<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Andrés Espinosa on Medium]]></title>
        <description><![CDATA[Stories by Andrés Espinosa on Medium]]></description>
        <link>https://medium.com/@andresespinosapc?source=rss-b6752d2c8611------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*rmA0dBBMgXmryv8X.jpg</url>
            <title>Stories by Andrés Espinosa on Medium</title>
            <link>https://medium.com/@andresespinosapc?source=rss-b6752d2c8611------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 15:04:48 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@andresespinosapc/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Haciendo rankings con BPR]]></title>
            <link>https://medium.com/@andresespinosapc/haciendo-rankings-con-bpr-a723112de1dc?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/a723112de1dc</guid>
            <category><![CDATA[spanish]]></category>
            <category><![CDATA[inteligencia-artificial]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[sistemas-recomendadores]]></category>
            <category><![CDATA[español]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Thu, 14 Dec 2017 09:01:01 GMT</pubDate>
            <atom:updated>2017-12-14T09:01:01.121Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FuG2iHJj41EDdIs5kDWOUQ.jpeg" /></figure><p>Muchas veces usamos <strong>modelos de Machine Learning</strong> sin entenderlos completamente, y terminamos usando parámetros predeterminados o los que usó otra persona en algún artículo que encontramos por ahí. Pero, aunque no queramos admitirlo, sabemos que sería mejor hacer un modelo que se ajuste mejor a nuestras necesidades.</p><p>En este artículo hablaré de un método que <strong>permite modificar un modelo</strong> ya existente para que funcione mejor para la tarea de hacer rankings, algo muy utilizado en sistemas recomendadores principalmente. Para esto voy a rescatar lo más importante a mi parecer del paper <a href="https://arxiv.org/pdf/1205.2618.pdf">BPR: Bayesian personalized ranking from implicit feedback</a>, así que tal vez hayan algunas partes un poco técnicas.</p><p>Primero, acá van algunas características importantes del método presentado en el paper:</p><ul><li>Usa como entrada <em>feedback</em> implícito, es decir, los usuarios no entregan una calificación explícita de los ítems, solo se infiere a partir de la interacción de los usuarios con la aplicación. Este se compone de una matriz usuario-ítem donde hay un 1 si el usuario ha interactuado con el ítem (<em>feedback </em>positivo) y un 0 en otro caso. Es importante notar que el 0 puede significar que no le interesa (<em>feedback </em>negativo) o que simplemente no lo ha visto (digámosle <em>feedback</em> faltante).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/386/1*B--VHK32HbbEeMG6DRndYw.png" /><figcaption>Matriz de <em>feedback implícito.</em></figcaption></figure><ul><li>Asumen que los usuarios prefieren los ítems positivos frente a los ítems con los que no ha interactuado. De esta forma, se construye un set de entranamiento formado por tuplas de la forma <em>(u, i, j)</em>, que representan que el usuario <em>u </em>prefiere el ítem <em>i </em>por sobre el ítem <em>j</em>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/372/1*HBNyf3mBoLo_eQdsFPfngw.png" /><figcaption>Transformación al set de entrenamiento utilizado.</figcaption></figure><ul><li>Lo anterior permite que haya información tanto positiva (ítem <em>i</em> es preferido sobre <em>j</em>), negativa (ítem <em>i </em>no es preferido frente a <em>j</em>) o incierta/faltante (ha interactuado con ambos ítems), mientras otros métodos tienen problemas diferenciando entre valores faltantes y valores negativos.</li><li>El set de entrenamiento es creado específicamente para el objetivo de hacer un ranking, que es una de las contribuciones más importantes de los autores en este paper.</li></ul><h3>BPR</h3><p>El método en sí consiste de dos partes principales:</p><ol><li><strong>BPR-Opt</strong>, que consiste en la función de pérdida a ser optimizada. Esta es formulada intentando maximizar la probabilidad o <em>likelihood</em> de que un usuario prefiera el ítem <em>i </em>sobre el ítem <em>j</em>.</li><li><strong>LearnBPR</strong>, que es el método utilizado para optimizarla eficientemente. Para este los autores simplemente utilizan el método del gradiente descendente estocástico pero tomando muestras aleatorias y con reemplazo.</li></ol><p>Los autores también mencionan una analogía entre su método y el método de optimización de área bajo la curva (AUC), la que pueden ver en la sección 4.1.1 de su <a href="https://arxiv.org/pdf/1205.2618.pdf">paper</a>.</p><h3>Aplicando BPR a modelos existentes</h3><p>Los autores prueban usando BPR con dos modelos existentes: Factorización Matricial (MF) y K-Vecinos más cercanos (KNN). Lo bueno de BPR es que, como es una función de pérdida, puede ser potencialmente aplicado a cualquier método que trate de minimizar una función de pérdida, incluyendo las famosas redes neuronales (aunque esto solo hace sentido si el objetivo es hacer un ranking).</p><p>No me voy a adentrar en MF y KNN, pero para los que conocen estos métodos es importante notar que: En el caso de MF, la optimización se usa para encontrar la aproximación de matrices que mejor se ajusta a los ratings a predecir, mientras que en KNN se usa para encontrar la mejor métrica de similaridad entre puntos.</p><p>La evaluación y comparación que voy a mostrar más adelante la hacen entre estos métodos, la modificación de ellos usando BPR y otros dos métodos existentes:</p><ul><li>WR-MF: Es un método para <em>feedback</em> implícito, pero la optimización está hecha utilizando de a una instancia en vez de a pares como se hace en BPR. Además, su optimización se adecúa mejor para problemas de regresión cuando debería serlo para clasificación (para la tarea de predicción de ítems).</li><li>MMMF: El problema con este método es que es específico para MF y esta diseñado para trabajar con datos muy dispersos.</li></ul><h3>Evaluación y conclusión</h3><p>Los métodos derivados de MF y KNN, a los que llamaron BPR-MF y BPR-KNN, fueron comparados con los otros métodos del estado del arte usando AUC como métrica de evaluación, aplicándolos a dos datasets distintos.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bn6fyLodNFvV6h2WXnvKkg.png" /><figcaption>AUC para los métodos evaluados.</figcaption></figure><p>Como se puede ver, BPR superó a todos los otros métodos. Algo que rescato del paper es fácil que puede ser integrarlo con otros métodos. Dada la modularidad de muchos de los métodos de recomendación existentes, es posible combinar métodos de distintos papers. En este caso, los autores trabajaron solo en una nueva función de pérdida en vez de un modelo completamente nuevo, pero las funciones de pérdida son usadas en muchos modelos ya existentes.</p><p>Finalmente, vuelvo a lo que dije al principio de este artículo. Citando a los autores: “La calidad de la predicción no solo depende del modelo sino también en gran parte del criterio de optimización”. A lo que quiero llegar es que, no siempre es bueno usar un modelo con sus valores predefinidos. Entenderlo y modificarlo para ajustarse a tus necesidades puede ser una muy buena idea.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2Fde58d5%3Fas_embed%3Dtrue&amp;url=https%3A%2F%2Fupscri.be%2Fde58d5&amp;image=https%3A%2F%2Fe.enpose.co%2F%3Fkey%3DdRXnS9Gplk%26w%3D700%26h%3D425%26url%3Dhttps%253A%252F%252Fupscri.be%252Fde58d5%252F%253Fenpose&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/f1342d9242d958b58363dfe521f61b02/href">https://medium.com/media/f1342d9242d958b58363dfe521f61b02/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MMX82VX3T3Ts2Sae6-toyw.gif" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a723112de1dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Today I’m writing about Slope One Predictors for Online Rating-Based Collaborative Filtering by…]]></title>
            <link>https://medium.com/@andresespinosapc/today-im-writing-about-slope-one-predictors-for-online-rating-based-collaborative-filtering-by-3c06677f75c6?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c06677f75c6</guid>
            <category><![CDATA[collaborative-filtering]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Sun, 03 Dec 2017 00:55:49 GMT</pubDate>
            <atom:updated>2017-12-03T00:55:49.090Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m writing about <a href="http://epubs.siam.org/doi/abs/10.1137/1.9781611972757.43">Slope One Predictors for Online Rating-Based Collaborative Filtering</a> by Daniel Lemire and Anna Maclachlan.</p><p>The purpose of this paper is to show that Slope One schemes have some desirable characteristics while getting a competitive accuracy compared to other methods for item rating prediction. The paper is from 2008 so it’s a bit outdated, but their methodology could work as inspiration for many. I will paste here these desirable characteristics I mentioned:</p><ol><li>Easy to implement and maintain: All aggregated data<br>should be easily interpreted by the average engineer and<br>algorithms should be easy to implement and test.</li><li>Updateable on the fly: The addition of a new rating<br>should change all predictions instantaneously.</li><li>Efficient at query time: Queries should be fast, possibly<br>at the expense of storage.</li><li>Expect little from first visitors: Auser with few ratings<br>should receive valid recommendations.</li><li>Accurate within reason: The schemes should be competitive with the most accurate schemes, but a minor gain in accuracy is not always worth a major sacrifice in simplicity or scalability.</li></ol><h3>Baselines</h3><p>They compare their approach with other four methods. I will explain the first two which are simpler, the other two are better explained in the paper itself:</p><ul><li>Per User Average: Predict that a user will rate everything according to that user’s average rating.</li><li>Bias From Mean / Non Personalized: The prediction is based on the user’s average plus the average deviation from the user mean for the item in question across all users in the training set.</li><li>Adjusted Cosine Item-Based.</li><li>Pearson.</li></ul><h3>The authors’ method</h3><p>Their first approach takes into account “both information from other users who rated the same item (like the Adjusted Cosine Item-Based) and from the other items rated by the same user (like the Per User Average)”, and they finally come up to a formula that only depends on the user’s average rating and which items has the user rated:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/388/1*FuCJWbbybRq5kiRXubpWqQ.png" /><figcaption>The Slope One scheme.</figcaption></figure><p>In this formula, <em>u</em> with a bar is the user average rating, <em>j</em> is the item to predict, <em>R </em>sub <em>j</em> are all the entries that don’t have item <em>j, card(S) </em>considering any set <em>S</em> is the number of entries in it, and <em>dev</em> is the average deviation of item <em>i </em>with respect to item <em>j</em> as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/340/1*tzxlRj6hKToUF8sWKVRNlA.png" /><figcaption>Average deviation.</figcaption></figure><p>Here’s a new term <em>S </em>sub <em>i,j </em>, which simply are all the entries with contain <em>i </em>and<em> </em>j.</p><p>In their second approach they include information of the number of ratings observed and name their approach Weighed Slope One. Here’s the new formula:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/541/1*uR3jmITYIVW5fSUE1phDtA.png" /><figcaption>Weighted Slope One scheme.</figcaption></figure><p>Here the new terms are: <em>X</em> which are all the evaluations in the training set and <em>S(u) </em>which are all the entries for the user <em>u.</em></p><p>Finally, they divide the training set into liked and disliked items, depending on if the item has a rating smaller or greater than the user’s average rating, and the formula is a weighted approach similar to the Weighted Slope One but considering the liked and disliked items separately weighted. They name it Bi-polar Slope One and here’s its formula:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/603/1*GXBAx5Vlf8OMu8ylye_WwQ.png" /><figcaption>Bi-Polar Slope One scheme.</figcaption></figure><p>Where <em>p</em> is just the same used in Weighted Slope One:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/198/1*oy6VngcMuiOthRBgkvRqQg.png" /></figure><h3>Evaluation</h3><p>They use All But One Mean Average Error as evaluation metric, for which they try to predict one item having the information of the rest and repeating this for all the items. Here’s what they found:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*6Ormi8TlI6v5uJZkiFy7Yw.png" /><figcaption>Results.</figcaption></figure><p>So their objective is completed. They achieved competitive results with a method that meet their requirements. With their Bi-Polar Slope One approach they event obtain the best results.</p><h3>What do I think?</h3><p>I found this paper very straightforward. First, it has a well defined and simple objective, mentioning all the characteristics that their method should have. Also, the explanations of related work and baselines are very good and compact. Then, the methodology used is very well explained, it would be easy for someone to replicate the experiments and understand why every step is for. And at the end, they return to their objective and show that it was accomplished.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c06677f75c6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[POI Location-Based Recommendations]]></title>
            <link>https://medium.com/@andresespinosapc/poi-location-based-recommendations-da128a6bab51?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/da128a6bab51</guid>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[point-of-interest]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Mon, 20 Nov 2017 05:43:43 GMT</pubDate>
            <atom:updated>2017-11-20T05:43:43.530Z</atom:updated>
            <content:encoded><![CDATA[<p>In this post I’m going to talk about <a href="https://www.aaai.org/ocs/index.php/WS/AAAIW15/paper/viewFile/10132/10253">A Survey of Point-of-Interest Recommendation in Location-Based Social Networks</a> by Yonghong Yu and Xingguo Chen.</p><p>This is a short paper that goes through a review of relevant work done in the field of point-of-interest (POI) or location-based recommendation systems and also tries to introduce some research directions regarding these systems. This type of recommendations have become more popular nowadays, in part because they “help LBSNs [location-based social networks] to increase revenues by providing users with intelligent location services, such as location-aware advertisements”.</p><p>They divide POI recommendation systems into four categories. They explain not only the best results but the results that have been published over time, and give a brief but good explanation of each one. Now I’ll present what I found more interesting or important about them.</p><p><strong>Pure check-in data</strong></p><p>It only takes into account what POIs have the user visited, and sometimes also the frequency. A user-POI similarity is defined and conventional collaborative filtering approaches are used, such as used-based, item-based, matrix factorization. There’s also a work that uses a method named probabilistic factor models (PFM). It was found that the frequency of the check-ins only deteriorate the results, son only binary check-in information is used.</p><p>There are only a few studies of this topic and generally it is better to also use additional information from other methods.</p><p><strong>Geographical influenced enhanced</strong></p><p>It takes into account the location of the POIs. The authors of the paper mention a lot of methods, which I recommend to look directly into the paper. Basically, these methods find a probability distribution that models the visits of places based on the distance between user-POI or POI-POI. For example, one of the methods applies a kernel density estimation (KDE) to model this distribution, being personalized for each user.</p><p>The methods in this category generally overcome pure check-in data methods.</p><p><strong>Social influence enhanced</strong></p><p>They “assume that friends of LBSNs share much more common interests than non-friends”. They only consider a binary matrix where it is represented only if a user has a social relation with other or not. Social influence is far less informative than pure check-in and geographical information, so it doesn’t take much importance in this kind of recommendation systems. This could be explained because the need of physical interactions between users and locations, unlike traditional recommendation systems that are accessible trough the web.</p><p><strong>Temporal influence enhanced</strong></p><p>They ”assume that users’ interests vary with time and users’ visiting behaviors are often influenced by time since users visit different places at different time in a day”. The difference with traditional recommendation systems that use this information is that they make recommendations for a specific temporal state, while traditional systems simply take away importance of too old information.</p><h3>Conclusion</h3><ul><li>POI-based recommendation systems are different than traditional ones.</li><li>Social influence is the less informative for this task. Geographical information is the most informative.</li><li>Personalized distance distributions per user is better than universal distributions.</li></ul><p>The authors present more conclusions, but these are the ones I found more relevant. Also some are only make sense if you read the whole paper.</p><h3>Research directions</h3><ul><li>Use check-in frequency data, not only binary check-ins.</li><li>Use rank-based collaborative filtering approaches (I talked about one of these techniques in my last post).</li><li>Consider different types of social relations, not only if the user has a social relation with another or not.</li><li>Use information from other social networks.</li><li>Design of a unified framework to join all kinds of information from the different categories.</li><li>More efficient methods, e.g. parallelized computing. “POI recommender systems suffer seriously from scalability problem”.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da128a6bab51" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learning to Rank: BPR]]></title>
            <link>https://medium.com/@andresespinosapc/learning-to-rank-bpr-5fe5561d48e0?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/5fe5561d48e0</guid>
            <category><![CDATA[loss-function]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Mon, 13 Nov 2017 05:43:36 GMT</pubDate>
            <atom:updated>2017-12-07T21:17:46.354Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m talking about <a href="https://arxiv.org/pdf/1205.2618.pdf">BPR: Bayesian personalized ranking from implicit feedback</a><em> </em>by Steffen Rendle, Christoph Freudenthaler, Zeno Gantner and Lars Schmidt-Thieme.</p><p>They start mentioning a bunch of existing methods for the task of item recommendation. Nevertheless, in words of the authors “none of these methods directly optimize its model parameters for ranking. Instead they optimize to predict if an item is selected by a user or not”. There are some methods that consider the optimization for ranking, but they are non-personalized, which means that they give the same recommendation for all users.</p><p>Some important things to consider about the authors’ method:</p><ul><li>They assume that the user prefers the positive item over all other non-observed items.</li><li>The training data is constructed by tuples in the form (u, i, j), which represent that user <em>u </em>prefers item <em>i </em>over <em>j</em>.</li><li>The above allows the usage of both positive and missing values, while other methods have trouble taking the missing values as negative ones.</li><li>Another thing to notice is that the training data is created for the actual objective of ranking, which is one of if not the most important contributions of the authors’ method.</li></ul><h3>BPR</h3><p>Then they explain the BPR method itself. They start with the BPR-OPT formula, which is the one that needs to be optimized. It is formulated by trying to maximize the likelihood that a user prefers item <em>i </em>to item <em>j</em>. After that the authors mention an analogy between their method and the AUC optimization method. And finally they explain the Learn-BPR method, which is used to optimize BPR-OPT efficiently. They simply use the stochastic gradient method but taking the samples randomly and with replacement.</p><h3>Application</h3><p>The authors use BPR with two different existing methods: MF and KNN. It’s worth noting that because BPR is a loss function and optimization method, it can potentially be applied to any method that tries to minimize a loss function (although it only makes sense for ranking methods). The MF method (before BPR) uses least square or other methods to find the best approximation of the predicted X to the real X value, where X is the rating of the items that would be used to rank them. In the same way, the KNN method uses a similarity measure C that can be learnt. In both cases, the authors use BPR instead of the loss function used in each method. They also have to add some model specific regularization parameters.</p><h3>Other methods</h3><p>Two other methods are described, with their corresponding disadvantages in regard of BPR.</p><ul><li>WR-MF: It is a method for implicit feedback, but the “optimization is on instance level (one item) instead of pair level (two items) as BPR”. I have already mentioned that this is one of the major advantages of BPR. Also, their optimization is best suited for regression, but it should be for classification (for the task of item prediction).</li><li>MMMF: It’s disadvantages are that it is specific to MF and is designed to work with sparse data.</li></ul><h3>Evaluation and conclusion</h3><p>The method was evaluated and compared to state-of-the-art methods using AUC as the metric for evaluation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bn6fyLodNFvV6h2WXnvKkg.png" /><figcaption>Figure 1. AUC for the evaluated methods.</figcaption></figure><p>As you can see in <em>Figure 1</em>, BPR outperformed every single method. So the biggest conclusion they make is that “the prediction quality does not only depend on the model but also largely on the optimization criterion. Both our theoretical and empirical results indicate that the BPR optimization method is the right choice for the important task of personalized ranking”.</p><h3>What do I say?</h3><p>I liked the paper a lot. It is easy to read and very well explained. Also, the method itself had a big impact in item-based recommendations. Due to the modularity of many of the existing recommendation methods, it is possible to combine works from different papers. In this case, the authors worked just in a new loss function instead of a whole model, but loss functions are used in many existing models. In my opinion, this makes BPR a very clean and useful method embedded in a great paper.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5fe5561d48e0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Link-Prediction Problem for Social Networks]]></title>
            <link>https://medium.com/@andresespinosapc/the-link-prediction-problem-for-social-networks-362e8355e555?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/362e8355e555</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[social-network]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[network]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Mon, 30 Oct 2017 06:10:28 GMT</pubDate>
            <atom:updated>2017-10-30T06:10:28.743Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m writing about <a href="http://onlinelibrary.wiley.com/doi/10.1002/asi.20591/full">The Link-Prediction Problem for Social Networks</a> by David Liben-Nowell and Jon Kleinberg.</p><p>In the paper, the authors try to contribute in the prediction of links in a social network for a given future time considering the network at a current time. They want to do that based only on the topology of the network, without considering additional information.</p><p>After the introduction, in the Data and Experimental Setup section, they describe the dataset they will be using, how will they split the training and testing sets (1994 to 1996 and 1997 to 1999, respectively) and the evaluation metric they will be using.</p><p>They then present the existing methods related with the attacked problem. In general, they are “adapted from techniques used in graph theory and in social-network analysis”. They mention three major types:</p><ul><li><strong>Based on node neighborhoods:</strong> “Two nodes x and y are more likely to form a link in the future if their sets of neighbors T(x) and T(y) have large overlap”.</li><li><strong>Based on the ensemble of all paths:</strong> Consider all the paths between two nodes.</li><li><strong>Higher level approaches:</strong> Approaches that can be combined with the previous mentioned methods. They mention three: Low-rank approximation, unseen bigrams and clustering.</li></ul><p>In the Results and Discussion they mention a lot of potential future work, describe the differences in performance between the different used datasets, and most importantly expose the limitations of their experiments. About the latter, they say that all of their experiments “were carried out on social networks formed via the collaborations of physicists”. However, then they tried with some other datasets. In this last experiments they found that the performance versus the random method increases as the “topical focus of our dataset widens”.</p><p>Finally, they talk about some future directions that the investigation could take. First of all, they used a graph that has no much information, so the usage of a graph that considers articles related to the authors could give relevant information to be used by the algorithm. They also mention the possibility of considering differences in different periods of time, for example “treating more recent collaborations as more important than older ones”.</p><h3>What do I think?</h3><p>I think they did a great job explaining the existing methods that tries to solve this problem. The explanation for each method was complete and clear.</p><p>In my opinion, the evaluation comparison method could have been better. They just compare with random because there are no other method that uses the same information as they want to use. But I think they should have implemented a combination of the proposed method with an existing one and compared it with the existing one alone. I think it is not useful enough to know that the topology gives new information, but it is necessary to use this information to improve some existing method.</p><p>Out of that, it is a very complete and detailed paper, especially in the description of the existing methods and the Results and Discussion section.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=362e8355e555" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Factorization Machines]]></title>
            <link>https://medium.com/@andresespinosapc/factorization-machines-51ef1d0ea5a?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/51ef1d0ea5a</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[svm]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[factorization-machine]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Mon, 02 Oct 2017 04:15:25 GMT</pubDate>
            <atom:updated>2017-10-02T04:15:25.579Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m reviewing a great paper: <a href="http://ieeexplore.ieee.org/abstract/document/5694074/">Factorization Machines</a> by Steffen Rendle.</p><p>Basically, the paper consists of a description of a new technique named Factorization Machines and a comparison with other existent algorithms.</p><p>It starts by mentioning the FM advantages:</p><ul><li>Works for very sparse data.</li><li>Has linear complexity.</li><li>Is a general predictor (like SVM).</li></ul><p>Then, it explains how can it work with sparse data. The key ingredient of FMs is to model all interactions between variables using factorized parameters, i.e. to estimate fewer parameters in the model. Intuitively, it makes sense because with poor data it could be more difficult to estimate a lot of parameters.</p><p>They divide the FMs description into multiple parts. They explain the model itself, how they control its expressiveness, give a better explanation of the parameter estimation under sparsity, and explain how they turn the computation complexity into a linear one. This last part is very interesting because they turn a potential complexity of <em>kn^d</em> into linear.</p><p>They continue by naming some applications in which FMs can be used: regression, binary classification, and ranking. Lastly, they explain how the model is trained and how it can be generalized for <em>d</em> “degrees of freedom” (they don’t use this term).</p><p>The next part is a comparison with other methods. The first one mentioned is SVM. They explain the method and why it fails for very sparse problems. Finally, they show how other popular methods can be emulated just by applying a transformation to the input data, what talk about the generality that FMs have. I found this generality to be other of the most impressive things about this model.</p><h3>What do I say?</h3><p>I found this paper extremely interesting and well written. They give all the details of the implementation in an intuitive way. So it is simple and detailed enough, nothing else to ask.</p><p>The only thing I missed was the lack of explanation of how to construct the input data. For example, it is not clear to me why the users can’t be represented as an integer in the input data instead of a <a href="https://en.wikipedia.org/wiki/One-hot">one-hot encoding</a>. However, this is something you could easily check by your own when building the model.</p><p>In conclusion, it is a great paper and I recommend it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=51ef1d0ea5a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The basics of Context-Aware Recommendations]]></title>
            <link>https://medium.com/@andresespinosapc/the-basics-of-context-aware-recommendations-5dd7a939049b?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/5dd7a939049b</guid>
            <category><![CDATA[context-awareness]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[context]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Mon, 25 Sep 2017 03:46:20 GMT</pubDate>
            <atom:updated>2017-09-25T03:46:20.099Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m talking about <a href="https://link.springer.com/chapter/10.1007/978-0-387-85820-3_7">Context-Aware Recommender Systems</a> by Gediminas Adomavicius, Bamshad Mobasher, Francesco Ricci, and Alex Tuzhilin.</p><p>The paper is divided into two significant parts. The first one gives a theoretical framework of context-aware recommender systems. The second one gives multiple examples of applications that use them. I’ll make a summary of the theoretical part.</p><p>There are two different views for context:</p><ol><li>Representational: Observable attributes that are known <em>a priori</em>. The structure does not change over time.</li><li>Interactional: The scope of the features is defined dynamically. Cyclical relationship between context and activity (which makes the context change).</li></ol><p>The authors then classify around two different axes:</p><p><strong>Knowledge of a recommender system about the contextual factors:</strong></p><ul><li>Fully observable: The contextual factors are known explicitly.</li><li>Partially observable: Only some information is known explicitly.</li><li>Unobservable: No explicit information. Needs to be modeled using latent variables.</li></ul><p><strong>How contextual factors change over time:</strong></p><ul><li>Static: The factors and their structure don’t change over time. Can be addressed by machine learning methods.</li><li>Dynamic: The factors change in some way. For example, if new categories are created for one factor. They are more complicated than static factors because they involve changing data, what makes them difficult to evaluate.</li></ul><p>In my opinion, the most important of the above is that fully observable factors are easier to handle than unobservable ones, and static easier than dynamic. Also, I’d like to point out that there’s a strong relationship between <em>representational</em> views and <em>static</em> factors, and the same between <em>interactional</em> and <em>dynamic</em>.</p><p>Then, the authors present three different ways of including context into a traditional recommender system model, which are better explained by an image of the paper:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/932/1*xbnfpQ0g9bltyflGNW-74Q.png" /><figcaption>Figure 1. Paradigms for Incorporating Context in Recommender Systems</figcaption></figure><p>The letter ‘c’ indicates when in the model the context is included. The letter ‘u’ shows the same for the user data.</p><ul><li>Prefiltering: Context is used to select some set of data and then predict like a traditional recommender system. One way of doing that is splitting items or users by context as if they were different items or users.</li><li>Postfiltering: Ratings are predicted and then the results are filtered using the context. This can be implemented by ordering the results depending on the context or by just filtering the results.</li><li>Modeling: The context is used right in the model. It is more complex and could be implemented by multiple machine learning models like SVM, matrix factorization or a markov chain.</li></ul><p>Until now, everything has been more theoretical than anything. For me, that was the most important part of the paper and where I learned the most. The next part is about various examples of context-aware recommender systems. It is not too specific, but it gave me a notion of how they could be used. I’m not going into details because a summary of this part doesn’t worth it. In the end, I think developing context-aware recommender systems is too attached to each particular application, so the important thing is getting a notion of how are they used instead of learning specific methods that can be applied.</p><p>Finally, I want to add a simple idea from the conclusion of the paper: The possibility of making combinations of <em>prefiltering</em>, <em>postfiltering</em> and <em>contextual modeling</em>. It had not been mentioned before in the paper, and I think it’s good to take it into account.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5dd7a939049b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hybrid recommendation systems]]></title>
            <link>https://medium.com/@andresespinosapc/hybrid-recommendation-systems-b9a8d45596e8?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9a8d45596e8</guid>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <category><![CDATA[hybrid]]></category>
            <category><![CDATA[collaborative-filtering]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Wed, 20 Sep 2017 02:50:11 GMT</pubDate>
            <atom:updated>2017-09-20T02:50:11.511Z</atom:updated>
            <content:encoded><![CDATA[<p>This is a review of <a href="https://link.springer.com/article/10.1023%2FA%3A1021240730564?LI=true">Hybrid Recommender Systems: Survey and Experiments</a> by Robin Burke.</p><p>This paper is divided into two main parts. First, it describes the different types of recommendation systems and makes a very good comparison between them, I will go into details later. Second, it presents a new type of hybrid system, that is, a combination of two different techniques that hasn’t been researched yet.</p><h3>Existing hybrid methods</h3><p>I must say that the first part is a complete paper about all the big existing recommendation techniques, so I recommend reading at least that part.<br>Before describing the different existing recommendation techniques, the author defines three parts included in any of them. I think he wants us to have a notion of how they can be connected as a hybrid system:</p><ul><li>Background data: “The information that the system has before the recommendation process begins”. Can be seen as analogous to the training data in a Machine Learning classifier.</li><li>Input data: “The information that the user must communicate to the system in order to generate a recommendation”. I make an analogy with the input data in a classifier.</li><li>“An algorithm that combines background and input data to arrive at its suggestions”. It is like the classifier algorithm itself.</li></ul><p>Then, the author gives a good brief explanation of each existing method, including their strengths or benefits in practice, which is extremely important and have been omitted in other works. I will write down the most important ones in my opinion for each technique:</p><ul><li>Collaborative: “Completely independent of any machine-readable representation of the objects being recommended, and work well for complex objects such as music and movies where variations in taste are responsible for much of the variation in preferences.”</li><li>Demographic: “It may not require a history of user ratings of the type needed by collaborative and content-based techniques.”</li><li>Content-based: No benefit is mentioned in the paper. But I can say that it fits well when a good representation of the item contents can be extracted.</li><li>Utility-based: “It can factor non-product attributes, such as vendor reliability and product availability, into the utility computation, making it possible for example to trade off price against delivery schedule for a user who has an immediate need.”</li><li>Knowledge-based: As for content-based, no benefit is mentioned. But it can be inferred from the rest of the paper that it can be useful when the designer of the application knows beforehand some relevant features that can be computed and directly used to make recommendation decisions.</li></ul><p>Then, the author presents six different ways of mixing the above techniques. In other words, six different hybrid methods. He shows a table summarizing them very well:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*esb2SJyqjXXTadjBV5j3Yg.png" /><figcaption>Table 1. Hybridization methods.</figcaption></figure><p>He also shows a table (that I found awesome) with white spaces on potential future work about hybrid systems. It’s a shame that this paper is from 2002 and many of these may be already implemented by now.</p><h3>A new hybrid method</h3><p>The second part of this paper presents an implementation of a hybrid system on top of an existing application named Entree. The problem with this application is that sometimes “the retrieval process returns a large set of items all of which are equivalent as far as the system’s similarity metric<br>can determine”. This leads immediately to think of a cascade system. Although it is not mentioned in the above table, the cascade method is generally used to discern between ties in the output of the primary recommendation system. The author also mentions that they don’t use other hybrid techniques because they don’t want to lose the semantic ratings (the reason behind the rating).</p><p>I didn’t like this part as much as the first one because it is just a specific case of a hybrid system. But I must admit that, as they include many details of the process, you can get an idea of how to develop a hybrid system more generally.</p><p>One thing that you can notice is that they had to use a custom similarity function specific to the application. Probably this is something that needs to be done in many hybrid systems.</p><p>A contra of the method described in this paper is that arbitrary ratings must be set for mapping each action done in the application. This is needed to use collaborative filtering on top of the knowledge-based method that was previously used in Entree and could lead to scalability problems because of the manual work needed.</p><p>In general, a very good paper that summarizes the previous work in hybrid methods and contributes with a proof of work of a knowledge-based/collaborative hybrid method.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9a8d45596e8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Social tagging recommendations: A collaborative filtering approach]]></title>
            <link>https://medium.com/@andresespinosapc/this-is-a-review-of-collaborative-filtering-for-social-tagging-systems-an-experiment-with-6504fccc1a0c?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/6504fccc1a0c</guid>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[tagging]]></category>
            <category><![CDATA[review]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Mon, 11 Sep 2017 04:27:24 GMT</pubDate>
            <atom:updated>2017-09-11T04:30:33.455Z</atom:updated>
            <content:encoded><![CDATA[<p>This is a review of <a href="https://www.researchgate.net/publication/221141057_Collaborative_filtering_for_social_tagging_systems_An_experiment_with_CiteULike">Collaborative filtering for social tagging systems: An experiment with CiteULike</a> by Denis Parra and Peter Brusilovsky.</p><p>The problem is very well introduced. There’s a very specific comparison between what is needed for social tagging and what is covered by the current methods. Classic content-based and collaborative filtering methods just don’t adjust well to social tagging systems, which are “more diverse in its nature and quality” and explicit feedback is not provided. These are some of the characteristics that make them different, and which can lead to the poor performance the classic methods get. Paraphrasing, while social links and tags improve the personalization quality, the precision stays in the range of 0.1–0.3.</p><p>The dataset is also well described. The method of obtaining the data is mentioned, and more importantly some relevant details about the data that could be a problem but were taken into consideration. This answers in advance some potential reader’s questions.</p><p>In the algorithms section, the Classic Collaborative Filtering algorithm is described. Then the Neighbor-weighted Collaborative Filtering, which is the same but taking into account the number of raters. Finally the BM25-based Similarity, for which the authors had to make an analogy of tags with queries and a set of tags with documents, being able to use an information retrieval algorithm.</p><p>For the study, a good explanation of the difficulties they got was given and how they were solved. A problem with the study could be the few number of subjects, that can lead to more poor metrics and comparison between algorithms. However, this is understandable because of the method used for the evaluation, in which the users had to manually evaluate the system. Besides, the study’s conclusions seem to point in a good direction so future work can be done from it.</p><p>A little detail that I noticed was with the visualization. Lines are used in the plots, while the users are independent to each other, so the lines don’t give relevant information. Lines should be used to indicate change, otherwise, other visualization types should be explored.</p><p>In general, I liked a lot the paper. The explanations went right to the point so I understood the purpose of everything perfectly. Also, good results were obtained, and three relevant conclusions were taken from the experiment:</p><ul><li>Rating scale must be carefully considered.</li><li>Just including the amount of raters can improve the CCF algorithm for social tagging systems.</li><li>A different and suitable similarity function was presented.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6504fccc1a0c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Content-Based Recommendation Methods]]></title>
            <link>https://medium.com/@andresespinosapc/content-based-recommendation-methods-d08da1aec487?source=rss-b6752d2c8611------2</link>
            <guid isPermaLink="false">https://medium.com/p/d08da1aec487</guid>
            <category><![CDATA[review]]></category>
            <category><![CDATA[recommender-systems]]></category>
            <category><![CDATA[content-based]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[recommendation-system]]></category>
            <dc:creator><![CDATA[Andrés Espinosa]]></dc:creator>
            <pubDate>Sun, 10 Sep 2017 21:50:45 GMT</pubDate>
            <atom:updated>2017-09-11T01:52:51.854Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I’m writing about <a href="https://link.springer.com/chapter/10.1007/978-3-540-72079-9_10">Content-Based Recommendation Systems</a> by Michael J. Pazzani and Daniel Billsus.</p><p>It starts mentioning two different ways of building a user profile: Modeling the user’s preferences or using the history of his interactions. I would say that the last one is just additional information that can be used to build the model. Then, the authors mention two approaches of manually providing information to the recommendation system: User customization, where the user can use the interface to provide information about his tastes; and rule-based, where the system is designed to recommend items to some users based on some simple rules, like recommending a song to people that follow the song’s artist in Spotify.</p><p>The rest of the paper is focused on different ways of learning a user model, as they are a “key component of content-based recommendation systems because they learn a function that models each user’s interests”.</p><p>The authors first write about <em>decision trees</em>, which have a “bias that is not ideal for unstructured text classification”, but they have a good “performance, simplicity and understandability”. However, RIPPLE, a <em>rule induction algorithm</em> that is an extension of decision trees, supports multi-valued attributes. They can be useful for text classification because each individual word can be represented as multiple feature values.</p><p>Then, they talk about <em>nearest neighbor methods. </em>The authors say that “despite the algorithm’s unquestionable simplicity, it performs competitively with more complex algorithms”. Nevertheless, I think they are missing an important characteristic of this method: The memory usage. This method is known for using a lot of memory, and this kind of thing must be considered into the performance of the algorithm. In contrast with decision trees, this method can be used for text.</p><p>I will not say anything about the other mentioned methods because they are more of the same. Some technical details are mentioned so you can know, at the surface, how the algorithm works, but no other relevant insight is given.</p><p>In general, the paper makes a strong emphasis on text classification, all of the algorithms but decision trees are said to have a good performance for text classification tasks. But there is no explanation of when to use each algorithm, neither a good comparison between them, and it is difficult to make it myself with the little explanation given for each one. So the paper at the end is just a summary of the existent text-classification methods.</p><p>My favorite section was the Limitations and Extensions. There are some tips of when content-based systems are not the best option. They give the example of jokes and poems, where there is not enough information in the word frequency. In that cases, a collaborative filtering recommender would be useful. They also give another use case of content-based systems: For filtering some results of other methods.</p><p>In conclusion, the most relevant part, in my opinion, was the last one, where some practical tips were given on content-based systems. But the rest of the paper was too informative with too little integration of the topics.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d08da1aec487" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>