A machine learning approach for result caching in web search engines

Highlights

• To the best of our knowledge, our work is the

  rst in literature to apply machine learning techniques to the result caching problem in search engines, for both static, dynamic, and state-of-the-art static-dynamic cache organizations.

• We evaluate a large set of features and illustrate that they can be exploited to increase the hit rate of result caches.

• We evaluate various oracle caching strategies to illustrate the potential room for improvement in the result caching problem.

• We show that the proposed machine learning framework can improve the hit rate of result caches, potentially reducing the energy consumption in search engines.

Abstract

A commonly used technique for improving search engine performance is result caching. In result caching, precomputed results (e.g., URLs and snippets of best matching pages) of certain queries are stored in a fast-access storage. The future occurrences of a query whose results are already stored in the cache can be directly served by the result cache, eliminating the need to process the query using costly computing resources. Although other performance metrics are possible, the main performance metric for evaluating the success of a result cache is hit rate. In this work, we present a machine learning approach to improve the hit rate of a result cache by facilitating a large number of features extracted from search engine query logs. We then apply the proposed machine learning approach to static, dynamic, and static-dynamic caching. Compared to the previous methods in the literature, the proposed approach improves the hit rate of the result cache up to 0.66%, which corresponds to 9.60% of the potential room for improvement.

Introduction

Scalability and efficiency are two crucial aspects of performance in search engines (Cambazoglu & Baeza-Yates, 2015). A commonly used technique for improving search engine performance is result caching (Baeza-Yates et al., 2007a). In result caching, precomputed results (e.g., URLs and snippets of best matching pages) of certain queries are stored in a fast-access storage. The future occurrences of a query whose results are already cached can be directly served by the result cache, eliminating the need to process the query using costly computing resources. Result caching has two immediate benefits to a search engine. First, by reducing the computational load on the server side, it enables higher query processing throughput. Second, it reduces the average query response time perceived by the users.

In the result caching problem, the goal is to maintain a set of previously computed query results in a limited-capacity cache such that some performance metric is optimized over the time as new queries are received. Although other performance metrics are possible (Altingovde, Ozcan, Ulusoy, 2009, Gan, Suel, 2009), the main performance metric for evaluating the success of a result cache is hit rate, i.e., the fraction of queries that are answered by the cache. In practice, increasing the hit rate requires careful selection of queries whose results are to be cached.

In the literature, there are two types of result caches: static and dynamic. In static caching, the cache is filled in an offline manner with the results of queries selected from the past query logs. The underlying assumption in static caching is the steady behavior in the query stream, i.e., queries which were popular in the past remain popular in the future. Therefore, popular queries are preferred over the rest when making the caching decisions. In dynamic caching, the caching decisions are online and the goal is to identify queries that are least likely to result in a cache hit and evict such queries to make room for other, more recent queries. The underlying assumption in dynamic caching is the bursty behavior in the query stream, i.e., queries which are submitted more recently have higher probability of reoccurring in the future. Fagni, Perego, Silvestri, and Orlando (2006) show that a hybrid caching approach combining these two types of caches performs better than using them in isolation.

Past research has tried to improve the performance of the result caching by relying on a single or limited number of features. The primary objective of this work is to devise a unifying caching framework which exploits an extensive set of features. To this end, we propose a machine learning approach that combines a large variety of features extracted from search engine query logs and evaluate the impact of this framework on the hit rate of a search engine result cache. We devise different learning models for static and dynamic result caching. For the former type of caches, we focus on the offline cache allocation problem. For the latter type of caches, we focus on the online eviction problem. As a secondary objective, we aim to identify the potential room for improvement in result cache hit rate. To this end, we propose and evaluate various oracle algorithms to understand the potential room for improvement in hit rate. Although machine learning is used in different caching tasks (e.g., time-to-leave prediction (Alici, Altingovde, Ozcan, Cambazoglu, & Ulusoy, 2012), refreshing (Jonassen & Bratsberg, 2012), and admission (Ozcan, Altingovde, Cambazoglu, Junqueira, & Ulusoy, 2012)), our approach is the first that employs machine learning in result caching and eviction.

Our experiments are conducted using a large, real-life query log obtained from Yahoo Web Search. We evaluate our models within the state-of-the-art static-dynamic caching framework proposed in (Fagni et al., 2006). Compared to this state-of-the-art framework, the proposed approach improves the hit rate by 0.47%, which corresponds to 7.8% of the possible improvement. Although the improvement in the hit rate is quite modest, it presents a large potential financial benefit for commercial search engines.

The rest of this paper is organized as follows. In Section 2, the previous work on result caching is surveyed. In Section 3, we present the features used in our machine learning models. Section 4 presents the proposed techniques. The details of our query log and experimental setup are explained in Section 5. In Section 6, we present the results of our experiments. We present an extended discussion on the result caching problem and the results of our experiments in Section 7. Finally, we conclude the paper in Section 8.