<?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 Arman Akbarian on Medium]]></title>
        <description><![CDATA[Stories by Arman Akbarian on Medium]]></description>
        <link>https://medium.com/@armanakbarian?source=rss-4299b21b34ec------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Arman Akbarian on Medium</title>
            <link>https://medium.com/@armanakbarian?source=rss-4299b21b34ec------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 18:19:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@armanakbarian/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[Expressiveness and Robustness of Graph Neural Networks in Causal Effect Estimation and Node Level…]]></title>
            <link>https://armanakbarian.medium.com/expressiveness-and-robustness-of-graph-neural-networks-in-causal-effect-estimation-and-node-level-71d57a2ab60c?source=rss-4299b21b34ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/71d57a2ab60c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[graph-neural-networks]]></category>
            <dc:creator><![CDATA[Arman Akbarian]]></dc:creator>
            <pubDate>Wed, 09 Oct 2024 21:03:13 GMT</pubDate>
            <atom:updated>2024-10-09T21:03:13.457Z</atom:updated>
            <content:encoded><![CDATA[<h3>Expressiveness and Robustness of Graph Neural Networks in Causal Effect Estimation and Node Level Regression on Large Scale Product Substitutability Graphs</h3><p>Submitted as course project for <em>Machine Learning with Graphs (CS224W 2023), Stanford University</em></p><p>Authors: Arman Akbarian, Arjun Srinivasan</p><blockquote><a href="https://colab.research.google.com/drive/14TXA-5uxhcBIGKWgAGw3qUficpLGiVcR?usp=sharing">Colab Link to Synthetic Data Generation</a></blockquote><blockquote><a href="https://colab.research.google.com/drive/1wLxLWba-zGPK5q_9vWJwM3BP8l_YEnVh?usp=sharing">Colab Link to Modeling</a></blockquote><h3><strong>1. Introduction</strong></h3><p>Applications of machine learning models in estimating the causal effect is an active area of research and has attracted multi-disciplinary community of scientists in economics, statistics, social science and computer science. In particular, the literature of causal inference has evolved from using linear models with clear advantage of interpretability and design (choice of control variables, interaction terms and interpretation of model weights as average treatment effect) to learners that are more complex and non-linear in nature (such as tree based models).</p><p>Recently choice of neural networks has been adopted for causal inference (1, 2) and deep learning based models are proven to generate state of the art results in estimating conditional average treatment effect (3). This is due to the fact that DL models are very low biased estimators and are able to approximate complex functions in high dimensions, making them great candidate to extract conditional causal effects on different parts of the high dimensional input space. However, to the best of our knowledge, graph neural networks have yet to be adopted in the field of causal ML. In this paper, we explore why GNNs are great candidate for causal inference in network of elements where SUTVA assumption of causal modeling no longer holds (i.e. treatment and outcomes of different items are not independent).</p><p>A prominent example of this is the causal estimation of demand change as a function of price change (i.e. elasticity) of highly connected items. There are many real world examples where the dependency of demand on price of an item cannot be studied in isolation due to connectivity and network of relations between items. For example, in online retailer such as Amazon or Walmart there are many products that are substitutable and a change of price in one item can shift the demand to other products. In hotels and accommodation industry such as AirBnB, the relation between demand and price cannot be easily applied for single item optimization of maximizing profit (or other financial metrics) since an incremental change to price of one accommodation can shift the demand to/from other accommodation that are in close radius of each other; this in turn renders the single optimal price estimation to be in fact sub-optimal for global maximization of profit on all offering of the company. This motivates us to explore graph neural networks as candidate learner that can estimate the demand of a “product” (in generic sense, could be a retail item, accommodation etc.) as a function its price and attributes and also as a function of price and attributes of neighboring products. In addition, we explore how this learner can be used to understand the causal effect of price change on items’ demand and how demand flows in the network as prices are adjusted.</p><p>In this project we will explore the expressiveness power of GNN models in the node level regression task of predicting the demand of each product. We will use a complex synthetic graph and data generation process with reasonable assumptions of how the attributes and prices of neighboring products impact the demand of the target node and see if the well-know GNN architectures can recover the underlying demand relation to the price and neighboring product prices. Synthetic data is a standard practice in measuring robustness of causal effect estimation since in real world datasets the counterfactual is never observed and knowing the underlying causal relation will help us to have a ground truth on the causal effect and study if the GNN models perform well in estimating individual treatment effect on graphs.</p><h4><strong>1.1 Substitutability Graph</strong></h4><p>Let’s first discuss the graph of product connectivity. We will use online retailer’s catalog of product as the target real world application of this study. As depicted in Fig. 1, we define an edge between two products if they are highly substitutable for each other. The concept of substitutability is not a binary value and we can consider a complete graph where the weight of each edge between product A and B defines the likelihood of a customer considering product A in their purchase decision when they consider product B. However, realistically, products serve a function (except perhaps items that are often gifted) and this clearly disallows many substitutability connections in the graph. For example when a customer uses an online retailer to purchase a TV, they will definitely not consider a book! This concept of a serving function of a product will define one of the key categorical attributes of the nodes: “<em>product function”.</em> This means we will consider the graph to be disjointed sub-graphs with rarely any edge between nodes that do not have the same product function (except commonly gift-able item, but for the sake of simplicity we ignore this reality).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/721/1*Oa1_mYbzqBhVWm8HJuWSeA.png" /><figcaption>Figure 1: Visualization of how the substitutability graph will be generated synthetically. Here we are demonstrating a collection of TVs with brands B_{i} and sizes S1 or S2. As you can see, there is a clique of 6 brands that are all connected and can be considered substitutable. In addition, a reasonable assumption could be that a slightly different sizes of these TVs (S2 ≈ S1) of the same brand is also a candidate for substitution and should be joined by an edge so the GNN model can create functional dependency of demand on their prices. The dotted lines denote the edges, colored dotted lines denote the edges in between cliques and solid lines with an arrow are denoting edges when regression task is being performed for the target node (B6, S1). The arrows are demonstrating the causal effect of increasing the price on this node and demand flowing to the neighboring edges, i.e. the price of this node impacts the demand (target variable) of the other nodes and can have secondary smaller effect on the demand of the other clique’s nodes with size S2 via the edge that connects B6 nodes. This effect can be captured only if we consider our GNN model to have depth larger than 1.</figcaption></figure><p>To be more concrete, we assume the graph has strong homophily where attributes of the nodes that are connected are highly similar. By definition we also expect the substitutability to be somewhat a transitive property (i.e. if A is substitute for B and B is a substitute for C, we except A to be also substitute for C). This effectively means we expect the nodes to often be part of cliques of multiple products. For example various premium brands of TVs of same size and resolution can be considered a clique of substitutability and all connected to each other. Here we can imagine a weak substitutability with TVs of a slightly different size (e.g. 65&quot; TV may be considered as alternative to 75&quot; TV of the same brand if the 75&quot; version is priced significantly higher). For simplicity, when we discuss the synthetic graph generation in section 3, we will assume the graph is unweighted and due to the homophily of the graph we expect the model to rely on attribute of the products and the attention (interaction term between two nodes’ attributes) mechanism to decide how important certain product’s price is in determining demand of the other connected product. In another word, the homophily nature of the graph makes edge weights to be a redundant co-variate that is highly correlated to the similarity of the node features and can be ignored in our synthetic graph generation without losing generality.</p><h3>2. Why Graph Neural Networks</h3><p>The first part of the project is to develop a node level regression model on the substitutability graph (G, V ). The task is conditional estimation of demand of a product d_{i} given its price p_{i} and attributes as well as price and attributes of the connected neighboring nodes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mOJjoXaa2ONUB7ZUijzfOg.png" /></figure><p>Here x_{i} denotes the feature vectors of the node (i.e. attributes of the product) and N_{i} denotes neighbors of node i. Note that, technically, price attribute can be absorbed to the rest of product attributes in x_{i}, but since we want to study the causal effect of change in price on demand, we keep the price explicit in our notation. As one can see this functional dependency of φ can be expressed in the from of message passing and aggregation framework of graph neural networks, for example we can consider the following computational graph:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W8QExGIfo0qf7RuUGi05fg.png" /></figure><p>where MLP denotes the prediction head of multi-layer dense NN layer for regression task at node level, h^{(l)} is the node embedding at level l of the GNN, AGG is the permutation invariant aggregation operation (can have learnable parameters) and MSG is the choice of learnable layer (with non- linearity) for message passing between nodes. Note that in this formulation if we choose the depth of L = 1 for our GNN, we have achieved the desired functional dependency in eq. 1 (i.e. the demand of node will depend only on price and attributes of a product and its neighbors). Increasing the depth of the network will yield a functional dependency on substitutes of the neighbors of the node that are mainly neighbors again as the nodes are often members of cliques in the graph. However higher depth in the network might help in capturing the dependency of the demand on the price of neighboring cliques as depicted in Fig. 1.</p><h3>3. Synthetic Training Dataset &amp; Validation</h3><p>We will discuss the synthetic training data generation in two parts: 1) how the graph is generated and 2) how the demand function will define a complex dependency between nodes and final target value (demand of a node given various prices of the node and its neighbors) will be generated.</p><h4>3.1 Synthetic Substitutability Graph</h4><p>We define the substitutability graph as a collection of cliques generated from 200 unique product functions (consider it as categorical variable / ID) and 2000 unique brands (again a categorical / integer ID). To construct the product cliques first we construct the brand cliques of sizes [2,… 6], [20, 60] and one large clique of 200 brands. These define the brand substitutability cliques and we consider this graph to be weighed by weights B_{ij}. We consider a self-loop (i.e. diagonal component of B_{ij} is 1.0) and sample random uniform weights for each clique of brand substitutability graph ranging between [0.5, 1.0]. These values later will be used in defining our synthetic training dataset. However, node that this is completely “hidden” from the model and is the underlying hidden variables that define our training data, and not provided to the model to learn from.</p><p>Given the brand substitutability graph, we define two products substitutable if they have the same product function ID, their brands are substitutable and their “other” attributes beside brand and product function IDs are within a certain distance from each other. Please see next section for how we randomly sample the values for these other attributes within a substitutability product clique. Last step is to define a bipartite graph that connects Brand IDs to Product Function IDs indicating that brand i offers product function j, here we ensure that brands that are within a the same brand substitutability graph have overlapping and commonality on what product functions they offer (e. g. “Apple” and “Samsung” are both electronic brands and offer “Laptop”, “Phone”, “Monitor” etc…, it doesn’t make sense to consider two brand IDs substitutable if their product function IDs they offer are not overlapping)</p><p>Our final graph includes 390K nodes, 27K cliques of products with size 2 to 23. Since our graph is already nicely clustered to cliques, we will use cliques in our batches of data for training and do not need sophisticated graph subsampling.</p><h4>3.2 Synthetic Node Attributes, Prices and Demands for Training</h4><p>To create ground truth, we need to formulate a complex dependency of demand on price and product attributes of a node and its neighbors. To define these closed form functions we will consider various hidden variables. In addition we explicitly split the node/product feature vectors to:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wl4S7IeJb45XghZ_wUC0oQ.png" /></figure><p>in which b is the categorical variable brand, f is categorical variable “product function” we discussed, s is other continuous attributes (e.g. size, weight etc.) and c is the other categorical variables (e.g. color, flavor<br>etc.). Now given the notation above the following equations can be used to generate demand values for a clique of substitutable products:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DH3mVsf1Um-P-fWzFSLXPQ.png" /></figure><p>in which H(.) is harmonic mean over the clique price values and d_{i} is calculated using:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eTUTMOnZvLlTg_HG9XgeaQ.png" /></figure><p>in which function o is defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AS7AXertPUI4OrrqqNYf0Q.png" /></figure><p>and Λ is a normalization factor:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XifC_6l_vj7jubZ1YUrzZQ.png" /></figure><p>B_{ij} is the brand substitutability weight of the product i and j in the clique. |N| denotes the size of the clique, and `[f]` denotes that parameters vectors ρ and σ depend on product function ID f (i.e. randomly sampled and are different for each product function ID). Note that indices in s_j and c_k are indices along the continuous and categorical dimensions of each node’s attribute and should not be confused with index i, j above that are indices along the nodes of the clique. Now are data generation process is as following:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5aVZJQ81cNVpzEEILvi3xg.png" /></figure><p>in which curly N denotes normal distribution, and within each clique we choose the same c and s attributes and randomly change values in the dimensions for which ρ or σ ‘s are randomly selected to be zero. This may seem tedious but allows for highly complex dependency of demand on product attributes and prices and in particular creates an important property when in a substitutability clique we change one node’s price as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UyO2JyrsIsieDCAExWWR6Q.png" /><figcaption>Figure 2. Ground truth relation between a node’s price and its demand as well as a neighboring node’s demand (here the neighbors’ prices are kept constant). The slope of the demand shift to neighboring nodes (i.e. “cross elasticity of price”) is a function of many hidden variables, including size of the clique itself, and is more rapid on smaller highly substitutable cliques of products.</figcaption></figure><h4>3.3 Price exploration and train/validation/test split</h4><p>Finally to create the demand, we sample 10 different prices for each node for each clique around the o_i values in the equations above. We call one sampling, a <em>“price exploration”</em>, i.e. for each combination of prices within a clique we will get a new values for demands. Note that our training dataset does *NOT* contain any controlled scenario where all prices of surrounding nodes are kept constant and we only change the price of the node. For example for a randomly selected clique of size 17 each node’s 10 price and demand l0 demand values look as following:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5E1Nonb-2dXCy1v4g7pmcQ.png" /><figcaption>Figure 3. Demand and prices of nodes when we no longer control for any of the prices in neighboring nodes of a clique. This is a clique size of 17</figcaption></figure><p>As you can see, there are some evidence of the inverse relation between demand and price, but it is no longer available to the model as the neighboring prices are also changing. This “noisy demand” when we do not control for the surrounding neighborhood is more prominent in smaller cliques:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2FWwhNmvR_xdsyFvjaUXFQ.png" /><figcaption>Figure 4. In smaller cliques, exploration of prices result in more “noisy” looking demand values as we no longer control for neighboring prices.</figcaption></figure><h4>3.4 Validation and Metric</h4><p>Since we are particularly interested in understanding if model is capable of recovering the price demand relationship, we split our dataset along the price exploration, we choose 6 instance of price exploration for training, 3 for validation and 1 for testing. In addition, we prepare a particular price vectors where we only vary the price of 1 node in the clique as shown in the Fig 2. to see if the model is capable of recovering the relationship. To measure model’s regression we use Mean Square Error (MSE) and R² on the test data.</p><h3>4. Modeling</h3><p>In this section we discuss our design choices of the neural network architecture, and particular designs to allow the model to be able to recover the price/demand relation. There are limited literature in application of neural networks (and in general adaptive function ML models) for causal effect estimation of continuous variable. <a href="https://arxiv.org/pdf/2103.07861.pdf">Nie et. al (4) </a>proposed VCNet where the representation of control X is guided by a multi-headed objective of predicting treatment (for us: price) and outcome (i.e. demand) at the same time. This allows the model to observe X (control parameters such as product attributes etc) separately than the treatment variable, in turn treatment parameter is added to predict the outcome.</p><p>We <a href="https://colab.research.google.com/drive/1U7yJunmWxTqojtB9kJZ2KiYnsl6CFNwK?usp=sharing">experimented with this idea</a>, but realized a simplification of it to late fuse the control parameter price after preprocessing of node features than providing the price at initially as a node feature to be equally powerful. Our architecture choice is shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IqW6fc44sXi8BlvjyEzT8Q.png" /><figcaption>Figure 5. The neural network architecture we explored. Here “Graph Convolution” layer will be experimented with (e.g. GCN, GraphSage, GAT etc.) while we identified two simple yet powerful ideas: a) Late fusion of the treatment variable price separate than rest of the parameters allows the model to recover the price/demand relation better and combined with idea (b) it is effectively as powerful as VCNet proposed by <a href="https://arxiv.org/pdf/2103.07861.pdf">Nie L., et. al [4].</a> b) The skip connection improves the speed of convergence and is effectively a non-graph-neural-network path where nodes’s demand is model only as a function of node feature itself. Without the GNN layer, this path also established our “baseline model”, which has small (yet nonzero) predictive power.</figcaption></figure><p>For the “Graph Convolution” layer we experiment with all the well known graph convolution layers: GCN <a href="https://arxiv.org/abs/1609.02907">[5]</a>, GraphSage <a href="https://cs.stanford.edu/people/jure/pubs/graphsage-nips17.pdf">[6]</a>, GINConv <a href="https://arxiv.org/abs/1810.00826">[7]</a> and Graph Attention <a href="https://arxiv.org/abs/2105.14491">[8]</a>.</p><h3>5. Results</h3><h4>5.1 Regression and Recovering Demand</h4><p>Let’s first see if model is capable of predicting the demand given its input. Table 1. summarizes various experiments we ran using well know GNN architectures and also how various model’s are capable of utilizing larger data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ycSvHl47KaFMgtw783EDUg.png" /><figcaption>Table 1. Performance of various choices of GNN layers measured as Mean Squared Error on test/validation data. Note that “baseline” is the architecture in Fig 5. where we eliminate the GNN layer and use the skip layer directly for each ASIN. As you can see, some of the variability can be explained by this baseline model, but obviously model cannot predict without knowing the neighboring information and message passing plays a clear role in improving regression. Few other observations here are: 1) We were surprisingly unable to get GINConv to work 2) Notice how adding skip layer improves all of our architectures 3) GraphSage is able be in par with Graph Attention model but was unable to utilize more data while GAT was able to continue improving given larger dataset and more training time. 4) To no surprised GAT performed the best as we have terms such as B_{ij} that are set to random and a node’s demand-price influence on other node depends on this interaction term in between them. 5) GAT with 5 head outperforms GAT with 10 head in small dataset, while as we increase training datasize the larger capacity of 10 head becomes more evident.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BFqMfAAr3BCA2A341rgffw.png" /><figcaption>Figure 6. Best model’s prediction vs actual demand and the residual, as you can see model is biased for larger demand values and has higher relative error for smaller values of demand.</figcaption></figure><h4>5.2 Recovering Demand and Price Relation in Controlled Setting</h4><p>In this section, let’s take a look at if our best model can recover the causal relation between price and demand for a node itself as well as its impact on neighboring nodes. See the diagrams below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ErtxY-jaDJkiGgtilDCltQ.png" /><figcaption>Figure 7. For a clique of size 17, we construct a dataset where only 1 nodes price is changing and see if model’s prediction is aligned with the ground truth of the causal impact of price on demand of the node and its neighboring node. As you can see model performs better on node itself than on the neighboring nodes. Note that the slope of this diagrams (in log scale) is defined as “price elasticity”, and as you can see the cross elasticity (one nodes price change on another node) can be highly biased.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oIXN9tw45FFiVORNUAncmw.png" /><figcaption>Figure 8. Same as Fig.7 for a clique of size 5. Here model performs better on cross elasticity estimation (slope of the diagrams) However performs worse on node’s elasticity estimation itself, here the demand shift between nodes in the training dataset is less smooth as there are less number of nodes. One approach to improve the model’s consistency in prediction between large cliques and small cliques would be to add two GNN layers with two different aggregation function such as “mean” and “sum”.</figcaption></figure><h3>6. Final Remarks</h3><p>In this study we showcased that GNNs are capable of recovering a highly complex relations between price and demands in a connected substitutability cliques of products. In particular, GAT is able to learn the relative cross elasticities in between two nodes that depends on both nodes. Note that we haven’t debiased the model yet, as you can see from the diagrams, the cross elasticities on smaller cliques can be consistently higher or lower than the true elasticity which can impact our overall estimate on the full population. Next step of this project is to explore how de-biasing frameworks such as DML (1) can be proposed in a graph setting.</p><h3>7. References</h3><ol><li>Chernozhukov, V., et. al. 2022 <a href="https://arxiv.org/abs/2110.03031"><em>RieszNet and ForestRiesz: Automatic Debiased Machine Learning with Neural Nets and Random Forests</em></a></li><li>Shi C., et. al. 2019 <a href="https://arxiv.org/abs/1906.02120"><em>Adapting Neural Networks for the Estimation of Treatment Effects</em></a></li><li>Shalit, U., et. al.2017 <a href="https://arxiv.org/abs/1606.03976"><em>Estimating individual treatment effect: generalization bounds and algorithms</em></a></li><li>Nie L. et. al. 2021 <a href="https://arxiv.org/pdf/2103.07861.pdf"><em>VCNET and functional targeted regularization for learning causal effects of continuous treatments</em></a></li><li>Kipf et. al. 2016 <a href="https://arxiv.org/abs/1609.02907"><em>Semi-Supervised Classification with Graph Convolutional Networks</em></a></li><li>Hamilton et. al. 2017<em> </em><a href="https://cs.stanford.edu/people/jure/pubs/graphsage-nips17.pdf"><em>Inductive Representation Learning on Large Graphs</em></a></li><li>Xu et. al. 2018 <a href="https://arxiv.org/abs/1810.00826"><em>How Powerful are Graph Neural Networks</em></a></li><li>Brody et. al. 2021 <a href="https://arxiv.org/abs/2105.14491"><em>How Attentive are Graph Attention Networks</em></a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71d57a2ab60c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hello World!]]></title>
            <link>https://armanakbarian.medium.com/trying-out-medium-45872c6dda2f?source=rss-4299b21b34ec------2</link>
            <guid isPermaLink="false">https://medium.com/p/45872c6dda2f</guid>
            <dc:creator><![CDATA[Arman Akbarian]]></dc:creator>
            <pubDate>Mon, 26 Jun 2017 03:22:55 GMT</pubDate>
            <atom:updated>2017-06-26T03:29:59.323Z</atom:updated>
            <content:encoded><![CDATA[<img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45872c6dda2f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>