Recommendation Systems - Science topic

Assuming that the user already put in a couple of query words, you could in a first step tryo to find out in what relationships these words are. For example, if the words are antonyms, you could suggest "what is the difference between * and *". You will need something like an ontology.

Regards,

Joachim

fair enough Mourad Raif

Thanks Babatounde Moctard Oloulade for the direction. But the main challenge lies in how we are defining the graph from both user and item point of view.

Ali Fallahi One of the most often used recommendation algorithms is collaborative filtering (CF) and its variants. Even inexperienced data scientists may use it to create their own personal movie recommender system, for example, for a CV project.

Recommender systems are used by Netflix, YouTube, Tinder, and Amazon. The systems attract users by making relevant suggestions depending on their decisions. Recommender systems may also improve the user experience of news websites.

I agree with Michael Frömmel that it is good to have this recommendation at the start of your contact with a new journal but I agree with Subrata Chakraborty and that you have to be very careful because some of those journals have ridiculous or exorbitant APC requirements.

Omar Tarek ,

I would first sit down and describe at a high level the features that are desired for each of the modules. Be sure to account for feature dependency between modules. After doing so, the stage for particular algorithm that targets the different features can be derived. You can contact me here or privately to continue the discussion.

Regards

Hi I am working with Simpsons Paradox and XAI. You explore it and if it is interesting for you, we can collaborate.

If you are interested in assessing the performance of modelling approaches, please let me recommend these works:

If your topic assumes the comparison of alternative approaches, the FEW-L1 workflow presented in the article above can be used.

Hi,

I have not notices it until now. That is a bit strange way to write a link.

Thanks for the answer, Hardy! Yes, please ask your colleague! The topic of generating recommendations for filling out data entry forms is interesting! Best regards, Benno

Dear Jamalla Zawia ,

The recommendation system is a very trending topic at present. It is basically an extension of semantic analysis.

You can follow these articles for better insight.

Hope it will help you.

With regards,

Sayan Surya Shaw.

Dear Muhammad Hassaan Farooq Butt ,

Regarding your question,

1. Recommender system is one of the hottest topics recently and good quality work in recommender system has a good opportunity to be published in high impact journals.

2. There are many hot topics in RS, such as using the user reviews to solve the two most famous problems of RS: data sparisty and cold start problem, you can check this recent survey and at the end of it there are some future trends,

Also, you can use deep learning techniques to enhance the recommender system accuracy/performance.

Good Luck

Use k- mean to cluster the people into groups. Set the number of groubs k to equal the number of plans you have. Finally, use these groups to build your system

1) If data is in the form of healthcare images (e.g. MRI results, X-rays, CT Scan Images etc.) then CNN is the best choice. See the following recent top-level experimental evidences:

i. According to a study published in June 2018 in “Annals of Oncology” [1] , a research work was carried out by a team of experts from German institutions. The goal was early detection of Melanoma (the most serious type of skin cancer) through dermoscopic images. They trained and tested a deep learning CNN for differentiating dermoscopic images of melanoma and benign marks. The team set out to compare its diagnostic performance to that of 58 dermatologists of various levels of expertise. The convolutional neural network (CNN) scored 10% higher in terms of specificity than human experts.

ii. According to an article of August 2018 [2], researchers at the Mount Sinai Icahn School of Medicine (USA) trained CNN for early detection of acute neurological events (like stroke, hemorrhage and hydrocephalus etc.). They used 37,236 CT scan images for its training and their algorithm detected disease in CT scan images in just 1.2 seconds that was 150 times faster than human radiologists.

Also, there are many more experimental studies in this context and you can see them if you want to work using deep learning in such healthcare data applications.

2) If the data is in the form of text (like medical reports or any other health care documentation) then those deep learning models will be more suitable that had already given good results in NLP (natural language processing) applications because such algorithms can understand the complex semantic meaning in the text. Such algorithms are like LSTM (which is in fact the special type of RNN) as these algorithms are capable of learning long-term dependencies and good in sequence prediction problems. Hence they are well suited for classifying, processing and making predictions regarding text analytics which demands understanding textual data based on the sentential context.

[1] https://healthitanalytics.com/news/deep-learning-tool-tops-dermatologists-in-melanoma-detection

[2] https://healthitanalytics.com/news/deep-learning-ids-neurological-scans-150-times-faster-than-humans

Shafagat Mahmudova

very useful, thanks

At its simplest, market segmentation is the process of dividing a market into segments and directing a marketing mix to a particular segment, rather than the total market. In line with your question, product segmentation has special relevance for the retail sector, where the range of the offering (products + services) is very wide. For a startup it is important to start with market segmentation or market niche.

Dear Sandhya,

Many models are indeed biased and you often get the same set recommendations over and over again due to your browse history in those systems. For example, consider that from September till May you watch horror movies, while from April till September, you watch comedies. Obviously, there is a change in your preferences but most models can't understand this evolution. Only a few research has been conducted in the literature related to this phenomenon that is known as "preference/temporal dynamics". It seems that in most online services this feature is not supported yet.

Or:

We still need basic data , which will come from crop-nutrient repsonse studies in field or the basic dataset in field need to be subjected to machine learning duly cross validated in order to set right the suitability for practical application in filed...

I would like to recommend you to go through two of my articles:

1. B.K.Tripathy and Sooraj, T. R.: An Interval Valued Fuzzy Soft set Based Optimization Algorithm for High Yielding Seed Selection, International Journal of Fuzzy Sets and Applications, IGI publications, vol.7, issue 2, (2018), pp. 44 - 61.

2. B.K.Tripathy, Sooraj, T. R.: Optimization of seed selection for higher product using Interval valued Hesitant Fuzzy Soft Sets, Songklanakarin Journal of Science and Technology (SJST), 40 (5), Sep. -Oct. 2018, (2018), pp.1125-1135 .

Hicham Chouikh Fertilizer will be recommended using nutrient status table stored in the database. By comparing values of nutrients with table classification will be done. And accordingly fertilizer will be recommended to the user. Our system will help farmers for better crop yield which in turn maximizes profit. I suggest you follow: http://www.informaticsjournals.com/index.php/mvpjes/article/download/18273/17610

Most common techniques used for signature matrix are jaccard similarity and cosine similarity. I would prefer to use cosine similarity for given circumstances.

I think springer and elsevier are the best publisher to find the suitable journal .

The field of recommender systems, especially in academia, has been focusing mainly on the accuracy-related measures of successful recommendation. Many people have worked on things like how to improve accuracy of the recommendations by a tiny bit while the real effect of this improvement is really not clear. However, the real-world industrial recommender systems have proved that this over-emphasis on accuracy is wrong and there are other measures that have a better correlation with the success of a business. Things like diversity, novelty, long-tail recommendations, and removing biases. In addition, the new notion of multi-stakeholder recommendation has opened a lot of doors for further research. For example, taking into account different stakeholders' needs of a given recommender system is critical for the long-term success of the business. Think about Spotify, for instance. Spotify's success depends on the satisfaction of both the listeners and the artists. Therefore focusing only on the listeners may not necessarily lead to long-term success. To learn about the research directions in Multistakeholder recommendation I suggest our paper at UMUAI journal which can be a great read for someone who is looking for novel ideas.

https://link.springer.com/article/10.1007/s11257-019-09256-1#:~:text=Definition%202,over%20the%20full%20user%20population.

The start may be a bit challenging, however., if you really need to implement it and keep using in your tasks it will be a fun, see some sources e.g., https://github.com/awarebayes/RecNN

At PUC Chile we developed a python software package, pyreclab, to quickly prototype and develop recommendation systems. It might help you to move forward using Colab notebooks to test traditional models:

If you have further questions let me know.

There are third pary journal recommendation systems like JANE (Journal/ Author Name Estimator), Edanz, C&K recommender (which was developed for a reaserch degree), Manuscript Matcher, and so on. These systems vary from one to another based on the similarity algorithms used, subject domains, journal databases used, recommendation methods (i.e. content-based, knowledge-based, hybrid, etc.) employed, and so on. You can find more systems and a comprehensive comparison among them from the following link:

https://edoc.hu-berlin.de/handle/18452/20889

Regarding conferences, there are a few services which I will recommend to check:

I would recommend to also look into Boolean matrix factorization (BMF). k-means/modes is only usable if clusters have convex shapes and every point belongs to exactly one cluster. BMF can compute clusters with overlap (might not be necessarily important for your application) but it also identifies the features which are important for creating one cluster together with the data points belonging to one cluster. Especially if you have a high-dimensional feature space, it is unlikely that the similarity between points is expressed over the whole feature space. BMF identifies the feature space in which the points form a cluster.

If you are interested, look at Chapter 2 (particularly Section 2.4) from

There is a possibility that you develop your own algorithm. What do you need for it?

* sufficiently large database of annotated facial expressions

* when you have none, you can build your own from a video recording of people changing facial expressions

* you will extract features, positions of special points at the face -- this you know how to do better than me

* you use AI and ML techniques to distill changes in features that correspond with given emotional states

* you can go for static evaluation of features

* or you can go to dynamic evaluation

It all sounds like a really interesting project. What I can recommend to you is to learn how biosignals are processed using complex systems and AI & ML techniques. It can give you a big kick in hoe to proceed in your case.

All the best :-)

A determinant of organizational performance is the ability to leverage expert knowledge: much of that knowledge is tacit and therefore difficult to capture, codify, and make available through search engines and database technologies. And so, when looking we usually turn to people we know for quick, reliable information (and chance conversations can help too.) However, in the connected economy, personal networks are no longer sufficiently diverse to identify all the right persons, much as reliance on random connections is a thing of the past. Staff directories are no longer adequate to the task: therefore, learning organizations thrive on rich and fluid linkages and use expertise location capabilities to put people in contact with one another.

@Raju_Balakrishnan3

Thank you so much.

The accuracy depends on the source of the data and the target objetive.

Most of cases, kNN is better in sets of data with low missing date proportion. And SVD is better in huge size data sets.

Daniel,

You have asked a very important question that I have been asking myself in terms of the perceived goals of RG. My take is that RG is not a site to post trivial questions or answers. Questions and answers should follow some rigor in formulation. In my view it is a site to post information to clarify and further the areas of knowledge.

Some people take to RG as if this was Facebook and I think it trivializes the purpose of the site.

Regards

Pavlos Kefalas Dear, I using MovieLens Data for experimentation in this regard, I cannot use the time taken to read the review, can you suggest any other additional information that can be incorporated that can be used as intention?

Hello,

I see that Marijn Koolen is a ResearchGate member; might it be worth asking him for his advice?

Very best wishes,

Mary

Hi Tareq,

my suggestion is to read the chapter about recommender systems in "Mining Massive Datasets" http://www.mmds.org/. There you can get the idea for which types of problems the two basic approaches are good for and that often you need an hybrid approach.

Adjacency as the reviewer pointed is the most key factor and relationship

You can conduct time-series analyses with ARIMA-GARCH models and use residuals or standardized residuals to examine any outlier effect. Also, you can examine change/break-points in these residual series with several statistical techniques. A proper ARIMA-GARCH model should provide you normally distributed residuals with no serial correlation.

Suppose there are three synonyms i1, i2, and i3 of the same product. The simplest solution is to use dictionary to merge them as item i4. However, if user u1 rates on i1 and i2 by values v1 and v2 then, rating value of u1 on i4 is the average value (v1 + v2) / 2.

Hello,

Recommender Systems could be applied for many large-scale e-applications like e-commerce, e-learning, etc.

In fact, Recommender systems are being more powerful

with the huge volume of customer data in existing corporate

databases (structured Big Data), and the

increasing volume of customer data available on the

Web and social media platforms (unstructured Big Data).

You can check these links for more details:

- https://dl.acm.org/citation.cfm?id=2792842

- https://ieeexplore.ieee.org/document/7381557/

-

Dear Ibrahim,

Comparison is always acceptable in view of justifying your claim. If it can be shown, it is well taken.

I think I can understand the problem.

Generally, in personalized movie recommendation system, we use either content-based collaborative filtering or item-to-item collaborative filtering, or both in hybrid two-layered recommendation engine. In brief, the content-based collaborative filtering provides recommendation based on the genre of that particular movie/es you watched or rated. That means, if you initially watched horror movie/es, you will continually be recommended the horror movies. On the other hand, item-to-item collaborative filtering provides recommendation based on the similarity of your movie watching pattern with the other users' watching/rating pattern.

I think the following link can help you with the necessary coding.

https://muffynomster.wordpress.com/2015/06/07/building-a-movie-recommendation-engine-with-r/

thanks Nikola S. Nikolov i will read it and give you my feedback, thanks

You can group them in terms of similarity in use or functions for the purpose of comparison.

Thanks Nidhi for your answer.

could you send to me the dataset, because of it is for a data mining assignment. If you still keep the related dataset. thanks a lot！

The Nikon Ti-E epifluorescence setup is very flexible, you can very easily programme different time courses, and it can reliably maintain focus and xy-position over extensive time periods. We have used it for continuous neuronal live cell imaging for up to five days.

To evaluate a recommender system, you require some ground truth. Split your dataset into say 80% training set and 20 % test set. Use the 20% for test. Often times, people design a confusion matrix and calculates metrics such as precision, recall, F-score and plot ROC. study each of those to know which is suitable to you.

Thanks Pavlos and Carla !

Dear Pengtao,

Another dataset match the description can be found in the following link:

It depends on the recommender system that you needs but  you can visits  eBOOK at boonbook for more informations

I suggest you take a look at MyMediaLite (http://mymedialite.net/). It's a library that contains a set of recommender algorithms (including UserKNN and ItemKNN).

from the question it appears that you know the solution. but you are asking that it is taking lot of time for computation. 

my answer is -

yes it takes time because you are trying to segregate the objects and creating lots of datasets. it is obvious. 

any machine learning tool is OK. but the latest ones that i experienced are 

SVM. 

your idea has been implemented in the following paper. 

here in the paper i have taken a face and divided the face in to different parts. but i used the same SVM tool as i used before dividing the face. 

you may get an idea if you look at it. 

Hi Mohammed,

I can refer following works that combines linked open data and recommendation systems.

Start by reading Tommaso Di Noia papers; he has an important work on the subject. Here papers that you can start with:

* Recommender Systems Meet Linked Open Data.

* Recommender Systems and Linked Open Data.

* Content-Based Recommendations via DBpedia and Freebase: A Case Study in the Music Domain. 

Take a look on his papers here:

You can also take a look on these papers:

[1] R. Meymandpour and J. G. Davis, “Enhancing Recommender Systems Using Linked Open Data-Based Semantic Analysis of Items,”

[2] A. Passant, “dbrec - Music Recommendations Using DBpedia,”

So . . . . your problem is your data schema . . . . the max quantity of 1 effectively claims that everything is unique which makes the concept of comparison nonsensical and  literally impossible.

I'm not quite sure *why* you have a table of items (Amazon certainly doesn't) but that is not what you should be running your recommender off of.  You should be running it off of an itemTYPE table (which would be a foreign key for your items table if such is truly necessary) -- which is pretty much what Behzad suggested.

There are multiple datasets on web which someone can use during evaluation step. Depending on you model and the auxiliary information used (tags, timestamps, ratings, etc.) you should choose the best dataset close to your needs. The most used datasets in literature can be found here:

You can follow some base authors in this field

1. http://sisinflab.poliba.it/ostuni/index.php?blockTitle=Publications&page=publications&author=Ostuni

Jena API would help you for writing a java code while interacting the RDF data through SPARQL querying.

Devin: I think behavioral data could be crucial because it shows more insight about the patient than analyzing medical data. I read one interesting case study about leading healthcare provider. They were finding real pain points behind increased readmission. After analysis of unstructured and structured data, they found that leading causes of readmission are behavioral such as loneliness, smoking, no exercise, etc.

If you can provide me some sample data, then I would like to add them into the predictive model that I am building to predict diabetes in advance.

I think you need to dig into graph matching problems

Weka works with these metrics: support, confidende, lift, leverage and conviction.

The default output is confidence, take a look:

Best rules found:

The biggest music database is The Echo Nest (http://developer.echonest.com/docs/v4).

Here you have an API and you are able to request all the metadata for a special song (e.g. beats per minute, genre, duration, valence, energy, ...). Otherwise you can request songs with special attributes, like genre = pop or beats per minute > 150. All you have to do is to request a developer key, which is for free for research purposes.

Here are some example queries:

Here you can find a paper about The Echo Nest: http://ismir2011.ismir.net/papers/OS6-1.pdf

Regards

Patrick

There is a research paper in Procedia, Elsevier 2014 focusing on the app for "A Tourist guide with a recommender system and Social INteraction" This paper discusses the various challenges of recommender system and has performed various types of testing to evaluate the effects of proposed approach. Might be you can find something useful in it.

The approach used for third suggest is quite nice i must admit.

User emotional feedback which is based on LIKE /Dislike can be upgraded to include sentiment analysis of the comments written by user. Finding the sentiment out of comment can provide us with more information about the level of satification rather than like or dislike. Sentiment Analysis could be performed with the help of Deep Learning. This work has been undertaken by researchers at MIT. You can take their reference and embed the same in your work to optimise the results of recommendation. 

Hello Sapan, if you're still looking for it, you may want to take a look at these resources:

Best,

Danilo

I suggest to perform a search using the following keywords:

entity resolution classification

or

entity resolution clustering

If you are not really bound to Weka (and so to classification or a clustering approach) there are plenty of methods to deal with the problem, which to my understanding seems to be an entity resolution (i.e. disambiguation) issue.

Hope it helps :)

You can try these tool-kits:

Java based: Lenskit, Mahout, EasyRec, PREA and Rapid Miner

C/C++ based: SVD Feature, CofiRank, GraphLab, LibFM, LibMF, Waffles

Python based: Grab, Python-recsys

Others: MyMediaLite, Recommendable, Recommenderlab

You can search in Google for more details.

Best regards.

Collaborative Filtering is often employed in Point-of-Interest (POI) recommendation, but several issues must be tackled before using the traditional approach. Information about visited POIs by users is often partial (data sparsity) and usually the number of POIs is very large in comparison with the user ratings for each POI. Additional information about POIs is always welcome (category, tags, reviews, number of visits, etc). Luckily, tourist attractions such as museums, galleries, fountains, are quite "stable" temporally so it is easier to implement a CF approach. Personalisation needs information about user preferences, so it is crucial understand how this info can be collected in your scenario (through mobile devices, web sites, purchase histories). There are several papers online, you can have a look at this survey:

A Survey on Recommendations in Location-based Social Networks

if you are looking for I recommend dataset http://archive.ics.uci.edu/ml/

I think that shifting from recommendation to prediction may not be that trivial: generating predictions implies predicting how much a user will like/find useful a given item, while content-based recommendation is more of an information filtering affair.

By multiplying the cosine similarity and calling it a prediction you are assuming that the percieved quality of an item is solely based upon its similarity with other items. There is plenty of examples in which such an assumption won't hold: consider for instance Horror movies, most of them have the same plot, the same situations and often the very same actors and directors, however despite the high similarity you can still find good horror movies and crappy horror movies.

On the other hand, there exists other domains in which the potential interest can be represented by the sole similarity, for instance the domain of Scientific literature, Legal papers, Patents, and other technical documentation (when writing a paper you may cite papers that you consider crap, or a lawyer may find useful laws he/she just doesn't like), and therefore in which such shifting operation may make much more sense.

So, in my humble opinion, everything boils down at asking yourself what are you recommending: if your system recommends papers, well, the "naive" multiplication solution may just work fine, otherwise consider introducing collaborative filtering elements to predict ratings.

For a free solution in Python: http://scikit-learn.org/stable/modules/generated/sklearn.grid_search.GridSearchCV.html

example:

from sklearn.neighbors import KNeighborsClassifier

from sklearn.grid_search import GridSearchCV

from sklearn.metrics import accuracy_score, classification_report

# here must be some code for your training set and test set

tuned_parameters = [{'n_neighbors': [1,2,3],

'weights': ['distance', 'uniform'],

'algorithm': ['ball_tree', 'kd_tree', 'brute']}]

scores = ['precision', 'recall']

performances = [];

for score in scores:

print("# Tuning hyper-parameters for %s" % score)

print()

clf = GridSearchCV(KNeighborsClassifier(), tuned_parameters, cv=10, scoring=score)

clf.fit(X_train, y_train)

print("Best parameters set found on development set:")

print()

print(clf.best_estimator_)

print()

print("Grid scores on development set:")

print()

for params, mean_score, scores in clf.grid_scores_:

print("%0.3f (+/-%0.03f) for %r"

% (mean_score, scores.std() / 2, params))

print()

print("Detailed classification report:")

print()

print("The model is trained on the full development set.")

print("The scores are computed on the full evaluation set.")

print()

y_true, y_pred = y_test, clf.predict(X_test)

print(classification_report(y_true, y_pred))

print()

.

not exactly _content_ suggestion but quite close, see "dueling bandits" by Y. Yue and T. Joachims

http://www.yisongyue.com/publications/jcss2012_dueling_bandit.pdf

the point is that content recommendation is almost always contextual : users'profile information is often more important than the overall success rate of an item (otherwise, keep recommending blockbusters and you're fine !)

this brings you to "contextual bandits", "bandits with side information" and so on ...

by the way, googling _Reinforced Learning for content recommendation_ will lead you to quite a crowd of interesting publications !

https://www.google.fr/search?q=Reinforced+Learning+for+content+recommendation&ie=utf-8&oe=utf-8&rls=org.mozilla:fr:official&client=firefox-a&gfe_rd=ctrl&ei=ux4eU7nKMInf8gf-0YHwCg&gws_rd=cr#q=Reinforced+Learning+for+content+recommendation&rls=org.mozilla:fr:official&safe=off

.