College Papers

Neural built by inte-grating the traditional models and

Neural Network Approach Motivation:– In recent years, deep learnings revolutionary ad-vances in speech recognition, image analysis andnatural language processing have gained significantattention– Also demonstrate its effectiveness in coping withinformation retrieval and recommendation tasks.Applying deep learning techniques into recom-mender system has been gaining momentum dueto its state-of-the-art performances and high-qualityrecommendations– Though we have been using collaborative andgraph based models for years, the accuracyfor recommendations have not changed. Manycompanies are moving towards to deep learning forfurther enhancing their recommendation quality.Covington proposed neural net based algorithmfor video recommendation on YouTube.Shumpei presented a RNN based newsrecommender system for Yahoo News.– Above deep models have shown tremendrous im-provement over traditional models.Neural Collaborative filtering:– This is neural network architecture built by inte-grating the traditional models and deep learning.– We use SVD to generate the latent vectors and usethem as input for the feed forward neural networkwhich learns the weights and bias and predictsratings for each user item combinations– the equation of the neural network is given by :r ij = f (U T , s u , s i , V )where s i and s u are feature vectors for items andusers and U and T are latent vectors for user anditem respectivelyFollowing Pis the loss function using rmse:’ 2L = alli,j w ij (r ij ? r ij)w ij is the weights associated with user i and itemjV. R ESULTS AND A NALYSISThis section describes the results of the experimentsperformed using the various recommendation systems de-veloped.Each of the methods used have their own set of pros andcons which are described in the sections below.A. Content based Filtering ResultsPros and Cons of Content Based Systems:Pro:Although we concluded that Collaborative systems are farbetter recommenders, CF isn’t always a viable option. If acustomer without an account or a new customer is viewing aproduct details page, we do not have any information aboutthe user’s likes and dislikes, we cannot build a matrix ofpurchases. What we can do is recommend similar productsusing content-based systems. This helps solve the “cold-start” problem that prevails in collaborative filtering methods.Cons:The aim of a recommender system is to recommend itemsthat drives incremental sale. On most e Commerce sites allthat a customer needs to do to get content based recommen-dations is to click the section of the website. For eg. its easyenough for folks to browse laptops in the laptop category.The motivation for recommendation systems is to driveincremental sales, i.e. sales that would not happen otherwise.If a customer has bought Harry Potter and the Philosopher’sstone, this recommender is going to recommend Harry Potterand the Chamber of Secrets. We can be almost certain thatthe customer already knew that there second Potter bookexists. Therefore incremental sales do not happen in thismethod of recommendations.B. Collaborative Filtering ResultsFig. 1: RMSE Values of Collaborative Filtering methods.•Figure shows the network architecture with latentand feature vectors as input and rating as output.•The bar graphs show us that the Jaccard similarity givesus the least RMSE value for all 3 error types i.e. Userbased, Item based and the Hybrid method.At each stage we decide which similarity method topick for reporting values based on the RMSE measure.•We pick the measure with the least similarity measurein each of the cases.Once the similarity method has been deduced in eachcase we can use the method to predict values forunknown data.•Popularity Bias: Lots of people give high ratings toa popular item. Collaborative filtering usually includesthis item in the neighbors that we find. So this item isgenerally recommended to the users.C. Singular value decomposition resultsSVD was performed on ml-100k data set with 90-10split and predicted ratings were compared against the actualusing RMSE.RMSE for k = 100 is 0.604014558278time taken to execute : 34.397260665893555 secondsPros and Cons of Singular value decomposition: Pro:••Fig. 2: Time taken for the different Collaborative Filteringmethods.Cons:•We notice that although the User-user based filteringmethods have the least precision, this method runs thefastest among the three.• The item-item based filtering takes up time in firstcoming up with the item-item matrix and finding asimilarity between the items. The item-item method isfar more accurate than the user-user method as itemsin general are simpler than users. Users with eclecticprove to be a pain point in recommending accurately.• The hybrid system combines both the user-user anditem-item methods to recommend items. The RMSEvalues seemed to improve as we provided more datato the system.Pros and Cons of Collaborative Filtering:Pro:• CF systems are more versatile, they can be appliedon any domain and with some care can provide crossdomain recommendations as well. CB recommendersrequire information about items, therefore they requiredomain modeling and extending the algorithm to differ-ent domains is problematic.• CF engines work well when the user space is large,as the user space proves to be the backbone of healgorithm.• CF engines are better at helping users escape from the”filter bubble problem”, as users draw bridges acrosssubspaces in the item space. CB filters provide rec-ommendation only within a item space, therefore thediversity by such systems is limited.Con:• CF does not work well for new users. This is oftenreferred to as the cold start problem• CF also suffers if new items are added to the catalogand have not been rated by any user.SVD was able to discover latent features which was notdirectly available in any of the above methodsAs SVD was able to decompose the matrix , we hadinherent found out the unknown ratings by just usingsingle valued vectors and latent vectors•••Time to decompose the matrix is pretty high and mem-ory intensiveNot all the latent features are useful and it might overfitthe current data.SVD can be used only if the dataset is fixed, If a newuser/item is added we might need to reconstruct thewhole matrix which is costly