of the neighborhood models are brought into the model, by introducing user and item biases at the item and user cluster level. The experimental evaluation showed that the proposed method outperforms rest of the state-of-the-art methods for top-N recommendation performance.

Related documents