In this talk we will give a very brief overview of Flipkart, highlighting the major milestones in its journey so far and some latest market size numbers. We will enumerate some of the important data science challenges in an e-commerce company like ours. Finally we will cover one of the problems, demand forecasting, in some details to highlight some of our recent work: Accurate demand forecasts can help on-line retail organizations better plan their supply-chain processes. The challenge, however, is the large number of associative factors that result in large, non-stationary shifts in demand, which traditional time series and regression approaches fail to model. We propose a Neural Network architecture called AR-MDN, that simultaneously models associative factors, time-series trends and the variance in the demand

Several approaches to collaborative filtering have been studied but seldom have studies been reported for large (several millionusers and items) and dynamic (the underlying item set is continually changing) settings. In this paper we describe our approach to collaborative filtering for generating personalized recommendations for users of Google News. We generate recommendations using three approaches: collaborative filtering using MinHash clustering, Probabilistic Latent Semantic Indexing (PLSI), and covisitation counts. We combine recommendations from different algorithms using a linear model. Our approach is content agnostic and consequently domain independent, making it easily adaptable for other applications and languages with minimal effort. This paper will describe our algorithms and system setup in detail, and report results of running the recommendations engine on Google News.

We present a novel Locality-Sensitive Hashing scheme for the Approximate Nearest Neighbor Problem under lp norm, based on p-stable distributions.Our scheme improves the running time of the earlier algorithm for the case of the lp norm. It also yields the first known provably efficient approximate NN algorithm for the case p<1. We also show that the algorithm finds the exact near neigbhor in O(log n) time for data satisfying certain "bounded growth" condition.Unlike earlier schemes, our LSH scheme works directly on points in the Euclidean space without embeddings. Consequently, the resulting query time bound is free of large factors and is simple and easy to implement. Our experiments (on synthetic data sets) show that the our data structure is up to 40 times faster than kd-tree.

The need to deal with massive data sets in many practical applications has led to a growing interest in computational models appropriate for large inputs. The most important quality of a realistic model is that it can be efficiently implemented across a wide range of platforms and operating systems. In this paper, we study the computational model that results if the streaming model is augmented with a sorting primitive. We argue that this model is highly practical, and that a wide range of important problems can be efficiently solved in this (relatively weak) model. Examples are undirected connectivity, minimum spanning trees, and red-blue line segment intersection, among others. This suggests that using more powerful, harder to implement models may not always be justified. Our main technical contribution is to show a hardness result for the "streaming and sorting" model, which demonstrates that the main limitation of this model is that it can only access one data stream at a time. Since our model is strong enough to solve "pointer chasing" problems, the communication complexity based techniques commonly used in showing lower bounds for the streaming model cannot be adapted to our model. We therefore have to employ techniques to obtain these results. Finally, we compare our model to a popular restriction of external memory algorithms that access their data mostly sequentially.

Systems for processing continuous monitoring queries over data streams must be adaptive because data streams are often bursty and data characteristics may vary over time. We focus on one particular type of adaptivity: the ability to gracefully degrade performance via "load shedding" (dropping unprocessed tuples to reduce system load) when the demands placed on the system cannot be met in full given available resources. Focusing on aggregation queries, we present algorithms that determine at what points in a query plan should load shedding be performed and what amount of load should be shed at each point in order to minimize the degree of inaccuracy introduced into query answers. We report the results of experiments that validate our analytical conclusions.

We study the index selection problem: Given a workload consisting of SQL statements on a database, and a user-specified storage constraint, recommend a set of indexes that have the maximum benefit for the given workload. We present a formal statement for this problem and show that it is computationally "hard" to solve or even approximate it. We develop a new algorithm for the problem which is based on treating the problem as a knapsack problem. The novelty of our approach lies in an LP (linear programming) based method that assigns benefits to individual indexes. For a slightly modified algorithm, that does more work, we prove that we can give instance specific guarantees about the quality of our solution. We conduct an extensive experimental evaluation of this new heuristic and compare it with previous solutions. Our results demonstrate that our solution is more scalable while achieving comparable quality.

The sliding window model is useful for discounting stale data in data stream applications. In this model, data elements arrive continually and only the most recent N elements are used when answering queries. We present a novel technique for solving two important and related problems in the sliding window model --- maintaining variance and maintaining a k-- median clustering. Our solution to the problem of maintaining variance provides a continually updated estimate of the variance of the last N values in a data stream with relative error of at most # using O( # 2 log N) memory. We present a constant-factor approximation algorithm which maintains an approximate k--median solution for the last N data points using O( N) memory, where # < 1/2 is a parameter which trades o# the space bound with the approximation factor of O(2 ).

In many applications involving continuous data streams, data arrival is bursty and data rate fluctuates over time. Systems that seek to give rapid or real-time query responses in such an environment must be prepared to deal gracefully with bursts in data arrival without compromising system performance. We discuss one strategy for processing bursty streams --- adaptive, load-aware scheduling of query operators to minimize resource consumption during times of peak load. We show that the choice of an operator scheduling strategy can have significant impact on the run-time system memory usage. We then present Chain scheduling, an operator scheduling strategy for data stream systems that is near-optimal in minimizing run-time memory usage for any collection of single-stream queries involving selections, projections, and foreign-key joins with stored relations. Chain scheduling also performs well for queries with sliding-window joins over multiple streams, and multiple queries of the above types. A thorough experimental evaluation is provided where we demonstrate the potential benefits of Chain scheduling, compare it with competing scheduling strategies, and validate our analytical conclusions.

In this overview paper we motivate the need for and research issues arising from a new model of data processing. In this model, data does not take the form of persistent relations, but rather arrives in multiple, continuous, rapid, time-varying data streams. In addition to reviewing past work relevant to data stream systems and current projects in the area, the paper explores topics in stream query languages, new requirements and challenges in query processing, and algorithmic issues.

We consider the problem of maintaining aggregates and statistics over data streams, with respect to the last N data elements seen so far. We refer to this model as the sliding window model. We consider the following basic problem: Given a stream of bits, maintain a count of the number of 1's in the last N elements seen from the stream. We show that, using $O(\frac{1}{\epsilon} \log^2 N)$ bits of memory, we can estimate the number of 1's to within a factor of $1 + \epsilon$. We also give a matching lower bound of $\Omega(\frac{1}{\epsilon}\log^2 N)$ memory bits for any deterministic or randomized algorithms. We extend our scheme to maintain the sum of the last N positive integers and provide matching upper and lower bounds for this more general problem as well. We also show how to efficiently compute the L p norms ($p \in [1,2]$) of vectors in the sliding window model using our techniques. Using our algorithm, one can adapt many other techniques to work for the sliding window model with a multiplicative overhead of $O(\frac{1}{\epsilon}\log N)$ in memory and a $1 +\epsilon$ factor loss in accuracy. These include maintaining approximate histograms, hash tables, and statistics or aggregates such as sum and averages.

Studies the problem of approximately answering aggregation queries using sampling. We observe that uniform sampling performs poorly when the distribution of the aggregated attribute is skewed. To address this issue, we introduce a technique called outlier indexing. Uniform sampling is also ineffective for queries with low selectivity. We rely on weighted sampling based on workload information to overcome this shortcoming. We demonstrate that a combination of outlier indexing with weighted sampling can be used to answer aggregation queries with a significantly reduced approximation error compared to either uniform sampling or weighted sampling alone. We discuss the implementation of these techniques on Microsoft's SQL Server and present experimental results that demonstrate the merits of our techniques.

Association-rule mining has heretofore relied on the condition of high support to do its work efficiently. In particular, the well-known a priori algorithm is only effective when the only rules of interest are relationships that occur very frequently. However, there are a number of applications, such as data mining, identification of similar Web documents, clustering, and collaborative filtering, where the rules of interest have comparatively few instances in the data. In these cases, we must look for highly correlated items, or possibly even causal relationships between infrequent items. We develop a family of algorithms for solving this problem, employing a combination of random sampling and hashing techniques. We provide analysis of the algorithms developed and conduct experiments on real and synthetic data to obtain a comparative performance analysis.

Association rule mining has heretofore relied on the condition of high support to do its work efficiently. In particular, the well-known a-priori algorithm is only effective when the only rules of interest are relationships that occur very frequently. However, there are a number of applications, such as data mining, identification of similar Web documents, clustering and collaborative filtering, where the rules of interest have comparatively few instances in the data. In these cases, we must look for highly correlated items, or possibly even causal relationships between infrequent items. We develop a family of algorithms for solving this problem, employing a combination of random sampling and hashing techniques. We provide an analysis of the algorithms developed and conduct experiments on real and synthetic data to obtain a comparative performance analysis.