Date of Award
Doctor of Philosophy (PhD)
Michael L. Nelson
Michele C. Weigle
Erika F. Frydenlund
With the proliferation of public web archives, it is becoming more important to better profile their contents, both to understand their immense holdings as well as to support routing of requests in Memento aggregators. A memento is a past version of a web page and a Memento aggregator is a tool or service that aggregates mementos from many different web archives. To save resources, the Memento aggregator should only poll the archives that are likely to have a copy of the requested Uniform Resource Identifier (URI). Using the Crawler Index (CDX), we generate profiles of the archives that summarize their holdings and use them to inform routing of the Memento aggregator’s URI requests. Additionally, we use full text search (when available) or sample URI lookups to build an understanding of an archive’s holdings. Previous work in profiling ranged from using full URIs (no false positives, but with large profiles) to using only top-level domains (TLDs) (smaller profiles, but with many false positives). This work explores strategies in between these two extremes.
For evaluation we used CDX files from Archive-It, UK Web Archive, Stanford Web Archive Portal, and Arquivo.pt. Moreover, we used web server access log files from the Internet Archive’s Wayback Machine, UK Web Archive, Arquivo.pt, LANL’s Memento Proxy, and ODU’s MemGator Server. In addition, we utilized historical dataset of URIs from DMOZ.
In early experiments with various URI-based static profiling policies we successfully identified about 78% of the URIs that were not present in the archive with less than 1% relative cost as compared to the complete knowledge profile and 94% URIs with less than 10% relative cost without any false negatives. In another experiment we found that we can correctly route 80% of the requests while maintaining about 0.9 recall by discovering only 10% of the archive holdings and generating a profile that costs less than 1% of the complete knowledge profile.
We created MementoMap, a framework that allows web archives and third parties to express holdings and/or voids of an archive of any size with varying levels of details to fulfil various application needs. Our archive profiling framework enables tools and services to predict and rank archives where mementos of a requested URI are likely to be present.
In static profiling policies we predefined the maximum depth of host and path segments of URIs for each policy that are used as URI keys. This gave us a good baseline for evaluation, but was not suitable for merging profiles with different policies. Later, we introduced a more flexible means to represent URI keys that uses wildcard characters to indicate whether a URI key was truncated. Moreover, we developed an algorithm to rollup URI keys dynamically at arbitrary depths when sufficient archiving activity is detected under certain URI prefixes. In an experiment with dynamic profiling of archival holdings we found that a MementoMap of less than 1.5% relative cost can correctly identify the presence or absence of 60% of the lookup URIs in the corresponding archive without any false negatives (i.e., 100% recall). In addition, we separately evaluated archival voids based on the most frequently accessed resources in the access log and found that we could have avoided more than 8% of the false positives without introducing any false negatives.
We defined a routing score that can be used for Memento routing. Using a cut-off threshold technique on our routing score we achieved over 96% accuracy if we accept about 89% recall and for a recall of 99% we managed to get about 68% accuracy, which translates to about 72% saving in wasted lookup requests in our Memento aggregator. Moreover, when using top-k archives based on our routing score for routing and choosing only the topmost archive, we missed only about 8% of the sample URIs that are present in at least one archive, but when we selected top-2 archives, we missed less than 2% of these URIs. We also evaluated a machine learning-based routing approach, which resulted in an overall better accuracy, but poorer recall due to low prevalence of the sample lookup URI dataset in different web archives.
We contributed various algorithms, such as a space and time efficient approach to ingest large lists of URIs to generate MementoMaps and a Random Searcher Model to discover samples of holdings of web archives. We contributed numerous tools to support various aspects of web archiving and replay, such as MemGator (a Memento aggregator), Inter- Planetary Wayback (a novel archival replay system), Reconstructive (a client-side request rerouting ServiceWorker), and AccessLog Parser. Moreover, this work yielded a file format specification draft called Unified Key Value Store (UKVS) that we use for serialization and dissemination of MementoMaps. It is a flexible and extensible file format that allows easy interactions with Unix text processing tools. UKVS can be used in many applications beyond MementoMaps.
"MementoMap: A Web Archive Profiling Framework for Efficient Memento Routing"
(2020). Doctor of Philosophy (PhD), Dissertation, Computer Science, Old Dominion University, DOI: 10.25777/5vnk-s536