Advanced information technologies have enabled the development of online marketplaces that connect businesses and people on a global scale.Much of the analysis of the adoption, growth and engagement on these marketplaces in the extant literature is based on the premise that they are characterized by network effects-a premise that has major implications for their deployment, implementation and management.In this paper we test this premise using data from Kiva, the world's largest online, peer-to-peer social lending marketplace.We find that while network effects are strong and significant during the early growth phase of the marketplace, they become weak or disappear once the marketplace stabilizes.

While apps on the whole increasingly drive and shape present-day Web usage, they individually thrive and fall based on rich patterns of user adoption and long-term engagement. In studying these patterns at the population level, researchers remain curtailed by practical difficulties in accessing consistently detailed data covering large, representative cross-sections of apps. We address this challenge by proposing an empirical framework for analyzing an app»s patterns of adoption and engagement. Importantly, our approach requires obtaining only an app»s daily, weekly, and monthly usership time-series data, which are popular metrics tracked and made available for many apps. Modeling how the nuanced co-dynamics of an app»s daily, weekly, and monthly usership measures (DAU, WAU, and MAU) reliably reveal user adoption and repeat engagement, we extend and demonstrably improve on the predictive performance of prior install- or DAU-based methods. To further study the mechanisms by which successful apps cultivate their userbases, we position apps along the dual axes of viral adoption and retentive engagement as mechanisms to potentially explain success. Applying our approach to data on Facebook apps, we show that these apps over time became less viral and more engaging. Interestingly, despite their diminished virality, their improved user retention subtly and counter-intuitively boosted these apps' rates of new user adoptions through user-activity-based word of mouth, because an engaged user contributes to word of mouth over a longer period of time. In a final case application, we introduce evidence that developers of apps that learn to successfully retain their users carry this valuable experience over into their new apps.

This paper develops a microeconomic framework for the determination of data-processing budgets over time, and, in particular, the allocation of these budgets between software and hardware. The model dynamically balances the value and cost of information services, given the prevailing cost trends. It regards software and hardware as inputs to the process of producing information services, and identifies the complementarity and substitution between them as major determinants of the efficient budget allocation. The theory provides a basis for understanding the budgeting process and for predicting future trends, and is applied to actual budget data.

This paper presents a methodology for trading-off the cost of incomplete information against the data-related costs in the design of database systems. It investigates how the usage patterns of the database, defined by the characteristics of information requests presented to it, affect its conceptual design. The construction of minimum-cost answers to information requests for a variety of query types and cost structures is also studied. The resulting costs of incomplete database information are balanced against the data-related costs in the derivation of the optimal design.

Extendible hashing is an attractive direct-access technique which has been introduced recently. It is characterized by a combination of database-size flexibility and fast direct access. This paper derives performance measures for extendible hashing, and considers their implecations on the physical database design. A complete characterization of the probability distribution of the directory size and depth is derived, and its implications on the design of the directory are studied. The expected input/output costs of various operations are derived, and the effects of varying physical design parameters on the expected average operating cost and on the expected volume are studied.

A new approach to the analysis of hash table performance is presented.This approach is based on a direct probabilistic analysis, where the underlying probabilities are derived by using the ballot theorem and its ramifications.The method is first applied to analyze the performance of the classical (cyclic) linear probing scheme, and the results are used to solve an optimal storage allocation problem.A scheme frequently used in practice where the table is linear rather than cyclic is then analyzed using the same methodology.

A random-access file with N storage locations ~s consmdered.Records are added to the file from time to tmme A record wroth key ~f ~ ~2 ms hashed to storage location F(,o).A collismon ms resolved by the foUowmg chaining method' All records hashed to the same locatmon are chained to each other to form an ordered list, ordered m ascending order of the keys The first record of a list is stored either at location F(~0) or at an alternative start ff IocaUon F(,o) ms occupied.For thins process the multidimensional ume-dependent generating funcuon is derived, and the expected values of various state variables are calculated.These values are used to obtain formulas for the expected number of I/O operations needed for retrieval, addmon, or updating of a record Two measures of retrieval performance are calculated: (i) The expected number of addiuonal probes needed to find a record m the file This measure ms uniformly bounded by t. (n) The expected number of addmonal probes required to dmscover that a record is not in the file.This performance measure IS always smaller than the first and ms uniformly bounded by I/e.Addmon of a record consists of three steps.(l) checking that a record with the same key does not exmst in the file, (2) finding an empty locatmon, and (3) wrmng the record and updating all the pointers revolved The number of I/O operatmons needed for record addmon depends on the amount of reformation avadable on the occupancy of the file For various mformatmn levels the relevant performance measures are calculated and compared KEY WORDS AND PHRASES: hashing, colhslon resolutmon, chammg method, ordered list, multsdlmensmonal generatmg function, remeval performance CR CATEGORIES 3.72, 3 74, 4 33, 5.